Artificial intelligence in biomarker discovery for diseases: diagnostic and therapeutic prospects

Rasit Dinc; Nurittin Ardic

doi:10.1038/s41392-026-02946-4

. 2026 Sep 25;11:407. doi: 10.1038/s41392-026-02946-4

Artificial intelligence in biomarker discovery for diseases: diagnostic and therapeutic prospects

Rasit Dinc ^1,^✉, Nurittin Ardic ²

PMCID: PMC13612816 PMID: 42786165

Abstract

Biomarkers are central to modern diagnostics and therapeutics, yet traditional discovery approaches suffer from single-modality analyses, weak mechanistic foundations, and low reproducibility. Artificial intelligence (AI) enables the integration of complex multimodal biomedical data, but the translation of AI-derived biomarkers into clinical applications remains inconsistent. This review examines how AI reshapes biomarker discovery across diseases, focusing on biological mechanisms, validation requirements, and therapeutic integration. We synthesize evidence across molecular, cellular, imaging, and digital biomarker approaches, evaluating AI methodologies including classical machine learning, deep learning, graph-based models, foundation models, and causal inference. Findings are organized using a pipeline framework encompassing discovery, external validation, robustness testing, clinical utility, and deployment. AI enables identification of multiscale signatures reflecting cellular programs, tissue remodeling, and disease trajectories. However, many biomarkers fail external validation. Biologically grounded approaches using network-based modeling, spatial profiling, and mechanistic constraints improve interpretability and therapeutic relevance in oncology, cardiovascular disease, neurodegeneration, and metabolic diseases. Validation frameworks with decision curve analysis and prospective evaluation are necessary to demonstrate clinical utility beyond predictive accuracy. AI’s impact will depend on integration into rigorous validation and therapeutic pathways. By transitioning from correlational pattern recognition to mechanistically informed, clinically evaluated systems, AI can support next-generation precision diagnostics and targeted interventions.

Subject terms: Predictive markers, Prognostic markers

Introduction

Biomarkers form the foundation of precision medicine by defining how diseases are detected, classified, and treated.¹ Yet the transfer of candidate biomarkers to clinical practice is disproportionately limited compared to their discovery output. Most biomarkers show promising performance in initial studies but are not reproducible in independent cohorts, degrade under real-world heterogeneity, or prove insufficient to change clinical decisions. This persistent implementation gap reflects weaknesses in validation processes rather than technological capability. This is a challenge repeatedly highlighted in translational biomarker research.² Reproducibility remains a central challenge in biomarker development, as measured signals are susceptible to cohort-specific effects, pre-analytical variability, platform drift, batch effects, and confounding factors related to concomitant diseases and treatment. High-dimensional feature domains further increase the risk of overfitting when sample sizes are modest and analytical flexibility is high.^3–5 Even when biomarker performance appears strong, evaluation often relies on discrimination measures such as the area under the curve (AUC) without demonstrating whether the biomarker actually improves patient management or outcomes. This highlights the distinction between statistical performance and clinical benefit.⁶ Many clinical contexts demonstrate that incremental gains in statistical performance have a negligible impact on actual clinical decisions, especially when baseline performance is strong, or workflow constraints limit implementation.^7,8 These limitations reveal that successful biomarker translation requires much more than discovery. It demands rigorously validated development processes, transparent reporting, and documented clinical benefit. Similar concerns have been emphasized in previous assessments of biomarker development and translation.⁹ These requirements become even more critical when artificial intelligence (AI) methods are applied to complex, high-dimensional datasets.

Biomarker discovery now occurs within an exponentially expanding ecosystem of measurement. Contemporary studies integrate genomic and epigenomic variation, bulk and unicellular transcriptomics, proteomics and post-translational modifications, metabolomics, circulating nucleic acids, extracellular vesicle content, quantitative imaging features, and longitudinal digital phenotypes derived from wearable sensors and electronic health records.^10–14 This diversity reflects an evolving understanding of disease biology, as most disorders arise not from isolated molecular abnormalities but from disruptions distributed across signaling pathways, cellular states, tissue microenvironments, and systemic host responses. Consequently, modern biomarker research faces numerous statistical and biological challenges, including high dimensionality relative to sample size, complex confounding relationships, heterogeneity among real-world populations, inter-regional protocol and device variability, and temporal changes that can alter biological relationships over time. Traditional statistical approaches remain indispensable for inference, but they are often inadequate for large-scale representation learning and multimodal data integration. AI, particularly deep learning, has demonstrated the ability to learn hierarchical representations from complex biomedical inputs and holds great promise, especially in imaging and multimodal clinical datasets.¹⁵ However, increasingly rich datasets also create new opportunities for bias and error. Region-specific signatures, hidden confounding factors, and label leakage can produce seemingly accurate models that fail when evaluated in independent cohorts or real-world settings.¹⁶ Similarly, models can learn superficial correlations that provide strong predictive performance during development but lack biological relevance and robustness. This contrast between unprecedented opportunities for biomarker discovery and the increasing risk of erroneous findings represents a central challenge in contemporary biomarker science. Accordingly, successful translation depends not only on predictive performance but also on rigorous validation, mechanistic interpretation, and demonstration of reproducible biological relevance.

AI is reshaping the entire biomarker development lifecycle. In the discovery phase, machine learning enables scalable feature screening and multivariate signature generation.^17–19 Whereas deep neural networks learn direct representations from raw data such as images, molecular profiles, and longitudinal measurements, capturing complex distributed biological signals missed by univariate approaches.^20,21 In the validation phase, the key challenge is demonstrating robustness, transparency, and generalizability across diverse cohorts and real-world settings. AI-based biomarkers require explicit specification of cohort definition, endpoint labeling, preprocessing decisions, handling of missing data, model development choices, and evaluation protocols.^22,23 The fact that a lack of transparency weakens reproducibility has led to the development of AI-specific reporting frameworks and bias assessment tools. In the clinical utility phase, biomarkers should demonstrate incremental decision value according to established workflows, quantified through decision analytics that translate predictions into actual patient outcomes. This requirement is particularly important in settings where modifying existing tools is impractical and incremental value needs to be precisely defined by the decision point and patient subgroup. At the therapeutic stage, an important objective is to move beyond prediction toward biological interpretation.^24,25 AI-derived biomarkers are increasingly being used to support pathway analysis, target identification, patient classification, trial enrichment, and treatment optimization. Foundation models trained on large-scale biological data point to more transferable and mechanistically meaningful representations of molecular function. However, realizing this promise requires careful separation of correlational signatures from causal or actionable mechanisms through biological validation and orthogonal experiments.

This review presents a mechanism-centric, clinically applied framework for AI-assisted biomarker discovery and adoption in disease domains. Rather than cataloging studies or algorithms, we focus on the principles that determine whether AI-derived biomarkers pass external validation tests, support biological interpretation, and provide clinical and therapeutic value. We organize the discussion according to major biomarker data types, including molecular and cellular measurements, circulating analytes, imaging phenotypes, and digital biomarkers. These types differ in terms of sources of variability, susceptibility to confounding factors, and capacity to support mechanistic interpretation. By mapping AI methods to these biological realities, we separate prediction performance from biologically grounded evidence and highlight common modes of failure. We propose an operational development pipeline (including study design, harmonization, leakage prevention, robustness testing, and biological validation) aligned with contemporary reporting standards and bias assessment frameworks for AI prediction models. The fundamental contribution of this work is to advance AI-assisted biomarker discovery toward the field of reproducible, mechanistically based precision medicine by integrating plausibility and therapeutic actionability within a unified evidence framework. We synthesize representative literature encompassing biomarker validation, AI, systems biology, and translational medicine, with particular emphasis on studies published in the last decade and landmark research shaping current practices. The review is deliberately selective rather than comprehensive, focusing on conceptual integration between biomarker modalities, AI methodologies, biological interpretation, validation requirements, and therapeutic relevance.

To provide an overview of how data modalities, AI modeling, validation, and clinical adoption interact throughout the biomarker lifecycle, Fig. 1 summarizes a generalized AI-enabled biomarker discovery and deployment pipeline.

Biomarker models and biological principles (inter-disease relationships)

Molecular biomarkers: genomics/epigenomics, transcriptomics, proteomics/posttranslational modifications, metabolomics/lipidomics

Molecular biomarkers remain the most mature and widely used class of biomarkers, providing a mechanistic basis for disease stratification and therapeutic targeting. Genomic variation can fix disease risk and subtype definition, while epigenomic regulation captures dynamic gene control states shaped by development, environment, and disease activity.^26–28 In many disorders, particularly cancer, immune-mediated diseases, cardiometabolic syndromes, and neurodegeneration, single-layer biomarkers are increasingly insufficient to explain phenotypic heterogeneity. This situation encourages the development of multi-omics strategies that integrate molecular layers into coherent biological narratives. Among molecular modalities, small molecule metabolites and lipid species provide a sensitive indicator of integrated physiology. This reflects pathway flux, organ crosstalk, inflammatory state, and exposure history.^29,30 Recent syntheses highlight how metabolite biomarkers can support diagnosis, prognosis, and therapeutic target discovery. However, these syntheses also highlight recurring pitfalls such as cohort effects, platform shift, diet/drug-induced confounding, and limited external validation in clinically realistic settings.^31–33 Recent multi-omics studies continue to demonstrate the value of integrated molecular signatures for biomarker discovery and disease stratification.^34,35

Transcriptomic biomarkers are increasingly extending beyond bulk expression profiles toward spatially resolved and single-cell readouts, enhancing biological interpretability while preserving tissue architecture and cell state context. For example, in brain disorders, recent reviews highlight how single-cell and spatial transcriptomics can link disease phenotypes to cell-type-specific programs and spatial microenvironments; this approach can be generalized to oncology, inflammatory diseases, and cardiometabolic organs.^26,27,36

Proteomic biomarkers remain extremely relevant because proteins and post-translational modifications are closer to signal execution than DNA or RNA. However, they present significant analytical complexity and reproducibility challenges.³⁷ Even if molecular signatures are statistically robust, their real-world implementation depends on biological plausibility and clinical utility rather than solely on discrimination metrics. Therefore, mechanistic fixation (pathways, cell programs) and orthogonal validation (independent platforms, targeted assays) are essential.

Cellular and immune biomarkers: immune repertoires, single-cell states, cytokine programs, and NETs

Cellular biomarkers reflect a shift from “single-molecule-derived causality” to disordered cellular states and tissue ecosystems. Single-cell profiling enables the separation of immune activation, depletion, lineage shifts, and inflammatory programs that often cut across classical diagnostic categories.^38–40 Spatial technologies further preserve tissue context, which is critical where pathology is dependent on cellular neighborhoods, barrier interfaces, microvascular niches, or tumor-immune geography.^36,41,42 As the technology landscape evolves rapidly, recent studies have focused on mapping the practical capabilities and limitations of commercial single-cell and spatial platforms, providing a useful basis for selecting methods that fit the biological question and current tissue constraints.^43–45 Recent reviews also synthesize spatially resolved single-cell omics methods, highlighting how computational integration and standardization gaps shape downstream biomarker robustness and cross-center generalizability.^42,45–47

Immune repertoires provide quantitative signatures of adaptive immune dynamics across infection, cancer, autoimmunity, and transplantation.⁴⁸ Whereas cytokine programs and innate immune states serve as integral biomarkers of host response. Neutrophil extracellular traps (NETs) are a concrete example of the cross-disease inflammatory biomarker axis: they associate innate immune activation with vascular injury, thrombosis, and tissue damage, and carry clear therapeutic implications in the context of systemic autoimmune and autoinflammatory diseases.^49,50

A recurring principle is that “cellular biomarkers” are rarely stable scalar variables. They are compositional, state-dependent, and context-sensitive.⁵¹ Therefore, translational robustness requires careful separation of the biological signal from technical variation (partition effects, dissociation artifacts, antibody panels, spatial resolution constraints) and explicit linking of biomarker patterns to cellular programs and mechanisms, rather than narratives of post hoc interpretation. The diversity of biological information captured by modern biomarker platforms necessitates a systematic classification of data modalities.

Because biomarkers are measurements of biological systems at molecular, cellular, tissue, and behavioral scales, AI discovery strategies must address the characteristics, constraints, and dominant noise sources of each data type. Accordingly, Fig. 2 summarizes the major biomarker classes and data sources used for AI-assisted biomarker discovery, encompassing molecular, cellular/immune, circulating, imaging, extracellular vesicle, and digital domains.

Fig. 2 — Biomarker modalities and data sources. Overview of six major biomarker domains used in AI-assisted biomarker discovery, including molecular, cellular/immune, circulating, imaging, extracellular vesicle, and digital biomarkers. Representative analytical methods and biological information captured by each modality are summarized. Figure created by the authors using Adobe Illustrator (Version 29.3.1). RNA-seq RNA sequencing, PTM post-translational modification, TCR T cell receptor, BCR B cell receptor, NETs neutrophil extracellular traps, ctDNA circulating tumor DNA, CT computed tomography, MRI magnetic resonance imaging, PET positron emission tomography, H&E hematoxylin and eosin, EV extracellular vesicle, MISEV Minimal Information for Studies of Extracellular Vesicles

Circulating biomarkers: cfDNA/cfRNA, circulating proteins/metabolites, liquid biopsy paradigms

Circulating biomarkers provide minimally invasive access to disease biology and are particularly valuable for longitudinal monitoring. Cell-free DNA (cfDNA) or RNA (cfRNA) and circulating tumor DNA (ctDNA) exemplify classes of biomarkers that can support early detection strategies, residual disease monitoring, response assessment, and clonal evolution tracking.^52–55 Yet their clinical value largely depends on pre-analytical processing, analytical sensitivity, and clinical context (tumor burden, shedding rates, treatment timing), all of which affect the stability of the derived biomarkers. Recent syntheses highlight the current status, advantages, and limitations of liquid biopsy across tumors, emphasizing both clinical opportunities (early diagnosis, dynamic monitoring) and translational bottlenecks (standardization, false positives, access, and evidence quality).^56–58 Emerging studies on minimal residual disease monitoring further support the clinical value of liquid biopsy approaches in longitudinal cancer management.^59–61

In non-oncological diseases, circulating proteins and metabolites remain clinically dominant due to assay maturity and implementation feasibility, but they often reflect downstream physiology rather than proximate causal mechanisms.^31,62–64 This creates a predictable mode of failure: biomarker performance appears promising in selected cohorts, but generalization is poor when concomitant diseases, drugs, and spectrum effects disrupt baseline distributions. Cross-mode integration (e.g., pairing circulating signals with imaging or tissue-derived molecular states) can improve biological interpretability and stability if validation is rigorous and leakage is avoided.

Extracellular vesicles/exosomes as biomarker carriers

Extracellular vesicles (EVs), including exosomes and microvesicles, offer a biologically attractive biomarker model because they carry a multimolecular payload reflecting cellular status and intercellular communication. Therefore, EV biomarkers are located at the intersection of molecular and cellular biomarkers: they can capture cell-of-origin signals while encoding systemic communication programs.^10,65

Yet the conversion of EVs into biomarkers remains constrained by heterogeneity in isolation methods, co-isolated contaminants (including lipoproteins), variable quantification standards, and a lack of consensus among laboratories.⁶⁶ A significant recent development is the updated consensus guideline MISEV2023 from the International Society for Extracellular Vesicles (IESV), which expands minimum information standards from basic to advanced approaches and explicitly addresses expectations of methodological diversity and reproducibility.^67,68

For AI-assisted discovery, EVs are a high-opportunity but high-risk method: the signal is rich and multilayered, yet the feature domain is highly sensitive to pre-analytical variation.^69–71 This makes EV biomarkers a strong test case for the validation ladder, specifically external replication, stability analyses, and biologically grounded validation. Recent AI-assisted studies have further revealed the diagnostic potential of extracellular vesicle-derived markers in multiple tumor types.⁷²

Imaging biomarkers: radiomics, functional imaging, quantitative imaging processes

Imaging biomarkers enable non-invasive phenotyping of tissue structure and function at the organ scale. They can encode clinically relevant heterogeneity (e.g., fibrosis architecture, vascular remodeling, metabolic activity, lesion tissue, and spatial distribution) and often capture biological variation that correlates with molecular status and prognosis. Radiomics and computational imaging methods systematize this quantification, but real-world implementation requires tight control of scanner variability, acquisition protocols, segmentation procedures, and inter-site harmonization.^12–14

A significant recent development is the emergence of “foundation model-ready” imaging biomarker discovery: self-supervised models pre-trained on large imaging datasets can learn transferable feature representations and then be adapted to specific biomarker tasks. Similar advances have been reported in digital pathology and multimodal imaging foundation models. These further expand the translational potential of imaging biomarkers.^73–75 In a representative study, Pai et al. published their work developing a foundational model for the discovery of cancer imaging biomarkers.⁷⁶ By evaluating this across clinically relevant applications, they demonstrated how representation learning can enable more generalizable imaging biomarker processes.

At the same time, recent critical reviews on precision oncology highlight that large-scale validation remains the limiting step for radiomic biomarkers, particularly in high-risk contexts such as immunotherapy response prediction.⁷⁷ These analyses reinforce the core theme of this review: imaging biomarkers are successful only when supported by biological anchoring and prospective evidence, not merely retrospective discrimination.

Digital biomarkers: wearable devices, passive sensing, digital phenotypes and linkage to outcomes

Digital biomarkers are derived from continuous or high-frequency data streams generated by wearable devices, smartphones, and ambient sensors. These methods extend biomarker science beyond molecular analyses to computational physiology and behavioral phenotyping. Eventually, they enable longitudinal monitoring of sleep, mobility, autonomic physiology, and symptom dynamics. Their translational appeal lies in their scalability and temporal resolution, particularly for chronic disease monitoring and early deviation detection.^78–80 Recent wearable device research continues to expand the scope of digital biomarkers for physiological monitoring and disease prediction.^81,82

Yet the concept of “digital biomarkers” remains heterogeneous. A recent systematic mapping of the biomedical literature reveals significant differences in how the term is defined, with direct implications for study design, validation, and clinical benefit claims.⁸³ This descriptive fragmentation interacts with real-world technical challenges such as device heterogeneity, incompleteness, behavioral confounding, and socioeconomic bias. This necessitates careful endpoint alignment and validation. Digital biomarkers become most defensible when matched with interpretable physiological constructs or functionally meaningful outcomes, and when their reliability is demonstrated under realistic application conditions rather than controlled research settings.

The “Biological plausibility” framework: distinguishing causal, mechanistic, and correlational biomarkers

Across all biomarker modalities, biological plausibility is the gatekeeper of translation. Many biomarkers are predictive correlations that track disease burden without participating in disease mechanisms. Such biomarkers can still be clinically useful (e.g., risk stratification), but they often fail as therapeutic guidelines or surrogate endpoints when mechanistic assumptions are overstated. A practical plausibility framework emphasizes several dimensions: consistency across cohorts and platforms; temporal alignment with disease progression; dose-response behavior; specificity to biological processes rather than technical artifacts; and consistency with known pathways, cell programs, and tissue context. Stronger mechanistic support comes from genetic linkage, perturbation evidence, spatial localization of the signal, and longitudinal dynamics that match causal hypotheses. This framework sets the groundwork for later sections on mechanistic AI. The aim here is not only to predict outcomes but also to map biomarker signatures to pathways and actionable therapeutic hypotheses.

To illustrate how different classes of biomarkers contribute to AI-assisted discovery and application, Table 1 summarizes the major types of biomarkers according to their biological scale, technical characteristics, commonly applied AI methods, and current application maturity stages.

Table 1.

Biomarker models and translational properties

Biomarker modality	Typical data sources	Biological level	Strengths	Key limitations	AI methods commonly used	Translation maturity
Genomics/epigenomics	WGS, SNP sequences, methylation profiles	Germline/regulatory	Stable, causal inference, lifetime risk	Limited dynamic monitoring, modest effect sizes	Regularized models, GNNs, polygenic modeling	High (risk prediction)
Transcriptomics (bulk/scRNA)	Blood, tissue biopsies	Cellular programs	Pathway-level insight, immune status	Batch effects, tissue access, cost	Clustering, VAEs, graph models	Medium
Proteomics/metabolomics	Plasma, CSF, tissue	Functional molecular activity	Proximal to phenotype, druggable pathways	Technical variability, platform dependence	Tree ensembles, representation learning	Medium
ctDNA/liquid biopsy	Plasma cfDNA, RNA	Tumor-derived signals	Minimal residual disease, mutation tracking	Low abundancy, test sensitivity	Deep sequencing ML, Bayesian models	High (oncology)
Imaging radiomics	MRI, CT, PET	Organ and tissue structure/function	Spatial context, longitudinal tracking	Scanner heterogeneity, interpretability	CNNs, vision transformers	High
Extracellular vesicles	Plasma, urine, CSF	Intercellular signaling	Mechanistic signaling insight	Isolation standards, low throughput	Feature learning, multi-omics fusion	Low-medium
Digital biomarkers	Wearable devices, smartphones, electronic health record streams	Behavioral/physiological	Continuous monitoring, scalable	Adherence bias, device variability	Time series models, DL	Medium

Open in a new tab

Key references for this table: refs. ^{25,67,147,170,278–280}

Collectively, these biomarker classes operate at distinct biological scales and exhibit significant differences in analytical stability, mechanistic interpretability, and clinical scalability. The comparison also underscores the need for multimodal integration strategies discussed in Sections “Artificial intelligence methods for biomarker discovery” and “Mechanistic AI: mapping biomarkers to pathways, targets, and therapeutic hypotheses”, highlighting that no single biomarker approach is sufficient to capture complex disease biology. It also emphasizes that translational readiness varies considerably among biomarker classes. As will be discussed in the following section, these modality-specific constraints directly shape the selection and performance of AI methods.

Artificial intelligence methods for biomarker discovery

Classical machine learning for high-dimensional biomarker data

Many biomarker discovery settings remain fundamentally “small-n, large-p settings where the number of variables being measured can far exceed the number of subjects. In this regime, classical machine learning (ML) methods such as regularized regression, support vector machines (SVM), tree ensembles, and sparse feature selection retain their practical value because they provide strong inductive bias and can be paired with stability analysis.^84,85 Their strengths are most evident when the scientific aim is not only prediction but also generating a suitable shortlist of candidates for orthogonal validation. However, these same environments are quite vulnerable to optimistic predictions driven by leakage, repeated tuning, and selection on outcome. This means a strict training/validation separation and nested evaluation are essential.⁸⁵

A second limitation is interpretability inflation. Feature significance or coefficient magnitude is often treated as biologically grounded evidence, even if related features are interchangeable and the learned signature is not unique. In high-dimensional biology, “explanations” can be unstable unless supported by stability based on resampling, external replication, and alignment with path structure.

Deep representation learning and self-supervised learning for omic and single-cell data

Deep learning (DL) is most advantageous when biomarker signals are distributed and nonlinear, or when inputs are not naturally tabular (e.g., arrays, images, graphs). A significant advance is the shift towards representation learning, where models learn reusable embedding vectors that capture biological structure and can be adapted to downstream tasks with limited labels.^86,87 In single-cell genomics, self-supervised learning (SSL) has emerged as a prime way to leverage large unlabeled datasets, but its benefits are not homogeneous across tasks. Benchmark-focused analyses highlight that SSL is most helpful when transfer is required across batches, tissues, or experimental settings and downstream task labeling is limited.⁸⁸

The practical implication for biomarker discovery is that SSL should not be framed as a universal upgrade. Its value depends on whether the pre-training objectives align with biological invariances, whether the batch structure dominates the embedding space, and whether the learned representations improve generalization rather than merely compressing technical variation.⁸⁸ When used appropriately, representation learning can support both prediction and discovery by enabling downstream probe of latent dimensions and stabilizing models trained on limited cohorts.

Multimodal learning in omics, imaging, and clinical data

Clinical biomarkers rarely exist in isolation. They are interpreted in conjunction with imaging, lab values, longitudinal clinical courses, and often text-derived context. Multimodal ML aims to formalize this integration, learning complementary information across data types while avoiding the pitfalls of naïve concatenation.^89–91 Modern multimodal pipelines are often defined in terms of merging strategies (early, mid, or late merging), and this choice greatly influences the tolerance for missing data and whether cross-modal interactions can be learned.⁹²

A particularly important model in biomarker discovery is transfer learning across modalities, where large observational EHR-scale datasets are used to pre-train components that are then transferred to smaller cohorts with paired omics and clinical data.^33,93 This approach directly addresses the power limitation of omics studies and can improve both prediction and biological insight when implemented with careful cohort separation and robust evaluation.⁹⁴ The broader lesson is that multimodal models should be designed around realistic clinical data availability (partial modality coverage, asynchronous sampling, and variable quality) rather than idealized assumptions of complete cases.⁹²

Graph neural networks and knowledge-guided learning

Many biomarker-relevant relationships are naturally relational, including protein-protein interactions, regulatory networks, cell-cell communication, pathway graphs, and multimodal linkages between genes, variants, drugs, and phenotypes. Graph neural networks (GNNs) enable learning on such non-Euclidean structures and have rapidly expanded across single-cell tasks, including imputation, batch correction, cell typing, spatial domain detection, and network inference.^71,95,96 Their most important advantage for biomarker discovery is their ability to add structural priors, which can constrain the hypothesis space and increase robustness in limited sample sizes.⁹⁷

A parallel direction is mechanism-aware architectures that explicitly encode signal or pathway structure to encourage biologically plausible computation. Methods such as GNNs have been proposed to model perturbation biology by constraining the flow of information through structured interaction graphs and pairing them with explanatory modules aimed at localizing the graph substructures responsible for predictions.⁹⁸ These approaches are particularly attractive for target discovery and therapeutic classification because they can align predictable features in interpretable ways, but still require rigorous validation to ensure that interpretability is not an afterthought narrative.

Foundation and large-scale pre-training in genomics and biomedicine

Foundation models trained on large biological datasets are beginning to reshape biomarker discovery by providing transferable representations across tasks and domains. In genomics, DNA “language models” trained with self-supervised targets can learn sequence representations, improving downstream prediction of molecular phenotypes and variant effects, and supporting biomarker processes starting from sequence-level function.^99,100 This class of models is rapidly evolving, and reviews now categorize foundation models across genomics, transcriptomics, proteomics, drug discovery, and single-cell analysis. The reviews also highlight both opportunities and unresolved limitations of these models in evaluation and interpretability.^46,101 These developments support the emerging concept of a biological “language of life” that can be learned through large-scale foundational models.¹⁰²

In practice, foundation models change biomarker workflows in two ways.^103–106 First, they can reduce reliance on task-specific labeled datasets by enabling fine-tuning or adaptation with limited supervision. Secondly, they encourage a shift from handcrafted features toward learned representations, which can improve generalization ability but also increase the need for auditing: pre-training corpora can encode batch, ancestry, or site signals that later appear as “biomarkers” unless explicitly controlled.

Generative and simulation-assisted approaches for biomarker mechanism and target discovery

Beyond prediction, generative modeling offers a pathway toward biologically grounded hypothesis generation. In biomarker discovery, generative models can be used to simulate counterfactual molecular states, predict perturbation responses, or suggest molecular patterns consistent with observed phenotypes.¹⁰⁷ The most credible uses are those tied to perturbation datasets, causal structure, or explicit biological constraints, as unrestricted generation can produce outputs that seem plausible but are not scientifically grounded. In perturbation biology, mechanism-guided graphical models demonstrate how generative or predictive learning can be anchored to curated biological structure and experimental perturbations, providing a more testable bridge between signatures and actionable pathways.⁹⁸

Generative and simulation-assisted outputs should therefore be interpreted as hypothesis-generating rather than evidentiary by default. Within the validation ladder proposed in this review, such models primarily contribute at the discovery stage by proposing candidate signatures, latent structures, pathway hypotheses, or putative mechanisms. Moving beyond this stage requires orthogonal confirmation: Level 1 requires internal analytical consistency and protection against leakage or memorization; Level 2 requires external replication across independent datasets or sites; Level 3 requires robustness testing under perturbation, subgroup shift, or alternative preprocessing pipelines; and advancement toward Levels 4 and 5 requires demonstrating that the generated hypothesis improves clinically meaningful decisions rather than merely producing biologically plausible outcomes. Within this framework, generative AI functions as a discovery engine, whereas predictive AI more directly supports risk estimation or classification in downstream clinical workflows.⁹⁸

To highlight which AI approaches have progressed from proof-of-concept studies to real-world clinical workflows, Table 2 summarizes representative biomarker-driven AI systems with documented clinical deployment and/or regulatory clearance across multiple disease domains.

Table 2.

AI methods with demonstrated clinical deployments in biomarker-based applications (representative examples)

Clinical use case (biomarker function)	Input method/“biomarker”	Dominant AI method	Deployment context	Regulatory/clinical status	Representative evidence (examples; not exhaustive)
Autonomous diabetic retinopathy screening (diagnosis/triage biomarker)	Retinal fundus photographs	CNN-based classifiers	Primary care/screening settings (autonomous operation)	FDA De Novo clearance (IDx-DR)	²⁸¹
Atrial fibrillation detection from wearable ECG (rhythm biomarker)	Single-lead wearable or clinical ECG	Signal processing + CNN	Consumer devices with clinician confirmation	FDA-cleared ECG AF detection algorithm	⁸⁰
Large-vessel occlusion stroke triage (imaging biomarker)	Head/neck CTA	CNN-based detection + workflow prioritization	Emergency stroke networks (alert routing)	FDA-cleared triage software (e.g., Viz.ai)	²⁸²
Intracranial hemorrhage triage (imaging biomarker)	Head NCCT	CNN-based detection	Emergency radiology triage	FDA-cleared notification system	²⁸³
Prostate cancer detection support (histopathological biomarker)	Whole slide images pathology image (H&E)	CNN/vision transformer	Pathologist assistive diagnosis	FDA-cleared assistive CAD tools	¹⁰⁴
Early sepsis deterioration prediction (physiological/EHR biomarker)	EHR time-series (vitals, labs)	Gradient boosting/RNN	Inpatient clinical decision support	Clinical deployment; mixed regulatory status	²⁸⁴
Low ejection fraction screening from ECG (functional biomarker)	Standard 12-lead ECG	CNN-based phenotype prediction	Opportunistic population screening	Deployment in hospital ECG system	¹⁹⁴
Pulmonary embolism triage from CT (imaging biomarker)	Chest CT/CTA	CNN-based detection + workflow prioritization	Radiology workflow triage	FDA-cleared triage software	²⁸⁵

Open in a new tab

CTA CT angiography, NCCT non-contrast CT, CNN convolutional neural network, SaMD software as a medical device, CAD computer-aided detection, EHR electronic health record

Explainability, uncertainty, and the risk of “false mechanistic comfort”

Explainability is essential in biomarker discovery because the output is rarely a final decision; it is a candidate hypothesis that must be subject to biological validation.^108,109 However, explanation methods can create false confidence if treated as mechanistic evidence rather than diagnostic tools. Recent domain-focused reviews highlight that post hoc explainability can be useful for triage and auditing, but are sensitive to correlated features, model unidentifiability, and dataset shift. These conditions are routinely encountered in omics.^110,111 Bioinformatics-focused survey further highlights the diversity of explainability techniques and the frequent mismatch between explanation outputs and actionable biological interpretation.¹⁰⁹ A practical principle is to treat interpretability as evidence that must replicate: do the same features/pathways appear under resampling, across cohorts, across platforms, and under perturbation? Wherever possible, explanations should be tested against orthogonal experiments (e.g., CRISPR perturbation, proteomic validation, targeted assays) rather than simply being reported.

Biomarker type and method selection as a translational objective function

No single AI method is optimal among biomarker modalities. Classical ML remains strong for small cohort tabular omics, especially when interpretability and stability are prioritized. DL is most justified when inputs are high-dimensional and structured (arrays, images, graphs) or when transfer learning and pre-training are available. Multimodal learning becomes essential when clinical action depends on the integration of heterogeneous streams of evidence, and graph/information-guided learning is most valuable when biological prior information can meaningfully constrain the hypothesis space.

Since different AI architectures are preferentially suited to distinct biomarker discovery tasks, Fig. 3 maps commonly used ML paradigms matched with fundamental analytical goals such as feature selection, multimodal integration, target identification, and mechanistic inference.

Across all settings, the core requirement is that the method selection be aligned with the intended claim, including predictive enrichment, diagnostic discrimination, biologically grounded pathway nomination, target discovery, or therapeutic response classification. Each claim requires a different standard of evidence, and each is weakened by different modes of failure. Positioning AI methods within the logic specific to this claim is the most reliable way to prevent biomarker discovery from drifting toward elegant models with fragile scientific meaning. While the methods in Section “Artificial intelligence methods for biomarker discovery” provide powerful tools for extracting predictive signatures from complex biomedical data, prediction alone is insufficient when biomarker claims extend to areas such as biological interpretation, therapeutic targeting, or surrogate endpoint development. For biomarkers to inform mechanism and intervention, AI models must be coupled to biological structure, causal hypotheses, and perturbation-sensitive validation. Therefore, mechanistic AI represents a distinct paradigm in biomarker science; one paradigm that aims not only to predict outcomes but also to map signatures to pathways, cellular programs, and actionable molecular processes. The following section examines AI strategies that explicitly incorporate biological priors, experimental perturbation, and causal structure to support mechanistically grounded biomarker discovery.

Since different AI architectures are suited to different biomarker tasks and data structures, Table 3 maps the major methodological classes according to their best applications, strengths, and risks in biomarker development.

Table 3.

AI method classes and suitability for biomarker tasks

AI method class	Optimal biomarker tasks	Data types	Strengths	Major risks	Interpretability Potential	Example use cases
Regularized regression	Risk scores, screening	Tabular clinical/omics	Stable, transparent, robust	Limited nonlinear capture	High	Cardiovascular risk, fibrosis scores.
Tree ensembles	Feature selection, response prediction	Mix tabular data	Handles interactions and missing data	Overfitting in small groups	Medium	Treatment response models
Convolutional neural networks (CNNs)	Imaging biomarkers	MRI, CT, histopathology	Spatial feature learning	Scanner/site shift	Low-medium	Tumor segmentation, plaque analysis
Transformers	Multimodal fusion, arrays	Imaging + omics + EHR	Cross-modal attention	High data demand, non-transparent structure	Low	Survival prediction
Graph neural networks	Pathway inference, cell interactions	Molecular networks, scRNA	Mechanistic modeling	Complex training	Medium	Immune network biomarkers
Foundation models	Transfer between diseases/sites	Imaging, omics, multimodal	Data efficiency, reuse	Bias spread, cost	Low	Cross-cohort imaging
Causal machine learning	Surrogate validation, treatment effects	Longitudinal multimodal	Causal reasoning	Strong assumptions	High (if applicable)	Biomarker-outcome linkage

Open in a new tab

Key references for this table: refs. ^{21,112,286,287}

Mechanistic AI: mapping biomarkers to pathways, targets, and therapeutic hypotheses

From correlational signatures to mechanistic hypotheses

Most AI-derived biomarkers are correlational by construction. They identify patterns associated with outcomes, but they do not specify why these patterns occur. While correlational biomarkers are clinically useful, they are fragile when used to guide treatment, predict response mechanisms, or serve as surrogate endpoints.^112–115 Mechanistic biomarker discovery requires additional constraints linking features to biological processes.

Recent reviews emphasize that mechanistic interpretation should integrate computational evidence with biological plausibility, perturbation consistency, and pathway coherence, rather than relying solely on pattern explainability techniques.⁹⁵ Without such integration, feature attribution risks generating compelling narratives not supported by causal biology. Therefore, mechanistic AI aims to directly encode biological structure into model architectures or training targets and narrow the hypothesis space to biologically plausible explanations.

Knowledge-guided and pathway-constrained neural architectures

One approach to mechanistic modeling directly embeds compiled biological information, such as signaling pathways, transcriptional regulatory networks, or metabolic graphs, into neural network topology. In these architectures, nodes correspond to genes or proteins, and edges reflect known interactions, thus restricting the flow of information to biologically meaningful pathways.^95,116

Graphical-structured neural networks have been used to model perturbation responses, drug sensitivity, and pathway activation by forcing predictions to traverse interaction networks, enabling the identification of subnetworks associated with phenotypic change.⁹⁸ This design enhances interpretability and stability when sample sizes are limited, as prior biological information reduces model flexibility.

Nevertheless, pathway databases are incomplete, context-dependent, and sometimes contradictory.^117,118 Consequently, knowledge-guided models should be interpreted as hypothesis generators rather than definitive causal evidence and supported by data-driven components that allow for the discovery of novel interactions.

Perturbation-based learning and causal signal identification

Perturbation datasets, such as CRISPR scans, drug response assays, and cytokine stimulation experiments, provide critical leverage for separating causal factors from downstream correlations.^119,120 AI models trained on perturbation responses enable counterfactual prediction and target prioritization by learning mappings between molecular states and phenotypic outcomes.

Recent reviews highlight that perturbation-based ML is particularly powerful when combined with network prioritization and multimodal measures, supporting the inference of pathway dependencies and drug mechanisms. These approaches are increasingly applied in oncology, immunology, and systems pharmacology, where biologically grounded hypotheses can be experimentally tested.^121,122 Importantly, perturbation learning only increases mechanistic reliability when the experimental design includes adequate controls, replications, and the relevant biological domain. Sparse or biased perturbation panels can produce misleading inferences, even in advanced models.

Causal representation learning and counterfactual biomarker modeling

Causal representation learning aims to identify latent variables corresponding to underlying causal factors rather than superficial correlations. In biomarker discovery, this framework aims to separate disease-triggering processes from confounding factors such as technical artifacts, comorbidities, or treatment effects.^112,123

Recent methodological syntheses explain how invariant risk minimization, instrumental variable approaches, and counterfactual targets can improve generalization across environments by prioritizing features that are stable under intervention.^112,124 When applied in biomarker development, these methods can help identify signatures that are more likely to reflect causal disease mechanisms and therefore maintain validity across populations. However, causal learning in observational biomedical data remains challenging due to unmeasured confounding factors, feedback loops, and dynamic disease processes. Causal AI methods should therefore be viewed as a reinforcing, not a replacement for, the need for experimental validation.

Linking biomarker signatures to drug response and therapeutic classification

A major objective of mechanistic biomarker discovery is to link biomarker signatures with biological processes relevant to treatment response. Unlike purely correlational biomarkers, mechanistic-based biomarkers can provide information about pathway activity, cellular states, and molecular interactions that contribute to disease progression or treatment sensitivity.

Network-based analyses, pathway-level representations, and causal modeling approaches can help identify biological programs associated with treatment response.^84,98,116 Knowledge-graph approaches incorporating perturbation transcriptomic data have been used to infer compound-protein interactions and connect molecular perturbations with candidate therapeutic targets. Therefore, they have improved mechanistic interpretation beyond conventional predictive modeling.¹²⁵ By integrating biomarker signals with existing biological information, these methods can generate hypotheses about potential points of intervention, resistance mechanisms, or patient subgroups with differing treatment sensitivities. Such analyses can enhance biological interpretability and facilitate the prioritization of candidate therapeutic targets for further investigation.

However, identifying mechanistically plausible targets does not guarantee clinical benefit. Biological associations need to be validated through experimental studies, perturbation-based investigations, and independent clinical datasets to confirm therapeutic significance. In conclusion, mechanistic AI approaches should be viewed primarily as tools for hypothesis generation and biological interpretation, rather than as direct evidence of treatment efficacy.

The translation of biomarker-guided hypotheses into clinical decision-making frameworks, patient classification strategies, accompanying diagnostic methods, and drug development applications is discussed in Section “Translational roles of AI biomarkers in therapeutic development and clinical decision-making processes”.

Limitations of mechanistic AI and the role of experimental validation

Despite rapid advances, mechanistic AI cannot establish causality on its own.^112,113 Biological systems exhibit redundancy, compensation, and context specificity, limiting inference from static data.^11,126 Especially in immune and neurodegenerative disorders, even perturbation experiments may not fully reflect in vivo complexity.¹²⁷

Therefore, mechanistic AI must be embedded in an iterative discovery cycle consisting of computational hypothesis generation, targeted experimentation, model refinement, and clinical correlation. This cycle reduces the risk of overinterpreting model descriptions and aligns biomarker development with experimental biology.^128,129 From a translational perspective, mechanistic rationality strengthens regulatory and clinical confidence, but only when supported by reproducible evidence across platforms and disease contexts. Therefore, mechanistic AI should be viewed as a facilitator of biologically based biomarker processes, rather than a replacement for rigorous validation.^112,130

While mechanistic AI provides the methodological foundation for moving beyond correlational signatures, the usefulness of translational biomarkers ultimately depends on whether AI-generated patterns can be mapped into actionable signal transduction states that inform therapeutic intervention.

While mechanistic AI provides the methodological foundation for moving beyond purely correlational biomarker signatures, its primary contribution lies in the inference of biologically significant pathway activity states and network relationships. These approaches enhance the mechanistic understanding of biomarker signaling by helping to transform high-dimensional molecular data into interpretable representations of disease biology. Therefore, the following section focuses on signal transduction networks and pathway state biomarkers as biological frameworks within which AI-derived biomarker signatures can be interpreted and contextualized.

Signal transduction networks, pathway-state biomarkers, and therapeutic targeting

Why pathway states outperform static biomarkers

Most conventional biomarkers quantify the abundance of molecules (gene expression levels, protein concentrations, or circulating analytes) at a single point in time. While such measurements can be correlated with disease states, they often fail to capture the functional activity of underlying biological systems, particularly in complex and dynamic diseases. In contrast, signaling pathways represent integrated, dynamic processes that encode how cells interpret and respond to internal and external stimuli. Consequently, pathway activity states, rather than static molecular levels, are often more directly linked to disease mechanisms and therapeutic response.

For example, activation of the PI3K-AKT-mTOR or MAPK pathways cannot be reliably inferred from the expression of individual components alone, due to post-translational regulation, feedback loops, and compensatory signaling. Similarly, inflammatory signaling pathways such as NF-κB and JAK-STAT exhibit context-dependent activation patterns influenced by cytokine gradients, cellular composition, and microenvironmental cues.^131,132 Consequently, biomarkers based solely on abundance measurements can misclassify patients with similar molecular levels but significantly different pathway activation states.

AI methods can assist in inferring latent pathway states by integrating transcriptomic, proteomic, imaging, and clinical information.^89,90 Instead of addressing features independently, these models can learn coordinated patterns across molecular networks, enabling the identification of pathway-level signatures that better reflect biological activity. For example, multi-omics integration approaches using machine learning frameworks have demonstrated the ability to capture the coordinated activation of signaling modules associated with disease progression and treatment response.^11,47

Importantly, pathway-state biomarkers carry greater translational significance by nature, as most therapeutic interventions act on signaling processes rather than isolated molecules. Targeted therapies, including kinase inhibitors, immunomodulators, and pathway-specific biological drugs, are designed to modulate pathway activity rather than molecular abundance. Therefore, biomarkers that measure pathway activation or dysregulation are more likely to predict treatment response, resistance mechanisms, and optimal therapeutic combinations, aligning biomarker discovery with mechanisms of action in precision medicine.^133,134 Figure 4 illustrates the hierarchy of biomarker evidence, demonstrating a progression from purely correlational relationships to therapeutically applicable biomarkers with robust mechanistic validation.

Fig. 4 — Biomarker evidence hierarchy: from correlation to clinical action. A four-level framework showing the progressive strengthening of the mechanistic basis and therapeutic significance. Level 1 (Correlational): statistical association without biological context. Level 2 (Pathway-aligned): consistency with known signaling networks and coordinated pathway organization. Level 3 (Perturbation-validated): functional validation through CRISPR scans, drug perturbations, or genetic models. Level 4 (Therapeutically actionable): directly guiding treatment selection and therapeutic response prediction. Arrows show the gradual increase in strength of evidence (left) and clinical impact (right). Higher-level biomarkers require more extensive validation but provide stronger therapeutic rationale and improved translational probability. Examples are given for each level. Figure created by the authors using Adobe Illustrator (Version 29.3.1)

Signal transduction complexity and biomarker instability

One of the fundamental challenges in biomarker development is that signal transduction networks are characterized by redundancy, feedback regulation, and context-dependent crosstalk; all of which can destabilize biomarker performance across populations and environments. These characteristics help explain why many biomarkers that perform well in discovery cohorts fail during external validation or clinical practice.^5,135

First, redundancy in signaling networks allows multiple pathways to produce similar phenotypic outcomes. For example, tumor proliferation can be driven by MAPK or PI3K-AKT signaling, and inhibition of one pathway can lead to compensatory activation of alternative pathways. This redundancy contributes to therapeutic resistance and limits the generalizability of biomarkers targeting single molecular traits.

Secondly, both negative and positive feedback loops can dynamically reshape pathway activity over time, especially under therapeutic pressure. Targeted inhibition of a signaling node may initially suppress downstream signaling but subsequently trigger feedback activation in upstream or parallel pathways. These adaptive responses reduce the stability of static biomarkers and make treatment prediction difficult, particularly in oncology and immune-mediated diseases.^98,136

Thirdly, context-dependent cross-talks between pathways create variability that is difficult to capture using unimodal or linear models. In inflammatory and thrombo-inflammatory states, signaling interactions between immune cells, endothelial cells, and circulating mediators create highly context-specific biological states. For example, NET formation demonstrates how cross-network interactions generate emerging phenotypes that cannot be reduced to a single marker by integrating signals from inflammation, coagulation, and endothelial pathways.^137,138

These features collectively imply that biomarker robustness depends not only on statistical performance but also on alignment with the underlying network structure of disease biology. AI models that ignore these features risk failing to learn dataset-specific correlations under distributional shifts; this is a well-known limitation in biomedical machine learning.^139,140 In contrast, approaches that incorporate network constraints, pathway priorities, or learning based on perturbation information are better positioned to identify features that remain stable across biological contexts.

From a translational perspective, recognizing signal complexity clarifies why biologically grounding improves generalizability. Biomarkers that reflect pathway-level processes rather than isolated molecular relationships are more likely to retain validity across populations, disease stages, and therapeutic interventions. This understanding motivates the shift from feature-centric biomarker design to network-aware biomarker design, where AI is used not only to detect patterns but also to map those patterns into biologically interpretable signaling processes and actionable therapeutic targets.

AI approaches for pathway-state inference and network-constrained biomarker modeling

A major limitation of traditional biomarker discovery is that individual molecular measurements often provide only indirect information about the functional state of biological systems.^11,141 Consequently, there is growing interest in pathway-state biomarkers that aim to capture the coordinated activity of signaling networks rather than isolated molecular features. In this context, AI primarily serves as an integrative framework for inferring latent biological states from heterogeneous data, rather than as an end in itself.

Pathway-state inference typically combines information from transcriptomic, proteomic, imaging, and clinical datasets to predict the activity of biological programs that cannot be directly measured. Recent multimodal interpretable deep learning frameworks have further demonstrated the ability to integrate transcriptomic and clinical information, while also providing insights into drug response mechanisms.¹⁴² Instead of focusing on single genes or proteins, these approaches identify coordinated patterns reflecting pathway activation, cellular interactions, or disorganization at the network level. Such representations are generally more robust than individual biomarkers because they capture collective biological behavior and are less sensitive to measurement variability.^143,144

The biological significance of pathway-state biomarkers is further enhanced by the fact that the inferred activity patterns remain consistent across multiple data types, independent cohorts, and experimental disruptions. Evidence from systems biology and network medicine shows that biomarkers consistent with coherent biological programs are more likely to remain stable across populations and disease contexts than signatures derived solely from statistical association.^145,146 Consequently, pathway-level inference provides a conceptual bridge between predictive modeling and mechanistic interpretation.

However, inferred pathway states should be considered probabilistic representations of biological activity rather than direct measures of causal mechanisms. Their reliability depends on the quality of the underlying data, the completeness of the available biological information, and the extent of external and experimental validation. Therefore, pathway-state biomarkers are most valuable when integrated into a broader validation framework that includes mechanistic consistency, evidence of disruption, and clinical utility assessment.^130,147

Various methodological approaches, including network constrained learning, graph-based modeling, perturbation-informed analysis, and multimodal integration, can contribute to pathway state inference.^129,148,149 Recently developed graph-based architectures, including graphical convolutional frameworks developed for predicting immune checkpoint inhibitor response, have successfully integrated biological relationships into treatment response prediction tasks.¹⁵⁰ However, the specific algorithm is often less important than the biological validity of the inferred pathway activity and its reproducibility across datasets and experimental settings. From a translational perspective, the main goal is not the selection of a specific AI architecture, but the identification of biologically consistent and clinically relevant pathway states that can support therapeutic decision-making.

Despite these advances, several limitations remain. Biological knowledge bases are incomplete and often context-dependent, which can lead network-constrained models to miss novel mechanisms while guiding them toward well-characterized pathways.^112,113 While perturbation datasets are informative, they are often sparse and may not fully capture in vivo complexity.^128,129 Furthermore, multimodal integration requires careful harmonization of data types with varying resolutions, aggregate effects, and noise structures. Consequently, pathway-state inference should be interpreted as a probabilistic approximation of biological activity rather than a precise representation of causal mechanisms.

However, the integration of network constraints, perturbation data, and multimodal learning represents a significant step toward mechanistically grounded biomarker modeling. By shifting from feature-centric prediction to pathway-level inference, these approaches enhance the stability, interpretability, and translational significance of AI-derived biomarkers and provide a foundation for linking biomarker signatures to therapeutic targeting strategies.^112,145

Neutrophil extracellular traps (NETs) as a pathway-state biomarker exemplar

Neutrophil extracellular traps (NETs) provide an illustrative example of how pathway-state biomarkers can capture dynamic biological processes more effectively than isolated molecular measurements. NET formation is regulated through coordinated signaling pathways involving reactive oxygen species production, PAD4 activation, chromatin decondensation, inflammatory cytokine signaling, and interactions between innate immunity and coagulation networks.^50,95,151 Consequently, biomarkers reflecting NET activity can provide information about the functional state of thrombo-inflammatory responses rather than the abundance of any single molecule.

From a biomarker perspective, NET-associated signatures can be derived from multiple data sources, including circulating extracellular DNA, citrullinated histones, myeloperoxidase-DNA complexes, transcriptomic profiles, proteomic datasets, and imaging-based assessments of inflammatory activity.^49,50,151 AI-powered integration of these heterogeneous data types can facilitate the identification of latent pathway states associated with immune activation, vascular damage, tissue remodeling, and disease progression.^89–91 Importantly, these approaches shift the biomarker development process from single-marker measurements to network-level representations of biological activity.

The translational significance of NET-associated biomarkers extends beyond disease detection. Since NET formation is involved in inflammatory, cardiovascular, autoimmune, thrombotic, and oncological processes, pathway-level NET signatures can contribute to mechanistic disease stratification and the identification of biologically distinct patient subgroups. Although significant experimental and clinical validation is needed, NETs demonstrate how AI-powered pathway state biomarkers can link molecular observations to broader disease mechanisms and support the development of biologically based precision medicine strategies.^49,50,151

Biological relevance of pathway-informed biomarkers

The ultimate translational value of pathway-informed biomarkers lies in their ability to guide therapeutic decisions, rather than simply predicting disease presence or prognosis. By capturing the activation states of signaling networks, these biomarkers provide a functional output of disease biology that can be directly linked to drug mechanisms of action, providing biologically meaningful characterization of disease heterogeneity.

In oncology, pathway-level biomarkers have been widely used to match patients with targeted therapies. For example, activation of the PI3K-AKT-mTOR pathway, MAPK signaling, or receptor tyrosine kinase networks can inform the use of pathway-specific inhibitors, while integrated molecular signatures can identify resistance mechanisms and compensatory pathway activation.^84,116 Importantly, these approaches are moving beyond single gene alterations toward network-based patient stratification that better reflects the redundancy and adaptability of signaling systems.

A similar paradigm is emerging in immune-mediated and inflammatory diseases, where pathway activation states, such as NF-κB, JAK-STAT, and cytokine signaling networks, can inform therapeutic targeting with biologics and small molecule inhibitors. For example, biomarkers reflecting interferon signaling or cytokine-driven immune activation have been associated with response to targeted immunomodulatory therapies, supporting their relevance as indicators of underlying immune activation states.^50,152 In cardiovascular and thrombo-inflammatory conditions, integrated pathway signatures such as NET formation and coagulation-inflammation crosstalk can reveal coordinated thrombo-inflammatory activity patterns associated with disease progression.

Linking biomarkers to treatment requires not only biological validity but also a demonstration of clinical benefit. Predictive performance alone is insufficient; pathway-based biomarkers must demonstrate that their use improves decision-making, patient outcomes, or healthcare efficiency. This requires prospective validation, decision curve analysis, and evaluation within clinical workflows, as highlighted in recent translational frameworks.^153,154 In this context, pathway-based biomarkers are particularly valuable because they align diagnostic classification with actionable points of intervention and strengthen the rationale for their adoption in clinical practice.

Despite these advances, several challenges remain. Signaling pathways are highly context-dependent, exhibiting significant variability across tissues, disease stages, and patient populations. Redundancy and feedback mechanisms can lead to adaptive resistance, limiting the effectiveness of therapies guided by single-pathway biomarkers. Furthermore, many pathway-based models rely on data from cell lines or controlled experimental systems that may not fully capture in vivo complexity. Consequently, robust translation requires the integration of experimental validation, real-world data, and iterative model refinement.

Overall, pathway-informed biomarker strategies represent a critical step toward mechanistically grounded precision medicine, where diagnostic signals are directly linked to therapeutic hypotheses and intervention strategies. By combining network-based modeling, perturbation-informed learning, and clinical validation, AI-powered biomarker frameworks can facilitate a shift from descriptive prediction to actionable biological insight, ultimately enabling more effective and individualized treatment approaches across disease domains.

Although pathway-based biomarkers provide biologically meaningful representations of disease activity, their practical value ultimately depends on their successful integration into biomarker development workflows, clinical validation frameworks, and decision support systems. The operational and translational requirements necessary to facilitate this transition are discussed in the following sections.

Translational workflow for AI biomarker development

Unlike Section “Artificial intelligence methods for biomarker discovery”, which focuses on methodological approaches to biomarker discovery, this section describes the operational workflow required to translate AI-derived biomarkers from the discovery to clinical implementation. The workflow encompasses study design, data harmonization, model development, biological validation, clinical utility assessment, reproducibility standards, and deployment infrastructure. The emphasis is not on specific algorithms, but on the sequential processes necessary to generate robust, clinically actionable biomarker evidence.

Study design and dataset generation: prospective and retrospective, endpoints, and spectrum effects

Biomarker processes should begin with a clear definition of the intended clinical task, including screening, diagnosis, prognosis, treatment selection, or response monitoring. Each task requires distinct tolerances for false positives, decision timing, and clinical consequences. Retrospective datasets are valuable for hypothesis generation, but often overestimate performance when spectrum bias, convenience sampling, and endpoint ambiguity present.^155,156

Case-control designs with extreme phenotypes inflate separability and may misrepresent true diagnostic ambiguity. Prognostic biomarkers are particularly susceptible to time-to-event definition, censoring, and treatment confounding. For AI models trained on resting heart rate (RHR)-linked data, temporal congruence between predictors and outcomes is critical to prevent accidental inclusion of future information.^15,157

Prospective cohort design remains the gold standard but is rarely feasible at scale for early discovery. Hybrid designs (retrospective discovery followed by prospective or pragmatic validation) represent a realistic translational pathway, provided that models are locked in before prospective testing and protocol deviations are minimized.

Data quality control and harmonization across platforms and centers

The robustness of AI biomarkers is primarily limited by the consistency of input data. In molecular studies, variations in sample processing, sequencing chemistry, and library preparation lead to batch effects that can dominate the biological signal. In imaging, scanner hardware, acquisition parameters, and reconstruction algorithms generate region-specific signatures. In digital biomarkers, firmware updates and sensor drift lead to non-stationary noise.

While batch-correction and harmonization methods can mitigate these effects, overcorrection risks eliminating disease-related variation. Therefore, harmonization strategies should be modality-specific, validated using technical iterations whenever possible, and implemented with sensitivity analyses demonstrating the stability of biomarker signals among correction methods.^158–161

Large-scale multicenter studies consistently demonstrate that harmonization quality and metadata completeness are the primary determinants of cross-site generalization, often outweighing algorithmic differences.¹⁵⁸ In conclusion, data engineering should be treated as a core scientific component of biomarker pipelines, not as a preprocessing step.

Feature representation, leakage prevention, and evaluation separation

Classical biomarker pipelines rely on predefined features such as gene panels, radiomic measurements, pathway scores, or summary digital metrics. Although these approaches are often interpretable and reproducible, they may fail to capture complex biological patterns. More recent workflows increasingly utilize data-driven feature representations, which can improve prediction performance but also increase the risk of leakage and spurious associations if evaluation procedures are not rigorously separated.

Leakage can occur through repeated patients, shared normalization between training and test sets, feature selection influenced by outcomes, or longitudinal contamination where future measurements affect previous predictions. In multimodal processes, leakage can occur when different modalities share descriptors or when preprocessing processes inadvertently transmit label information.

Best practice requires complete process separation between training and evaluation, including preprocessing, missing data completion, normalization, and feature selection. Nested cross-validation is essential when hyperparameter tuning is extensive. Independent external validation remains the most reliable indicator that leakage is not inflating performance.¹³⁰ Performance degradation after implementation has been documented even in widely adopted clinical prediction systems.¹⁶²

Model development: calibration, uncertainty, and subgroup performance

Beyond discrimination, clinical biomarker models must be well-calibrated:¹⁶³ predicted risks or probabilities should correspond to observed event rates. Incorrect calibration can lead to inappropriate treatment escalation or false reassurance, even if the AUC is high. Recalibration is often necessary when models are ported across populations or care settings.

Quantitative uncertainty is particularly important when biomarker outputs guide high-risk decisions or trial enrollments. Bayesian methods, ensemble variance, and conformal estimation provide complementary approaches to express predictive confidence, but currently, few biomarker studies explicitly report uncertainty.^164,165

Subgroup performance analysis is necessary to identify differential accuracy rates across strata of sex, ancestry, age, disease stage, and comorbidity.¹⁶⁶ Apparent global performance can mask clinically significant modes of failure in specific subpopulations, which can have direct implications for equity and regulatory approval.

Biomarker selection and stability: reproducibility across cohorts and resampling

In discovery settings, thousands of associated features may appear predictive, but only a subset will be stable under perturbation. Feature stability analysis using bootstrapping resampling, cross-cohort replication, and perturbation testing helps distinguish robust biological signals from sampling artifacts.¹⁶⁷

Stability is particularly important when biomarkers are proposed as mechanistic indicators or therapeutic guidelines. Unstable feature sets weaken biological interpretation and render subsequent validation inefficient.¹⁶⁸ Ensemble feature selection, pathway-level aggregation, and consensus modeling can increase robustness at the expense of lower detail.

From a translational perspective, inter-site generalization is more important than intra-cohort performance. Therefore, biomarker pipelines should prioritize inter-site stability rather than maximum intra-agency discrimination.

Biological validation cycle: functional evidence and orthogonal confirmation

Statistical validation does not determine biological relevance. Functional validation is essential for biomarkers that aim to provide information about mechanisms or treatments.¹⁶⁹ Strategies include correlation with gene perturbation, pharmacological modulation, spatial co-positioning with pathology, and orthogonal molecular assays.

Importantly, validation must be bidirectional: AI-generated hypotheses should inform experiments, and experimental results should iteratively improve models. This closed-loop approach reduces the risk of computational signatures remaining disconnected from biological causality.

Functional validation should demonstrate that biomarker-defined states correspond to reproducible biological processes and remain consistent across orthogonal experimental systems.

Clinical validation and benefit: from AUC to decision impact

Even biologically plausible biomarkers can fail clinically if they don’t improve decisions. Decision curve analysis and net benefit frames assess whether biomarker-guided strategies outperform standard care within plausible threshold ranges.²⁵ These methods are particularly important when biomarkers are proposed for risk stratification or treatment enhancement.

Clinical workflow integration should also be considered. Biomarkers requiring complex testing, long processing times, or specialized interpretation may not be adopted even if they are accurate. Human factors, interpretability, and integration into existing laboratory and IT systems strongly influence real-world efficacy.

Prospective clinical evaluation remains the strongest evidence of benefit, but pragmatic trials and embedded application studies may offer scalable alternatives when continuous model updates are expected.¹⁷⁰

Reporting standards, reproducibility, and auditability

Transparent reporting is essential for independent validation and regulatory confidence. TRIPOD-AI specifies minimum reporting requirements for predictive model development and validation, including dataset description, handling of missing data, and evaluation methodology.^130,147 PROBAST-AI provides a structured risk-bias assessment with an emphasis on leakage, confounding factors, and applicability.¹⁴⁷

For imaging biomarkers, modality-specific standards such as CLAIM further specify acquisition and evaluation reporting requirements. Regulatory bodies expect traceability across data, model releases, and performance audits, especially when adaptive or continuously learning systems are proposed.

Historical community-wide benchmarking initiatives, including Microarray Quality Control (MAQC) and subsequent Sequencing Quality Control (SEQC), and the more recent SEQC2 programs, have consistently demonstrated that biomarker signatures derived from high-dimensional datasets are highly sensitive to analytical decisions such as preprocessing, feature selection, and model development.^171–173 Importantly, different analytical workflows applied to the same datasets can produce significantly different biomarker signatures, even while achieving similar prediction performance.¹⁷⁴ These findings highlight the importance of standardized protocols, independent external validation, benchmark reference materials, and transparent reporting practices. The key lesson for AI-based biomarker discovery is that reproducibility across analytical pipelines, datasets, and study populations is just as important as prediction accuracy itself. Similar reproducibility concerns have been demonstrated in microbiome biomarker studies.¹⁷⁵

Data infrastructure and multi-center reproducibility ecosystems

Sustainable biomarker clinical adoption requires an infrastructure that supports external validation and lifecycle monitoring. Federated learning, secure data enclaves, and trusted research environments enable cross-center evaluation while preserving privacy. Shared benchmark datasets and challenge platforms can accelerate method comparison, but must reflect realistic clinical variability rather than idealized curated subsets.

Regulatory-grade biomarker development increasingly relies on longitudinal performance monitoring, deviation detection, and recalibration mechanisms. These operational requirements blur the traditional line between exploratory research and application engineering, reinforcing the need for biomarker processes to be designed with the end-use in mind from the outset.^176,177

Levels of evidence and verification: a unified grading framework (cross diseases)

To support a unified translation framework for AI-assisted biomarker development, Fig. 5 summarizes the progressive evidence requirements needed to advance biomarkers from the discovery to clinical implementation. The framework emphasizes internal validation, external replication, robustness testing, clinical benefit assessment, and real-world deployment readiness.

Levels of evidence for AI-generated biomarkers

AI-powered biomarker studies encompass a wide range of evidence maturity; however, published claims often fail to make a clear distinction between exploratory associations and clinically applicable tools. To support consistent interpretation, biomarker evidence can be conceptualized through a series of validation stages reflecting increasing resistance to bias and increasing suitability for real-world application.

At the exploratory level, biomarkers are derived from a single cohort or from aggregated retrospective datasets without independent testing. These studies are essential for hypothesis generation but are highly vulnerable to overfitting, aggregate effect, and cohort-specific confounding factors. Performance predictions at this stage reflect primarily intrinsic separability rather than generalizable disease biology.

External validation requires evaluation in independent cohorts collected in different locations, time periods, or clinical contexts. Successful external replication demonstrates that biomarker signals are not constrained by local technical or demographic characteristics. However, external validation alone does not guarantee clinical significance if the patient spectrum, treatment models, or outcome definitions differ from the intended use cases.¹⁵⁵

Prospective validation evaluates biomarkers in prospective studies where data collection, endpoints, and analysis plans are determined before outcome observation. Prospective designs reduce bias from post-model adjustment and allow for the assessment of operational feasibility, deficiency, and workflow integration. However, prospective accuracy does not automatically imply clinical benefit.^66,178,179

Evidence of clinical benefit addresses whether biomarker-guided decisions improve patient outcomes or resource allocation compared to standard care. Decision curve analysis and net utility frames offer a more direct link between biomarker output and patient impact by explicitly assessing clinical outcomes at decision thresholds.²⁵ Benefit can also be assessed through pragmatic trials or embedded application studies where randomized trials are impractical.

Finally, regulatory and clinical adoption evidence reflect sustained performance in real-world conditions, integration into clinical workflows, and post-market safety and efficacy oversight. For adaptive or continuously learning systems, regulatory frameworks increasingly emphasize lifecycle oversight rather than one-off validation.¹⁷⁰

Importantly, progression through this evidence hierarchy is not guaranteed: many biomarkers demonstrate strong exploratory performance but fail during external or prospective evaluation due to biological heterogeneity, dataset shift, or changes in clinical practice. While validation is generally discussed in general terms, evidence requirements for AI biomarkers change significantly throughout the development phases. To functionalize the concept of real-world implementation readiness, Table 4 presents a unified evidence framework for AI biomarkers that links the study design to the type of clinical claims that can be justified at each validation phase.

Table 4.

Levels of evidence (validation framework) for AI biomarkers

Evidence level	Study design	Typical dataset	What is demonstrated	Common pitfalls	Suitable publication claims
Discovery	Retrospective	Single cohort	Signal association	Overfitting, leakage	Candidate biomarker
External validation	Retrospective multi-center	Independent cohorts	Generalizability	Dataset shift ignoring	Validated biomarker
Prospective validation	Prospective observational	Real-world workflow	Stability in practice	Behavior confounding	Clinical readiness
Clinical Utility	Interventional	Decision-guided trials	Outcome improvement	Poor trial design	Clinical benefit
Regulatory adoption	Regulatory submission	Audited datasets	Safety and effectiveness	Limited post-market data	Approved diagnostic

Open in a new tab

Key references for this table: refs. ^{25,130,147,155,170}

This framework is consistent with emerging reporting and bias assessment standards and provides a practical reference for interpreting claims in the biomarker literature, particularly when evaluating whether models have progressed beyond the discovery phase. Most published AI biomarker studies are limited to early validation phases, contributing to the high loss rate observed during robustness testing and external replication. In conclusion, biomarker translation should be viewed as an incremental evidence-generating process, rather than a one-off model development effort.

Common failure modes in AI biomarker translation

Across biomarker types and disease domains, translation failures cluster around a limited number of recurring issues. Recognizing these failure modes is crucial for designing studies that generate reliable evidence.

Spectrum bias occurs when discovery cohorts include extreme phenotypes instead of clinically ambiguous cases, a phenomenon long recognized in diagnostic test evaluation.^180,181 This inflates apparent discrimination while underestimating real-world diagnostic complexity. In screening and early diagnosis settings, spectrum bias is particularly harmful because patterns of disease prevalence and comorbidity differ sharply from compiled datasets.

Dataset shift occurs when relationships between features and outcomes vary across populations, time periods, or care settings. Shifts can reflect demographic differences, evolving clinical guidelines, changes in treatment models, or technology updates. Models that exploit misleading correlations, such as site identifiers or acquisition artifacts, may perform well in development cohorts but degrade rapidly under deployment.^139,182

Confounding factors related to treatment are common in prognostic and response-predictive biomarkers.¹⁸³ When treatment assignment depends on disease severity or physician assessment, outcome relationships may reflect treatment effects rather than the intrinsic biology of the disease. Without careful causal framing, AI models may learn to predict treatment patterns rather than disease trajectories.

Label noise and endpoint ambiguity are common in datasets derived from EHRs. Diagnoses may be inconsistently coded, outcomes may depend to documentation practices, and surrogate endpoints may not reflect meaningful clinical situations. Models trained on noisy or surrogate labels may learn institutional workflows rather than pathophysiology.¹⁸⁴

Multicenter generalization failures reflect incomplete harmonization, unrecognized protocol differences, and demographic heterogeneity.^4,135,185 Even modest changes in imaging parameters, sample processing, or firmware can destabilize biomarker signals if models are not explicitly trained for robustness. These failure modes are often driven by biological variability, heterogeneity in clinical practice, and data generation processes rather than algorithmic limitations alone. Consequently, improvements in model architecture cannot fully compensate for weaknesses in study design, cohort selection, or validation strategy.

To explain why many AI-based biomarkers fail to progress beyond the discovery or early validation phase, Table 5 summarizes common failure modes in biomarker methods, their corresponding mitigation strategies, and proposed validation practices.

Table 5.

Common failure modes and proposed mitigation strategies in AI-based biomarker discovery

Failure mode	What it looks like	Most affected methods	Why it matters	Practical mitigations	Evidence stage where it must be addressed
Batch effects/site effects	Signature tracking points focus on the center/platform rather than biology	Omics, EVs, Imaging	External validation failure	Harmonized SOPs; batch-aware modeling; leave-site-out exclusion testing	External verification + robustness
Spectrum bias	Trained in “clean” endpoints; fails in mixed reality clinics	All	Inflated AUC; poor calibration	Sequential sampling; spectrum-aware evaluation; subgroup reporting	External + prospective
Tag noise/endpoint ambiguity	Misdiagnosis, inconsistent labels, weak underlying reality	EHR, Imaging, Digital	Unstable biomarkers; low clinical confidence	Adjudication; multi-rater labels; sensitivity analyses	Robustness + prospective
Leakage (temporal or data leakage)	Future-oriented information in training; duplicate across splits	EHR, Omics	Unrealistic performance	Patient-level splitting; strict temporal splits; leakage audits	Discovery → all stages
Treatment-related confounding factors	Biomarkers reflect treatment exposure, not disease	EHR, Omics	Failure in novel application models	Pre-treatment sampling; causal adjustment; target trial emulation	Prospective+ utility
Workflow confounding factors	Model learns care pathways (test intensity)	EHR	Nontransferable biomarkers	Model input constraint; causal checks; multi-site validation	External + prospective
Missingness bias	“Missing” severity or reach is encoded	EHR, Digital	Biased prediction	Model missingness explicitly; multiple imputation; sensitivity tests	Robustness
Dataset shift/drift	Population or measurement changes over time	All	Performance decay, safety issues	Monitoring, recalibration, change control plans	Post-deployment
Lack of biological integrity	Features are not mapped to pathways/cell states	Omics, EVs	Low therapeutic significance	Pathway priors; network models; perturbation/replication tests	Mechanistic validation
Poor portability	Works only in one ethnicity/region	All	Equity harm	Diverse cohorts; subgroup calibration; fairness reporting	External + prospective

Open in a new tab

Key references for this table: refs. ^{140,288–291}

Generating biomarker evidence at the regulatory level

The transition from research-level biomarkers to regulatory-acceptable tools requires a shift in development culture from exploratory modeling to controlled evidence generation. There are several principles that distinguish regulatory-level biomarker processes from academic discovery studies.¹⁸⁶ Evidentiary requirements for biomarker qualification continue to evolve across regulatory jurisdictions.¹⁸⁷ Regulatory agencies are increasingly promoting collaborative frameworks to accelerate biomarker and AI validation pathways.¹⁸⁸

First, pre-specification of protocols is essential. Intended use, endpoints, inclusion criteria, preprocessing processes, and evaluation metrics must be defined before starting the analysis. Pre-specification limits analytical flexibility and supports the interpretability of results by reducing selective reporting. Second, model locking is necessary before prospective or pivotal validation. Continuous tuning during evaluation weakens evidence integrity and obscures true performance. Locked models enable repeatable auditing and meaningful performance comparison between different centers. Third, auditability and traceability must be integrated into data and model processes. Version control of datasets, preprocessing scripts, model parameters, and evaluation code is increasingly expected by regulators and journal reviewers. Audit trials also support post-deployment monitoring and deviation detection. Fourth, change management for adaptive systems must be predefined. For AI systems that are updated over time, regulators emphasize predefined change protocols that specify what types of updates are permitted, how performance will be re-evaluated, and when recertification is required. This total product lifecycle (TPLC) perspective addresses validation continuously, rather than episodically. Considering the TPLC, regulators expect predefined update guidelines, traceability, and post-market performance monitoring plans for AI-powered device software functions (US Food and Drug Administration, 2021; US Food and Drug Administration, 2025).^189,190

Finally, clinical integration studies are necessary to demonstrate that biomarker outputs can be reliably used by clinicians and that workflow, interpretability, and turnaround times are consistent with application realities. Without such evidence, even analytically powerful biomarkers may fail to achieve adoption.

Together, these principles shift biomarker development from retrospective pattern recognition toward structured evidence generation capable of supporting regulatory evaluation, clinical decision-making, and real-world implementation.

To distinguish clinically established applications from promising but still developing approaches, Table 6 summarizes the current evidence status of major AI-powered biomarker strategies in translational use cases.

Table 6.

Representative evidence status of selective AI-enabled biomarker strategies in translational use cases

Biomarker strategy/domain	Typical data modality	Current evidence status	Main translational use	Main limitation	Validation requirement before broad clinical use	Representative references
Autonomous diabetic retinopathy screening	Fundus imaging	Proven	Screening/Triage	Device, population, and workflow transportability	Prospective deployment studies, external validation, workflow integration	²⁸¹
AI-assisted ECG phenotyping (e.g., AF, low EF, diabetes risk)	ECG + clinical metadata	Proven to emerging	Screening, early diagnosis, risk stratification	Calibration drift, prevalence dependence, implementation variability	External validation, calibration, decision curve or workflow benefit analyses	^80,194,292
Imaging radiomics for oncological risk stratification	CT/MR/PET imaging	Emerging	Prognosis, response prediction, trial enrichment	Limited reproducibility, site/scanner variability, weak external validation	Multicenter harmonization, external validation, prospective utility studies	^12,14,77
ctDNA minimal residual disease monitoring	Liquid biopsy	Emerging to proven (context-dependent)	Recurrence detection, treatment monitoring, trial adaptation	Tumor type dependence, assay sensitivity, false positive/negative results, context-of-use constrained	Analytical validation, prospective clinical utility, disease-specific implementation studies	^{58,197,210,211}
Multi-omics pathway-state biomarkers	Genomics/transcriptomics/proteomics/metabolomics	Emerging	Patient stratification, mechanism inference, therapy selection	Batch effects, overfitting, incomplete pathway knowledge	External validation, robustness testing, perturbation/mechanistic support	^11,47,89
Digital pathology/computational histology biomarkers	Whole slide histology	Emerging	Diagnostic support, subtype classification, biomarker discovery	Staining/site shift, annotation dependence, pathologist workflow integration	Multi-site validation, reader-assisted studies, deployment evaluation	^87,104,206
Spatial omics and single-cell AI biomarkers	Single-cell and spatial transcriptomics/proteomics	Emerging	Cellular state mapping, tissue programming, target discovery	Cost, sparsity, technical heterogeneity, limited clinical standardization	Cohort replication, orthogonal biological validation, clinical relevance testing	^36,41,42
Wearable/digital biomarkers	Continuous sensor data, smartphones	Emerging	Monitoring, flare prediction, risk estimation	Signal noise, adherence, device heterogeneity, behavioral dependence	Longitudinal prospective validation, addressing missing data, implementation studies	^79,262,293
Perturbation informed pathway biomarkers	CRISPR, drug degradation, stimulation assays	Emerging	Target prioritization, resistance mapping, therapeutic hypothesis generation	Experimental context mismatch, sparse perturbation coverage	Orthogonal validation, in vivo translation, pathway consistency testing	^116,152,207
Foundational model/simulation-guided biomarkers	Multimodal large-scale data, virtual representations	Speculative	General purpose biomarker discovery, virtual clinical trial support, adaptive modeling	Limited prospective validation, opaque failure modes, uncertain clinical utility	Benchmarking, external validation, prospective decision studies	^101,107,276

Open in a new tab

Proven = supported by prospective validation through regulatory approval and/or clinical practice in defined contexts; Emerging = supported by external validation and increasing translational evidence, but prospective benefit and workflow integration are under evaluation; Speculative = proof-of-concept or early translational stage, requiring significant experimental and clinical validation before routine use

Cross-disease applications (synthesis, not catalog)

Although disease-specific biology shapes biomarker selection and clinical goals, the most informative cross-domain comparison is not simply which biomarkers are used, but which combinations of methods, AI strategies, and biologically grounded frameworks are closest to clinical application. Recurring challenges related to validation, confounding factors, portability, and workflow integration are evident across all domains, supporting the need for unified methodological and regulatory frameworks. To facilitate cross-disease comparison, Table 7 summarizes the dominant biomarker methods, dominant AI approaches, major clinical applications, and key translation barriers across representative disease categories.

Table 7.

Cross-disease comparison of AI biomarker translation

Disease domain	Dominant biomarker modalities	Predominant AI approaches	Principal mechanistic insight	Key clinical applications	Major translation barriers
Cardiovascular	Imaging, plasma proteomics, inflammation	CNNs, ensembles, multimodal fusion	Plaque instability, fibrosis, thrombosis pathways	Risk classification, heart failure phenotyping	Dataset shift, treatment confounding
Oncology	ctDNA, pathology, immune profiling	Transformers, foundation models	Tumor-immune ecosystems, resistance	Treatment selection, monitoring of minimal residual disease	Biological heterogeneity
Neurodegenerative	PET/MRI, plasma p-tau, EVs	CNNs, Bayesian fusion	Glial activation, synaptic loss	Early diagnosis, progression monitoring	Mixed pathology
Autoimmune	scRNA, cytokines, imaging	Graph models, clustering	Immune cell programs, cytokine circuits	Treatment response, exacerbation prediction	Tissue and blood incompatibility
Metabolic	Multi-omics, imaging, digital	Ensembles, representation learning	Lipotoxicity, fibrogenesis	Fibrosis risk, lifestyle response	Behavioral confounding
Infection	Host transcriptomics, EHR time series	Time series DL, sparse ML	Immune dysregulation and hyperinflammation	Worsening prediction	Rapid dataset shift

Open in a new tab

Key references for this table: refs. ^{11,21,213,294}

Despite disease-specific biological differences, recurring challenges regarding validation, confounding factors, and clinical integration are evident across all fields, supporting the need for unified methodological and regulatory frameworks.

Cardiovascular diseases

Cardiovascular biomarker discovery integrates circulating molecular markers, multi-omics profiling (genomics, proteomics, metabolomics), immune-inflammatory phenotyping, quantitative cardiac imaging, and increasingly digital physiological monitoring. Plasma proteomics and metabolomics capture systemic inflammation, metabolic stress, and extracellular matrix remodeling associated with the progression of atherosclerosis and heart failure, while CMR and CCTA provide organ-level phenotypes of myocardial remodeling, plaque composition, and perivascular inflammation.^31,191,192 Imaging-derived Inflammatory biomarkers have also shown prognostic value for residual cardiovascular risk.¹⁹³

Wearable and ECG-based biomarkers further expand this framework by enabling longitudinal detection of arrhythmia, decompensation, and dynamic cardiometabolic risk. While ensemble models and regularized regression remain useful for tabular biomarker panels, deep learning is particularly effective in ECG and cardiovascular imaging, where spatial and temporal structure encodes clinically significant phenotypes.^194,195 Large-scale plasma proteomics studies continue to identify novel pathways associated with the development and progression of heart failure.¹⁹⁶ Multimodal learning is increasingly used to integrate imaging, omics, and clinical data, although performance gains are largely dependent on harmonization across institutions and devices.

Mechanistically, most informative cardiovascular biomarker programs focus on immune-inflammatory activation, endothelial dysfunction, fibrotic remodeling, thrombosis, and metabolic stress. Pathway-level analyses correlate circulating and tissue-derived biomarker with cytokine signaling, extracellular matrix turnover, lipid processing, and mitochondrial dysfunction, all of which contribute to plaque instability and myocardial remodeling.¹³¹

Current clinical use includes ECG-based screening for left ventricular dysfunction and atrial fibrillation, multi-marker risk stratification for adverse cardiovascular outcomes, phenotyping of heart failure subtypes, and longitudinal monitoring via wearable or implantable sensors.^80,194 The main translational barrier is not the lack of a predictive signal, but rather its instability across scanners, institutions, treatment contexts, and real-world populations. Many cardiovascular AI biomarkers perform well in carefully selected cohorts, but experience performance degradation in situations such as dataset drift, treatment confounding, and disparities in device access. For meaningful therapeutic integration, biomarkers need to demonstrate external validity, calibration stability, and utility at the decision level, especially when used for trial enrichment and identifying modifiable inflammatory or fibrotic pathways for targeted intervention.

Oncology (tumor-immune ecosystems, liquid biopsy signatures, response)

Oncology biomarker discovery is inherently multimodal, integrating tumor genomics and transcriptomics, immune microenvironment profiling, circulating tumor DNA (ctDNA), EV content, and quantitative imaging phenotypes. Tissue-based molecular profiling remains fundamental for identifying oncogenic drivers and resistance mechanisms, while liquid biopsy approaches enable non-invasive longitudinal monitoring of tumor burden, clonal evolution, and minimal residual disease (MRD). Accumulation evidence indicates that ctDNA-based MRD detection is emerging as a robust biomarker for early recurrence and treatment response across multiple cancer types, including colorectal, lung, and breast cancers.^{56,58,197,198} Recent systematic reviews, meta-analyses, and comparative assessments of tumor-informed and tumor-agnostic approaches have further strengthened the evidence base for ctDNA-guided MRD assessment.^199–202 Genomic MRD strategies are increasingly being integrated into precision oncology workflows.²⁰³ Tumor-agonistic MRD approaches have also shown promising results in non-small cell lung cancer.²⁰⁴ Individual clinical reports continue to demonstrate the feasibility of MRD monitoring in hereditary cancer syndromes.²⁰⁵

Immune biomarkers, including spatial immune context, T cell receptor diversity, cytokine schedules, and myeloid polarization, are central to predicting immunotherapy response. Whereas imaging and digital pathology provide spatially resolved representations of tumor-immune architecture. EV-based biomarkers offer potential for early detection and resistance monitoring, although test standardization and interlaboratory reproducibility remain significant translational barriers.⁶⁷

AI methods in oncology reflect the multimodal nature of modern cancer biology. Deep learning has demonstrated particular utility in digital pathology and radiological imaging. On the other hand, multimodal architectures integrating molecular, imaging, and clinical information are increasingly supporting prognostic assessment and treatment response prediction.^103,106,206 However, robust clinical implementation depends on external validation, standardization, and prospective demonstration of clinical utility. Multimodal fusion architectures incorporating histological, genomic, and clinical features are continuously improving treatment response and survival prediction by reflecting the hierarchical organization of tumor biology.

Mechanistically, the most informative oncology biomarkers capture tumor-immune ecosystem states rather than isolated tumor-intrinsic changes. Single-cell and spatial analysis reveal coordinated programs of immune exclusion, T-cell depletion, stromal activation, angiogenesis, and metabolic adaptation shaping treatment response. AI-enabled network and perturbation models facilitate mapping of biomarker signatures to pathways governing DNA repair, immune evasion, epigenetic plasticity, and drug tolerance.^207,208 Mechanistic relevance is stronger than relying solely on post-exposure feature attribution when biomarker signals are consistent across molecular, cellular, and tissue scales and align with experimentally validated resistance mechanisms.

Clinically, AI-derived oncology biomarkers support diagnosis, prognosis, treatment selection, and disease monitoring. Digital pathology models can aid tumor classification by extracting molecular subtypes from routine slides and enable prioritization of molecular testing.²⁰⁶ Integrated multimodal signatures improve prognostic classification beyond traditional staging, while immune and pathway-level biomarkers guide the selection of checkpoint inhibitors and targeted therapies. Elevated systemic inflammatory markers have also been shown to be associated with reduced benefit obtained from immune checkpoint blockade.²⁰⁹ ctDNA dynamics enable early detection of relapse and emerging resistance, often preceding radiographic progression, and are increasingly being considered as surrogate endpoints in clinical trials.^136,210,211 However, the regulatory characterization of such endpoints, particularly for MRD-based decision-making, remains.

Despite rapid progress, significant clinical adoption barriers persist. The integration of biomarkers into routine clinical practice remains an ongoing challenge across all specialties.²¹² These include interlaboratory variability in pathology processing and imaging, differences in scanners and staining, the effects of tumor purity on molecular measurements, and the evolution of tumor biology due to treatment, which limits the validity of static biomarkers. Liquid biopsy sensitivity is influenced by tumor burden, vascularization, and anatomical context, which restricts its use for universal early diagnosis. More generally, many AI biomarkers demonstrate strong retrospective performance but fail to prospectively improve clinical outcomes due to treatment confounding, biological heterogeneity, and inadequate integration into clinical workflows. Therefore, reliable clinical practice requires standardized tests, externally validated models, predefined analytical processes, and a clear demonstration of added value compared to existing molecular diagnostic methods. Biomarker-driven strategies should ultimately be compatible with viable therapeutic pathways and evaluated within prospective, decision-driven study designs.

Neurodegenerative disorders (early diagnosis, progression, and multimodal integration)

Neurodegenerative biomarker discovery increasingly integrates liquid biomarkers, neuroimaging phenotypes, genetic risk profiling, EV content, and digital behavioral measurements. Key biomarker modalities include plasma biomarkers, advanced neuroimaging, extracellular vesicle cargo, and digital behavioral measurements. All of which together support earlier diagnosis and long-term disease monitoring.^67,213–217 Digital biomarkers derived from speech, gait, and cognitive testing further extend this framework by enabling high-frequency monitoring of functional decline, albeit with limited cross-platform validation.

AI approaches reflect this multimodal structure. DL models are particularly effective for neuroimaging, capturing spatial patterns of atrophy, connectivity disruption, and ligand distribution across disease stages. They often outperform traditional morphometric methods in classification and progression.²¹⁸

For molecular and clinical data, ensemble and Bayesian approaches enable calibrated integration of plasma, cerebrospinal fluid, genetic, and demographic characteristics. Multimodal fusion models combining imaging and molecular data improve the prediction of transition from mild cognitive impairment to dementia and support disease subtyping; however, generalization is sensitive to scanner protocols, cohort composition, and real-world variability.⁶⁷

Mechanistically, the most informative neurodegenerative biomarkers capture interactions among protein aggregation, synaptic insufficiency, neuroinflammation, vascular dysfunction, and glial activation. Growing evidence supports the central role of neuroinflammation in the progression of neurodegenerative disease.²¹⁹ Single-cell and spatial transcriptomic studies highlight microglial activation programs, astrocytic inflammatory states, and complement-mediated synaptic pruning as factors contributing to regional vulnerability and disease progression.^220,221 AI-assisted network analyses offer a more integrated perspective on disease biology by correlating circulating and imaging-derived biomarkers with pathways governing proteostasis, mitochondrial stress, lipid metabolism, and blood-brain barrier (BBB) integrity. Importantly, biomarkers reflecting neuroinflammatory and vascular processes can capture modifiable disease processes beyond amyloid and tau load, particularly in mixed and vascular dementia phenotypes.

Clinically, AI-powered biomarkers support early diagnosis, prognostic stratification, disease classification and longitudinal monitoring. Plasma and imaging markers enable the identification of preclinical or prodromal disease states, while multimodal models improve differentiation between Alzheimer’s disease, Lewy body disease, frontotemporal dementia, and mixed pathologies. Biomarkers are increasingly used in clinical trials for participant selection, disease staging, and pharmacodynamic assessment. However, characterizing them as surrogate endpoints is insufficient and limits their role in regulatory decision-making processes. One of the biggest obstacles to the widespread adoption of blood-based Alzheimer’s biomarkers remains the practical application challenges.²²² The fundamental clinical adoption barrier is biological and clinical heterogeneity: mixed pathologies, comorbidities, and population diversity complicate biomarker specificity and generalizability. Many AI models trained on select datasets fail to consistently perform in community settings where diagnostic uncertainty and multiple diseases are prevalent. Therefore, reliable clinical practice requires consistent testing and imaging protocols, representative validation cohorts, and demonstration that biomarker-driven strategies improve not only diagnostic accuracy but also clinical decision-making processes.

Autoimmune and inflammatory diseases (immune-state biomarkers, therapeutic stratification)

Autoimmune and inflammatory diseases are characterized by dysregulated immune activation across innate and adaptive compartments; this makes immune cell phenotyping, cytokine profiling, autoantibody repertoires, transcriptomic signatures central biomarker modalities. Single-cell and bulk RNA sequencing captures heterogeneity in interferon signaling, T and B cell activation, and myeloid polarization, revealing disease- and tissue-specific immune programs in conditions such as rheumatoid arthritis, systemic lupus erythematosus, inflammatory bowel disease, and multiple sclerosis.^132,223,224 AI methods are increasingly used to integrate heterogeneous metabolic, imaging, and behavioral data for risk stratification and treatment response assessment.^225,226 Nevertheless, performance often remains sensitive to cohort composition, reference standards, and healthcare system-specific factors. Therefore, effective clinical adoption requires standardized testing, consistent single-cell pipelines, and prospective validation demonstrating that biomarker-guided treatment selection improves outcomes compared to empirical treatment strategies.

Metabolic disorders (multimodal metabolic phenotyping and fibrosis-risk stratification)

Metabolic disorders, including type 2 diabetes and MASLD/MASH (Metabolic Dysfunction-Associated Steatotic Liver Disease) are systemic conditions. These conditions include cardiometabolic comorbidities such as adipose tissue dysfunction, insulin resistance, hepatic steatosis, and fibrogenesis, and therefore multimodal biomarker integration is essential. Informative biomarker frameworks combine routine clinical chemistry, multi-omics (genomics, proteomics, metabolomics), inflammatory and immune markers, non-invasive imaging or fibrosis scores, and increasingly digital physiological and behavioral data. In MASLD/MASH, biomarker development focuses on identifying clinically significant fibrosis risk and treatment-responsive disease states using non-invasive tools supported by omics and imaging.^227–229 Advanced metabolomic signatures have recently improved the identification of patients at risk of progressive MASH.^230,231 Updated disease terminology and diagnostic frameworks continue to reshape biomarker development studies in steatotic liver disease.²³² Machine learning analysis of gut microbiota signatures may further support the diagnosis of MASLD.²³³

Mechanistically, metabolic biomarkers reflect the interaction of adipose tissue inflammation, hepatic lipotoxicity, insulin signaling disruption, mitochondrial stress, and fibrogenesis remodeling, and multi-omics studies reveal coordinated pathway-level changes that capture disease heterogeneity better than single markers.^234,235 Recent multi-omics studies have identified the underlying biological pathways of childhood obesity and metabolic dysfunction.²³⁶ Adipose tissue-microbiome interactions appear to play a central role in defining metabolic obesity phenotypes.²³⁷ Obesity-related inflammation and its resolution pathways continue to emerge as promising biomarker targets.²³⁸

AI approaches, including ensemble models and representative learning, are used to integrate heterogeneous data sources and improve the prediction of metabolic risk, disease progression, and treatment response; however, performance gains are often cohort-dependent and sensitive to reference standard variability.^32,239,240 Digital biomarkers derived from wearable devices, including activity, sleep, and glucose dynamics, extend metabolic phenotyping to longitudinal tracking, but are still affected by adherence, device variability, and behavioral confounding factors. Clinically, AI-derived metabolic biomarkers are used in risk stratification, complication prediction, fibrosis assessment, and treatment monitoring, and are increasingly being investigated for trial enrichment by identifying patients with active target pathways. However, translation is limited by dataset shifts across health systems, treatment and lifestyle confounding factors, variability in reference standards such as liver histology, and poor portability from specific cohorts to larger populations.²⁴¹ Therefore, reliable clinical practice requires standardized processes, external validation across care settings, calibration reporting, and demonstration that biomarker-guided decisions improve outcomes in real-world metabolic care pathways.

Infectious diseases and host-response states (dynamic biomarkers and stress testing of translation)

Infection and sepsis present a challenging test case for biomarker translation due to rapidly evolving biological and clinical conditions, including pathogen variability, treatment timing, and changing care practices. Recent reviews continue to identify significant unmet needs in the discovery and validation of sepsis biomarkers.²⁴² The most informative biomarkers reflect host response states, including cytokine and chemokine profiles, immune cell activation markers, and transcriptomic signatures that capture the balance between hyperinflammation and immunosuppression.^243,244 AI approaches often integrate electronic health record time series with host response data to predict clinical deterioration and course changes rather than static diagnosis. Machine learning approaches have identified additional candidate biomarkers associated with immune dysregulation at various stages of sepsis.²⁴⁵ Mechanistically, these biomarkers capture coordinated immune programs involving interferon signaling, myeloid activation, lymphocyte dysfunction, and endothelial damage, but interpretation is complex due to strong treatment and timing confounding factors. In clinical use, the focus is on early risk stratification, immunophenotype classification, and monitoring of treatment response. However, infectious diseases consistently reveal fundamental failure modes of AI biomarkers, such as rapid dataset drift, label instability, and poor cross-reproducibility. Therefore, robust real-world implementation requires time-sensitive validation, standardized processes, and explicit handling of treatment effects, highlighting that the biomarker utility depends on stability in evolving clinical settings rather than performance on static datasets.

While specific biomarkers and data modalities differ substantially across disease domains, the determinants of successful translation are remarkably consistent. Biomarkers with the greatest clinical impact are those that are biologically plausible, have external validation, are reproducibly measurable, and provide a clear level of benefit at the decision-making level. Conversely, common causes of failure include dataset drift, treatment confusion, biological heterogeneity, and inadequate integration into clinical workflows. These recurring patterns demonstrate that disease-specific innovation alone is insufficient and that progress toward clinical implementation depends on shared methodological and regulatory principles applicable across biomarker domains (Fig. 6).

Fig. 6 — Cross-disease AI biomarker applications. Examples of biomarker-driven applications across cardiovascular, oncology, neurodegenerative, autoimmune, metabolic, and infectious diseases. Despite disease-specific objectives, all applications rely on shared methodological foundations, including multimodal integration, baseline models, external validation, and clinical benefit assessment. Figure created by the authors using Adobe Illustrator (Version 29.3.1). AI artificial intelligence, ECG electrocardiogram, NETs neutrophil extracellular traps, ctDNA circulating tumor DNA, NfL neurofilament light chain, p-tau phosphorylated tau, TRIPOD + AI Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis with AI extension, PROBAST + AI The prediction model Risk Of Bias ASsessment Tool with AI-extended

Translational roles of AI biomarkers in therapeutic development and clinical decision-making processes

Moving from purely predictive markers to biologically based and therapeutically applicable biomarkers requires the explicit integration of mechanistic modeling and validation, as conceptually shown in Fig. 7.

Fig. 7 — Conceptual transition from predictive to mechanistically grounded AI biomarkers. Correlation-based biomarkers provide predictive signatures but are often lack biological interpretability. Mechanistically informed biomarkers may better support therapeutic decision-making and clinical application by integrating biological context, causal reasoning, and pathway-level understanding. Figure created by the authors using Adobe Illustrator (Version 29.3.1)

Biomarkers for study enrichment, adaptive designs, and outcome selection

AI-derived biomarkers are increasingly being used to enrich clinical trials by selecting patient subgroups that are more likely to express targetable disease biology or demonstrate measurable progression within feasible follow-up periods. In oncology, ctDNA dynamics, immune phenotypes, and tumor microenvironment signatures are used to identify biologically responsive subgroups, increase statistical power, and reduce sample size requirements.^246,247 A key example is the TIDE framework, which predicts response to immune checkpoint blockade by integrating signatures of T cell dysfunction and exclusion, demonstrating how biologically informed computational biomarkers can support patient enrichment and treatment stratification.²⁴⁸ Similarly, the IMPRES framework further supports the value of mechanistically informed biomarker signatures in therapeutic decision-making by showing that immune interaction patterns can strongly predict response to checkpoint inhibitor therapy in melanoma cohorts.²⁴⁹

In neurodegenerative and metabolic diseases, enrichment strategies often address the dilution of treatment effects in heterogeneous populations by focusing on selecting individuals with biomarker-validated pathology or rapid progression trajectories.^214,228

Figure 8 illustrates how AI-enabled biomarkers can support enrichment, stratification, longitudinal tracking, and adaptive treatment modification in modern clinical trial designs.

Fig. 8 — Biomarker-guided adaptive clinical trial design: AI-enabled biomarker enrichment is used to identify high-probability responders, stratify enrollments, and guide treatment allocation. Longitudinal biomarker monitoring supports adaptive decisions, including sample size re-estimation, arm modification, dose adjustment, and endpoint refinement. Continuous feedback improves trial efficiency and therapeutic accuracy. Figure created by the authors using Adobe Illustrator (Version 29.3.1)

Adaptive clinical trial designs are increasingly integrating biomarker feedback to modify enrollment criteria, randomization ratios, or treatment arms during trial execution. AI models predicting early response or toxicity can support response adaptive randomization and Bayesian borrowing among subgroups, but operational complexity and regulatory acceptance remain limiting factors.²⁵⁰ Importantly, to avoid bias in treatment effect prediction, biomarker-driven enrichment should avoid selection based on post-randomization variables.

Endpoint selection represents another crucial translational role for AI biomarkers. Imaging-derived radiomic features, molecular response scores, and digital functional endpoints are being considered as intermediate or surrogate endpoints to accelerate drug development. However, regulatory acceptance requires evidence that biomarker modifications reliably predict clinical benefit rather than merely reflecting pharmacodynamic activity. While predictive biomarkers can improve enrichment and risk stratification, mechanistically grounded biomarkers are more likely to provide information about drug targeting and adaptive therapy strategies. Table 8 compares these paradigms in terms of interpretability, validation requirements, and therapeutic benefit.

Table 8.

Predictive and mechanistically grounded AI biomarkers: Translative implications

Dimension	Predictive biomarkers	Mechanistically grounded biomarkers
Primary goal	Outcome prediction	Biological program/pathway inference
Typical outputs	Risk score, classifier, latent signature	Pathway activity, cell state program, network module
Dependence type	Correlational association	Causal/structural relationships (explicit or restricted)
Interpretability	Often post hoc	Direct biological interpretability (when validated)
Stability across therapies	Often unstable	More robust to treatment perturbations (if causal)
Trial enrichment	Common, pragmatic	Especially strong for especially mechanism-matched therapies
Treatment selection	Limited	Directly informative (companion/complementary role)
Potential surrogate endpoints	Often weak	Higher if tied to causal pathway
Emphasis on validation	Discrimination/calibration	Biological coherence + perturbation evidence + clinical utility
Key risk	Spurious association leakage	Over-claiming mechanism without experimental support
Best use case	Screening/prognosis	Drug development, stratification, adaptive monitoring

Open in a new tab

Key references for this table: refs. ^{4,112,268,289}

Accompanying diagnosis: linking AI biomarkers to treatment selection

Beyond trial enrichment, AI biomarkers are increasingly functioning as complementary or accompanying diagnostic tools, linking patients to targeted therapies or combination strategies. In oncology, predictive biomarkers derived from tumor genomics, immune context, and ctDNA dynamics inform the selection of targeted agents, immunotherapies, and rational combinations.^{248,249,251,252} AI models integrating these layers can identify resistance mechanisms and suggest adaptive treatment strategies during therapy. Recently, graph-based and multimodal learning approaches have further improved treatment response prediction by integrating molecular features with biological relationships and clinical context, thereby enhancing both prediction performance and biological interpretability.^142,224

In autoimmune and inflammatory diseases, immune status biomarkers can guide the selection between biological drugs and small molecule inhibitors by identifying dominant cytokine or cell type pathways, potentially reducing the empirical treatment cycle.²²⁵ In cardiovascular and metabolic diseases, phenotypic clusters arising from inflammatory, fibrotic, or metabolic signatures may shed light on the use of anti-inflammatory, antifibrotic, or metabolic modulating therapies, but clinical implementation is still in its early stages.

Dose optimization and treatment sequencing offer additional opportunities for biomarker-guided therapy.²⁵³ Longitudinal biomarker trajectories may reveal inadequate target interaction or the risk of early relapse and may support proactive dose adjustment or treatment escalation. However, most AI-assisted treatment selection frameworks are observational and require prospective validation to demonstrate clinical benefit.

Biomarkers as surrogate endpoints: when do they succeed and why do they fail?

The use of biomarkers as surrogate endpoints remains controversial, particularly in chronic and heterogeneous diseases. Successful surrogate endpoints should lie on the causal pathway between intervention and clinical outcome and capture the dominant mechanism of therapeutic effect.^115,254,255 Classic examples include viral load suppression in HIV and blood pressure reduction in cardiovascular disease, where strong causal links exist between biomarker change and clinical events.

Historical experience in therapeutic fields shows that only a small subset of biomarkers reliably functions as surrogate endpoints. Table 9 summarizes representative examples, highlighting the biological conditions under which surrogate strategies succeed or fail.

Table 9.

When do biomarkers succeed or fail as surrogate endpoints

Disease era	Biomarker surrogate	Biological rationale	Outcome correlation	Trial success?	Reasons for success/failure.
HIV	Viral load	Direct causal driver	Strong	Yes	Capture mechanism
Hypertension	Blood pressure	Mediator of vascular damage	Strong	Yes	Linear causal pathway
Oncology	Tumor shrinkage	Tumor burden proxy	Variable	Mixed	Resistance pathways
Alzheimer’s diseases	Amyloid load	Plaque hypothesis	Weak	Mostly no	Downstream neurodegeneration
MASH	Fibrosis scores	Predicts mortality rate	Moderate	Emerging	Slow biological response

Open in a new tab

Key references for this table: refs. ^{115,254,255,295,296}

These examples demonstrate that biomarker performance as a surrogate endpoint depends primarily on biological causality rather than measurement precision. Biomarkers that directly capture dominant disease mechanisms are more likely to predict clinical benefit. Otherwise, markers that reflect only partial or downstream aspects of disease biology often fail despite strong statistical associations. AI can improve biomarker sensitivity, integration, and temporal resolution, but it cannot overcome the weak causal relationships between biomarker modulation and patient-relevant outcomes. Therefore, AI-derived surrogate biomarkers should be evaluated within causal frameworks and validated across interventions operating through diverse biological mechanisms.

Longitudinal monitoring and dynamic treatment adaptation

AI-enabled longitudinal biomarker monitoring enables dynamic assessment of disease course, treatment response, and relapse risk. Time series models integrating molecular, imaging, and digital biomarkers can detect early deviations from expected response patterns and potentially enable preventive treatment modification.¹⁵ In oncology, serial ctDNA monitoring can identify molecular relapse before radiographic progression, while in metabolic diseases, continuous glucose and activity monitoring support adaptive lifestyle and pharmacological interventions. Longitudinal validation is still an underdeveloped area. Similar challenges have recently been highlighted in the validation of biomarkers of biological aging.²⁵⁶

Longitudinal monitoring, especially when digital biomarkers are included, also carries risks of overdiagnosis, alert fatigue, and behavioral confounding. Therefore, effective clinical practice requires careful calibration of alert thresholds, integration with clinical workflows, and continuous performance monitoring.

Real-world evidence (RWE) and post-implementation monitoring

After transitioning to clinical practice, AI biomarkers must be evaluated under real-world conditions characterized by heterogeneous populations, evolving clinical practices, and variable data quality. Real-world evidence (RWE) studies can assess whether biomarker-guided decisions improve outcomes, reduce costs, or change treatment models compared to standard care.^257–259

Post-application monitoring should evaluate not only prediction performance but also clinical impact, adverse events, and equity among subpopulations. Model drift, test changes, and shifts in disease epidemiology can degrade performance over time and may require recalibration or retraining strategies under regulatory oversight.

More importantly, RWE can contribute to the improvement of biomarker thresholds, decision rules, and patient selection criteria by revealing discrepancies between controlled clinical trial performance and real-world efficacy. Continuously learning health systems offer a framework for iterative improvement but require governance structures that balance innovation with patient safety and transparency.

Taken collectively, these applications demonstrate the transition of AI biomarkers from discovery tools to clinical decision support instruments. Their greatest value emerges when biological interpretation, rigorous validation, therapeutic suitability, and real-world performance are considered together, rather than as separate objectives. Ultimately, successful implementation will depend not only on predictive accuracy but also on demonstrable improvements in patient outcomes, treatment selection, and healthcare delivery.

Regulatory, ethical, and implementation considerations for AI-based biomarkers (biomarker-focused; algorithm + analysis combined system)

Biomarker characterization and algorithm validation: crossroads and interaction

Since regulatory requirements depend on intended use, clinical risk, and model adaptability, Fig. 9 summarizes a simplified decision tree regarding the FDA and EMA classification pathways applicable to AI-based biomarker software and devices.

Regulatory evaluation of AI-based biomarkers requires characterization of both the biological signal and the algorithmic transformation applied to that signal.^187,260 Unlike traditional biomarkers, AI biomarkers can integrate hundreds to thousands of features across various methods, complicating biological interpretability and analytical validation. Therefore, regulators are increasingly emphasizing the need to define intended use, biological plausibility, and pathway relevance, especially when biomarkers guide therapeutic decisions or serve as trial enrichment tools. Algorithm validation must demonstrate stability across clinically relevant subgroups, testing platforms, and disease spectra. When biomarkers integrate multi-omics or imaging data, analytical variability arising from upstream testing propagates into model uncertainty, requiring joint validation of laboratory procedures and algorithm performance. Interaction effects between biomarkers and treatments (e.g., predictive and prognostic roles) should also be clearly characterized to prevent misapplication of biomarkers beyond their validated contexts.

Biomarker characterization also requires a clear definition of the Context of Use (COU), including the purpose for which the biomarker will be applied, the patient population, and the clinical decision-making context. Regulatory pathways for biomarkers developed as drug development tools rather than diagnostic devices differ from the SaMD or IVD/CDx frameworks; in this case, the FDA Biomarker Qualification Program (BQP) provides a mechanism for formally qualifying biomarkers for specific use contexts in clinical trials and therapeutic development.²⁶¹

SaMD and AI governance for biomarker tools: locked versus adaptive models

Many AI-enabled biomarker systems are classified as SaMD when used for clinical decision-making purposes. Since many AI biomarker tools fit the Software as a Medical Device (SaMD) definition, regulatory expectations are increasingly emphasizing lifecycle management, transparency of intended use, and structured oversight of model updates.¹⁸⁹ Regulatory frameworks distinguish between locked algorithms (which remain unchanged after deployment) and adaptive or continuously learning systems (which can be updated based on new data). While adaptive models offer potential performance improvements, they also present challenges related to version control, traceability, and post-market surveillance.^186,189 For adaptive biomarker algorithms, the FDA has proposed the concept of a PCCP to define permitted updates, relevant validation/verification evidence, and post-update monitoring expectations.¹⁹⁰ Since AI-powered biomarkers are implemented through different regulatory and translational pathways depending on their intended use, Table 10 summarizes the main categories related to clinical implementation, the relevant evidence requirements, and representative application contexts.

Table 10.

Regulatory and translational pathways for AI-powered biomarker tools

Translational category	Typical intended use	Regulatory/translational pathway	Main evidence expectations	Example context	Representative references
Software as a Medical Device (SaMD)	Risk prediction, triage, assistive clinical decision support	device-software pathway with lifecycle oversight	Technical verification, clinical validation, workflow integration, post-deployment monitoring	AI-powered ECG, stroke triage, autonomous screening systems	^186,260,297
In vitro diagnostics/companion diagnostics (IVD/CDx)	Patient selection for treatment, assay-related treatment decisions	Assay-specific diagnostic pathway; often linked-targeted therapy approval	Analytical validation, clinical validation, intended-use definition, context-specific performance	PD-L1 type selection paradigms, biomarker-linked oncology treatment decisions	^134,253
Biomarker qualification/drug development tool	Trial enrichment, pharmacodynamic biomarkers, surrogate development	Context-of-use qualification framework	Reproducibility, context-of-use definition, evidentiary package, trial-level utility	Enriching biomarkers in targeted therapy development	^115,255,295
Research use only (RUO) biomarker models	Discovery-stage signatures, exploratory mechanism studies	Preclinical/exploratory pathway	Internal rigor, transparent methods, replication where possible, no clinical claims	Early multi-omics discovery models	^5,85,178
Adaptive/ continuously updated AI biomarker systems	Recalibrated or learning systems for longitudinal prediction or support	Lifecycle governance with update control and monitoring	Change management, traceability, predefined update rules, post-market oversight	Adaptive clinical decision support tools	^{260,298–300}

Open in a new tab

For biomarker applications, regulatory expectations generally favor locked models for diagnostic and therapeutic decision support, with predefined retraining and recertification pathways for updates.¹⁸⁷ While continuous learning may be more feasible in low-risk monitoring contexts, reconciling it with traditional validation paradigms remains challenging when biomarker thresholds directly impact patient management. Navigating the regulatory landscape for AI-based biomarkers requires systematic assessment of the intended use, clinical risk level, and information significance.²⁶²

Data privacy, fairness, consent, and population equity in multimodal biomarkers

AI biomarker development often relies on large-scale, multimodal datasets derived from vulnerable patient populations.²⁶³ Genomic and imaging data increase the risks of re-identification, while wearable and behavioral data raise concerns about ongoing surveillance. Therefore, ethical practice requires robust consent frameworks, transparency regarding secondary data use, and safeguards against misuse beyond clinical intent. Fairness is a critical regulatory and ethical issue, as biomarker performance can differ significantly across ancestry groups, socioeconomic strata, and contexts of access to healthcare. Failure to assess and correct such disparities risks exacerbating health inequalities, particularly when biomarkers determine eligibility for expensive targeted therapies. Multiple ancestry polygenic approaches can partially reduce ancestry differences in biomarker performance.²⁶⁴ Regulators expect subgroup performance reporting and mitigation strategies when clinically significant disparities are observed.²⁶⁵

Clinical workflow integration, interpretability thresholds, and trust

Even analytically validated biomarkers can fail clinically if they cannot be integrated into routine workflows. Practical application requires compatibility with laboratory information systems, imaging platforms, and electronic health records, as well as clearly defined clinical action thresholds. For AI-enabled biomarker systems that generate continuous risk scores, translating them into actionable categories remains one of the biggest barriers to adoption. Interpretability requirements are context-dependent: mechanistic transparency may not be necessary for low-risk screening tools, while therapeutic decision support often requires the description of contributing features or biological pathways.¹⁰⁸ Importantly, over-reliance on post-annotations lacking a biological basis can create false reassurance instead of true understanding. Clinician trust is shaped not only by model performance but also by perceived alignment with standards of clinical reasoning and evidence.

Accountability and responsibility in biomarker-driven decisions

AI biomarkers complicate traditional medical accountability concepts by involving algorithm developers, data providers, and healthcare institutions in decision-making paths historically driven by clinician judgment. When biomarker outputs influence diagnosis, treatment selection, or clinical trial appropriateness, responsibility for errors is distributed among technical and clinical actors.

Regulatory and legal frameworks are increasingly emphasizing common accountability models that require documentation of intended use, decision support limitations, and clinician override mechanisms. Ethically, clinicians should retain ultimate responsibility for patient care decisions, while institutions must ensure that the biomarker systems used meet comparable standards of evidence and safety to other diagnostic technologies.

Effective use of AI biomarkers requires coordinated validation of analytical tests, algorithms, and clinical workflows. Table 11 summarizes regulatory and implementation requirements throughout the development phases and highlights common failure points that prevent real-world deployment.

Table 11.

Regulatory and implementation requirements for AI biomarkers

Requirements	What needs to be validated	Who is responsible	Typical evidence	Common failure points
Analytical validity	Analysis + preprocessing	Laboratory + vendor	Accuracy, reproducibility	Batch effects
Clinical validity	Biomarker-outcome correlation	Developers	External cohorts	Spectrum bias
Clinical utility	Decision benefit	Trial Sponsors	Interventional trials	No impact on workflow
Algorithm governance	Model stability	Manufacturers	Version control audits	Drift post-distribution
Post-marketing surveillance	Safety and bias	Institutions	Real-world monitoring	Recalibration neglected

Open in a new tab

Key references for this table: refs. ^{170,189,298,301,302}

Meeting these requirements is vital not only for regulatory approval but also for sustainable clinical reliability, particularly in environments where adaptive models and evolving data distributions are expected.

Challenges and debates in AI biomarker discovery and translation

Spurious biomarker phenomenon: confounding effects, leakage, and batch effects

A significant limitation in AI-powered biomarker discovery is the frequent identification of signals in discovery datasets that correlate with outcomes but do not reflect true disease biology. Such spurious biomarkers arise from technical batch effects, confounding effects stemming from treatment or care pathways, and information leakage across training and testing sets.²⁶⁶ In multi-omics studies, differences in sample processing, sequencing platforms, or laboratory protocols can dominate biological variation, leading to signatures that reflect study design rather than disease mechanisms.¹⁵⁹

Clinical confounding effects are particularly problematic when biomarkers represent healthcare utilization, disease severity at presentation, or therapeutic interventions rather than intrinsic disease processes. For example, laboratory tests requested only in critically ill patients may appear predictive simply because they encode clinician behavior. Without careful causal framing and data source analysis, AI models can inadvertently learn such non-biological relationships and produce biomarkers that collapse under external validation.

Interpretability discussions: post hoc explanations versus mechanistic evidence

Interpretability remains one of the most controversial aspects of AI biomarker research. Post hoc explanation techniques, such as feature attribution maps, SHAP values, and saliency methods, aim to identify effective inputs but do not establish causal relationships or biological relevance.²⁶⁷ In biomarker discovery, such explanations can give a false impression of biological insight by highlighting associated but mechanistically irrelevant features.

Mechanistic interpretability requires alignment between learned representations and known biological pathways, cellular states, or molecular interactions. Approaches involving pathway prioritization, graph-based biological networks, or causal discovery frameworks offer greater potential for biologically grounded inference but remain computationally and data-intensive.

Generalizability and population bias: multi-ethnic and multi-regional performance gaps

Many AI-enabled biomarker studies rely on datasets that incompletely represent minority populations and low-resource healthcare settings, raising concerns about unequal performance across demographic groups. Differences in genetic origin, comorbidities, environmental exposures, and access to healthcare can significantly alter biomarker distributions, weakening model calibration and clinical reliability.^185,268

Therefore, external validation across geographic regions, healthcare systems, and origin groups is essential but is often neglected due to data access barriers and logistical constraints. When performance disparities are identified, mitigation strategies such as reweighting, subgroup-specific thresholds, or transfer learning can partially address disparities but cannot completely replace inclusive data collection strategies. Failure to address these issues risks embedding structural disparities into diagnostic and treatment pathways.

Reproducibility and evidentiary standards

Reproducibility remains one of the most significant unsolved challenges in AI-powered biomarker research. Biomarkers that perform well in discovery datasets often show declines in accuracy, calibration, or clinical utility when evaluated across independent cohorts, health systems, or measurement platforms. Large collaborative consortia, benchmark datasets, and multicenter validation initiatives offer valuable opportunities to assess robustness, but they often fail to fully reflect the diversity and complexity of real-world clinical settings.²⁶⁹

An ongoing debate concerns the level of evidence required for AI-derived biomarkers to support clinical decision-making. Predictive performance alone is increasingly seen as insufficient. Instead, reproducibility should encompass stability across populations, consistency across technical platforms, and the preservation of clinical utility in independent settings. These challenges highlight the need for evidence frameworks that balance methodological rigor with practical clinical application and acknowledge that acceptable thresholds of evidence may differ depending on the intended biomarker application.

Emerging trends and future directions in AI biomarker discovery

Foundation and self-supervised learning for biomarkers: promises and exaggerations

Foundation models trained on large-scale unlabeled biomedical data offer a pathway for transferable representations across diseases, modalities, and institutions. In imaging, self-supervised pretraining improves performance in low-label regimens and enhances the ability to generalize across scanners and populations. In molecular biology, protein, transcriptomic, and multimodal foundation models capture structural and regulatory patterns adaptable to biomarker discovery tasks.^101,107 Precision oncology represents one of the clearest examples of this convergence between biomarkers, machine learning, and clinical decision support.²⁷⁰ Future integration of electronic health records with post-genome datasets could further accelerate precision medicine initiatives.²⁷¹

However, foundation models inherently do not address issues of cohort bias, confounding factors, or clinical utility. Without careful fine-tuning and external validation, pre-trained representations may encode large-scale correlations rather than disease-related mechanisms. Furthermore, the computational and environmental costs of training such models raise questions about sustainability and accessibility. Therefore, future assessment frameworks should evaluate not only predictive gains but also robustness, interpretability, and clinical impact.

Multiscale systems medicine: linking molecules, cells, organs, and phenotypes

A significant frontier in biomarker science is the integration of molecular, cellular, tissue-level, and organism-level phenotypes into coherent disease models. Advances in single-cell and spatial technologies enable mapping of cellular state programs in anatomical and microenvironmental contexts, while imaging and digital biomarkers capture organ-level and functional manifestations.^272,273 Single-cell technologies are expected to play an increasingly important role in therapeutic target discovery and drug development.²⁷⁴ Interpretable frameworks for integrating single-cell and spatial omics data are likely to become increasingly important in future biomarker processes.^45,273

AI architectures with hierarchical and graph-based reasoning capabilities offer tools to connect these scales, making it possible to infer how molecular perturbations propagate through cellular networks and produce clinical phenotypes. Such multiscale models can support the identification of causal pathways, cross-tissue interactions, and emerging disease programs that are not visible in unimodal analyses. However, bringing together sufficiently large, well-explained multiscale datasets remains a significant bottleneck, and experimental validation of the inferred mechanisms will remain essential.

Joint learning and privacy-preserving analytics for multicenter validation

Data sharing barriers remain a fundamental obstacle to robust biomarker validation. Federated learning, secure multi-party computation, and differential privacy frameworks enable collaborative model development without centralizing sensitive patient data, thereby supporting multicenter validation while respecting regulatory and ethical constraints. By exposing models to more diverse populations and health systems, these approaches can increase generalizability and reduce center-specific biases.

However, distributed learning frameworks introduce additional challenges, including heterogeneous data distributions, variable local data quality, and evolving regulatory requirements for certification and auditing.²⁷⁵ Future efforts should focus on establishing practical standards for governance, validation, and interoperability across participating institutions.

Digital twins and in silico trials: realistic short-term use cases

Digital twins, computational representations of individual patients or disease trajectories, have been proposed as tools to simulate treatment response and optimize trial design.²⁷⁶ In the near term, realistic applications are likely to focus on population-level or cohort-level simulations rather than fully individualized physiological replicates.

In silico trials can support hypothesis generation, dose-finding studies, and the exploration of trial design alternatives, potentially reducing costs and accelerating development cycles.²⁷⁷ However, digital twins are critically dependent on the validity of the underlying mechanistic and statistical models; without strong empirical grounding, simulations risk reinforcing existing assumptions rather than revealing new insights. Therefore, digital twin approaches should complement rather than replace experimental validation strategies, serving as hypothesis-generating and trial-optimization tools while maintaining reliance on prospective biological and clinical validation.

A pragmatic roadmap for the next 5 years

Over the next 5 years, progress in the implementation of AI-enabled biomarker systems will depend less on algorithmic innovation and more on infrastructure, standards, and collaborative frameworks. Key priorities include⁹²: (1) data foundations (large, diverse, longitudinal, and multimodal cohorts with standardized acquisition and annotation); (2) methodological rigor (routine external validation, subgroup performance reporting, and decision curve-based assessment of clinical benefit); (3) biologically grounded integration (embedding biological information into model architectures and validation strategies to support causal inference); (4) regulatory alignment (early engagement with regulators to define acceptable evidence pathways for AI biomarkers); (5) implementation science (prospective studies evaluating how biomarker-guided decisions impact workflows, outcomes, and equity).

Therefore, investment strategies should prioritize interoperable data ecosystems, cross-industry consortia, and reproducibility infrastructure, rather than isolated model development efforts. Success will be measured not by benchmark performance, but by demonstrable improvements in patient outcomes and treatment development efficiency.

Conclusion and strategic perspective

Across disease domains and biomarker modalities, several consistent principles distinguish robust, translatable AI-powered biomarkers from that remain confined to discovery settings. First, biological validity and mechanistic grounding are central when biomarkers are intended to provide information on therapeutic development, patient stratification, or treatment selection. Biomarkers reflecting causal disease processes are more likely to generalize across populations and remain stable under clinical intervention than those derived solely from correlational signatures. Second, multimodal integration offers clear advantages when pathological processes span molecular, cellular, tissue, and organ-level scales. However, integration should be driven by biological hypotheses and empirical probability, rather than the indiscriminate aggregation of heterogeneous features, risking increasing noise and confusion instead of revealing disease mechanisms. Third, external validation across independent cohorts and clinically relevant subgroups is indispensable, as predicted performance in discovery datasets routinely overestimates real-world efficacy. Finally, success should be defined not only by prediction accuracy but also by demonstrable clinical benefit. Evaluation frameworks incorporating decision analysis metrics, workflow integration, and patient-relevant outcomes are necessary to differentiate biomarkers that meaningfully improve care from those that merely improve risk stratification without altering management.

Evidence across multiple disease domains indicates that translational success depends as much on study design, validation strategy, and clinical integration as on algorithmic performance. Biomarker development studies are most effective when aligned with explicitly stated diagnostic or therapeutic decisions and when model building incorporates biological information supporting reasonable biologically grounded interpretation. Evaluation strategies should extend beyond retrospective accuracy estimations to include prospective assessments of biomarker-guided decision paths, particularly in settings where treatment selection, dosing strategies, or trial enrichment are intended applications. On the other hand, common pitfalls include assuming statistical association has therapeutic significance, relying solely on endogenous cross-validation, and neglecting population diversity and health system variability during development. Similarly, while interpretability techniques are valuable for transparency, they cannot replace biological validation or experimental verification of disease mechanisms. These considerations are equally valid for molecular, imaging, and digital biomarkers, highlighting that translation barriers are often methodological and organizational rather than purely technical.

Despite substantial methodological advances, several barriers remain that limit the transfer of AI-derived biomarkers to clinical practice. Data heterogeneity, limited cohort diversity, incomplete biological annotation, and inconsistencies in reporting standards continue to be major challenges across disease domains. Addressing these limitations will require harmonized data collection frameworks, prospective multicenter validation studies, transparent model reporting, and closer integration of computational scientists with clinicians and experimental researchers. Equally important is the adoption of standardized evaluation frameworks that assess not only predictive performance but also robustness, calibration, fairness, and clinical utility. Such measures can mitigate the risk of overfitting, improve reproducibility, and facilitate regulatory acceptance.

Future progress will likely be driven by the convergence of foundation models, multimodal learning architectures, causal inference approaches, and systems biology frameworks. These technologies have the potential to move biomarker discovery beyond pattern recognition to the identification of biologically coherent disease programs that can inform therapeutic intervention. However, successful implementation will depend on rigorous external validation, continuous post-deployment monitoring, and prospective demonstration of patient benefit. Consequently, the most valuable AI-powered biomarkers will be those that not only accurately predict outcomes but also improve clinical decision-making, support therapeutic development, and contribute to more effective and personalized healthcare.

Despite substantial methodological progress, important barriers continue to limit the clinical translation of AI-derived biomarkers. If these conditions are met, AI-powered biomarkers can evolve from exploratory research outputs into reliable components of therapeutic development and routine clinical decision-making. Their greatest value will stem not only from increasingly sophisticated predictions but also from their ability to generate reproducible, biologically based, and clinically applicable knowledge. Consequently, the future success of AI-driven biomarker science will depend on integrating methodological innovation with rigorous validation, mechanistic understanding, and demonstrable patient benefit.

Author contributions

R.D. and N.A. designed the study, drafted the manuscript, and supervised and edited it. N.A. wrote the manuscript. All authors read and approved the review article.

Data availability

All supporting data are included in the article.

Competing interests

R.D. is the president of the INVAMED Institute for Medical Innovation. N.A. is retired and works as a volunteer consultant for Med-International UK Health Agency Ltd.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1.Ahmad, A., Imran, M. & Ahsan, H. Biomarkers as biomedical bioindicators: approaches and techniques for the detection, analysis, and validation of novel biomarkers of diseases. Pharmaceutics15, 1630 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
2.Drucker, E. & Krapfenbauer, K. Pitfalls and limitations in translation from biomarker discovery to clinical utility in predictive and personalised medicine. EPMA J.4, 7 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
3.Zhao, Q. et al. Applications and challenges of biomarker-based predictive models in proactive health management. Front. Public Health13, 1633487 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Ioannidis, J. P. Why most published research findings are false. PLoS Med.2, e124 (2005). [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Rifai, N., Gillette, M. A. & Carr, S. A. Protein biomarker discovery and validation: the long and uncertain path to clinical utility. Nat. Biotechnol.24, 971–983 (2006). [DOI] [PubMed] [Google Scholar]
6.AbdulRaheem, Y. Statistical significance versus clinical relevance: key considerations in interpretation medical research data. Indian J. Community Med.49, 791–795 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
7.Parikh, C. R. & Thiessen-Philbrook, H. Key concepts and limitations of statistical methods for evaluating biomarkers of kidney disease. J. Am. Soc. Nephrol.25, 1621–1629 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
8.Cook, N. R. Methods for evaluating novel biomarkers: a new paradigm. Int. J. Clin. Pract.64, 1723–1727 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Poste, G. Bring on the biomarkers. Nature469, 156–157 (2011). [DOI] [PubMed] [Google Scholar]
10.Kalluri, R. & LeBleu, V. S. The biology, function, and biomedical applications of exosomes. Science367, eaau6977 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Hasin, Y., Seldin, M. & Lusis, A. Multi-omics approaches to disease. Genome Biol.18, 83 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Gillies, R. J., Kinahan, P. E. & Hricak, H. Radiomics: images are more than pictures, they are data. Radiology278, 563–577 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Lambin, P. et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur. J. Cancer48, 441–446 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Aerts, H. J. et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat. Commun.5, 4006 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Fu, W., Ding, J., Gao, K., Ma, S. & Tian, L. A likelihood ratio test on temporal trends in age-period-cohort models with applications to the disparities of heart disease mortality among US populations and comparison with Japan. Stat. Med.40, 668–689 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Bu, K. et al. Identifying correlations driven by influential observations in large datasets. Brief. Bioinform.23, bbab482 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Al-Ewaidat, O. A. & Naffaa, M. M. Emerging AI- and biomarker-driven precision medicine in autoimmune rheumatic diseases: from diagnostics to therapeutic decision-making. Rheumato5, 17 (2025). [Google Scholar]
18.Awari, A. et al. Obesity biomarkers: exploring factors, ramification, machine learning, and AI-unveiling insights in health research. MedComm6, e70169 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Carletti, M. et al. Multimodal AI correlates of glucose spikes in people with normal glucose regulation, pre-diabetes and type 2 diabetes. Nat. Med.31, 3121–3127 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
20.Rajpurkar, P., Chen, E., Banerjee, O. & Topol, E. J. AI in health and medicine. Nat. Med.28, 31–38 (2022). [DOI] [PubMed] [Google Scholar]
21.Topol, E. J. High-performance medicine: the convergence of human and artificial intelligence. Nat. Med.25, 44–56 (2019). [DOI] [PubMed] [Google Scholar]
22.Norgeot, B. et al. Minimum information about clinical artificial intelligence modeling: the MI-CLAIM checklist. Nat. Med.26, 1320–1324 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Haibe-Kains, B. et al. Transparency and reproducibility in artificial intelligence. Nature586, E14–E16 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Diniz, M. A. Statistical methods for validation of predictive models. J. Nucl. Cardiol.29, 3248–3255 (2022). [DOI] [PubMed] [Google Scholar]
25.Vickers, A. J. & Elkin, E. B. Decision curve analysis: a novel method for evaluating prediction models. Med. Decis. Mak.26, 565–574 (2006). [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Manolio, T. A. et al. Finding the missing heritability of complex diseases. Nature461, 747–753 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Klarin, D. & Natarajan, P. Clinical utility of polygenic risk scores for coronary artery disease. Nat. Rev. Cardiol.19, 291–301 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Roadmap Epigenomics Consortium et al. Integrative analysis of 111 reference human epigenomes. Nature518, 317–330 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Nicholson, J. K. et al. Metabolic phenotyping in clinical and surgical environments. Nature491, 384–392 (2012). [DOI] [PubMed] [Google Scholar]
30.Lin, C. et al. Metabolomics for clinical biomarker discovery and therapeutic target identification. Molecules29, 2198 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Qiu, S. et al. Small molecule metabolites: discovery of biomarkers and therapeutic targets. Signal Transduct. Target. Ther.8, 132 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Ghosh, S., Zhao, X., Alim, M., Brudno, M. & Bhat, M. Artificial intelligence applied to ‘omics data in liver disease: towards a personalised approach for diagnosis, prognosis and treatment. Gut74, 295–311 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
33.Cheerla, A. & Gevaert, O. Deep learning with multimodal representation for pancancer prognosis prediction. Bioinformatics35, i446–i454 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Ahmed, I. et al. Plasma multi-omics and machine learning reveal predictive biomarkers for type 2 diabetes and retinopathy in Qatar Biobank Cohort. J. Transl. Med.23, 1159 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
35.Liu, J. et al. The role of multi-omics in biomarker discovery, diagnosis, prognosis, and therapeutic monitoring of tissue repair and regeneration processes. J. Orthop. Transl.54, 131–151 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Piwecka, M., Rajewsky, N. & Rybak-Wolf, A. Single-cell and spatial transcriptomics: deciphering brain complexity in health and disease. Nat. Rev. Neurol.19, 346–362 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
37.Aebersold, R. & Mann, M. Mass-spectrometric exploration of proteome structure and function. Nature537, 347–355 (2016). [DOI] [PubMed] [Google Scholar]
38.Stuart, T. & Satija, R. Integrative single-cell analysis. Nat. Rev. Genet.20, 257–272 (2019). [DOI] [PubMed] [Google Scholar]
39.Boxer, E. et al. Emerging clinical applications of single-cell RNA sequencing in oncology. Nat. Rev. Clin. Oncol.22, 315–326 (2025). [DOI] [PubMed] [Google Scholar]
40.Tirosh, I. & Suvà, M. L. Cancer cell states: lessons from ten years of single-cell RNA-sequencing of human tumors. Cancer Cell42, 1497–1506 (2024). [DOI] [PubMed] [Google Scholar]
41.Kleshchevnikov, V. et al. Cell2location maps fine-grained cell types in spatial transcriptomics. Nat. Biotechnol.40, 661–671 (2022). [DOI] [PubMed] [Google Scholar]
42.Dezem, F. S., Arjumand, W., DuBose, H., Morosini, N. S. & Plummer, J. Spatially resolved single-cell omics: methods, challenges, and future perspectives. Annu. Rev. Biomed. Data Sci.7, 131–153 (2024). [DOI] [PubMed] [Google Scholar]
43.De Jonghe, J. et al. scTrends: a living review of commercial single-cell and spatial ‘omic technologies. Cell Genom.4, 100723 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
44.Si, Y., Lee, J. S., Jun, G., Kang, H. M. & Lee, J. H. Spatial omics enters the microscopic arena: opportunities and challenges. Trends Genet.41, 774–787 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
45.Dinc, R. & Ardic, N. Mapping the immune environment: spatiotemporal dynamics in cardiovascular events. Front. Immunol.17, 1809941 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
46.Yiu, T. et al. Transformative advances in single-cell omics: a comprehensive review of foundation models, multimodal integration and computational ecosystems. J. Transl. Med.23, 1176 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
47.Baião, A. R. et al. A technical review of multi-omics data integration methods: from classical statistical to deep generative approaches. Brief. Bioinform.26, bbaf355 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
48.Robins, H. S. et al. Comprehensive assessment of T-cell receptor beta-chain diversity in alphabeta T cells. Blood114, 4099–4107 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Wigerblad, G. & Kaplan, M. J. Neutrophil extracellular traps in systemic autoimmune and autoinflammatory diseases. Nat. Rev. Immunol.23, 274–288 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
50.Ardic, N. & Dinc, R. Relationship between neutrophil extracellular traps and venous thromboembolism: pathophysiological and therapeutic role. Br. J. Hosp. Med.86, 1–15 (2025). [DOI] [PubMed] [Google Scholar]
51.Busarello, E. et al. Cell Marker Accordion: interpretable single-cell and spatial omics annotation in health and disease. Nat. Commun.16, 5399 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
52.Corcoran, R. B. & Chabner, B. A. Application of cell-free DNA analysis to cancer treatment. N. Engl. J. Med.379, 1754–1765 (2018). [DOI] [PubMed] [Google Scholar]
53.Wan, J. C. M. et al. Liquid biopsies come of age: towards implementation of circulating tumour DNA. Nat. Rev. Cancer17, 223–238 (2017). [DOI] [PubMed] [Google Scholar]
54.Crosby, D. et al. Early detection of cancer. Science375, eaay9040 (2022). [DOI] [PubMed] [Google Scholar]
55.Li, F. Q. & Cui, J. W. Circulating tumor DNA-minimal residual disease: an up-and-coming nova in resectable non-small-cell lung cancer. Crit. Rev. Oncol. Hematol.179, 103800 (2022). [DOI] [PubMed] [Google Scholar]
56.Ma, L. et al. Liquid biopsy in cancer current: status, challenges and future prospects. Signal Transduct. Target. Ther.9, 336 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
57.Qureshi, Z. et al. Liquid biopsies for early detection and monitoring of cancer: advances, challenges, and future directions. Ann. Med. Surg.87, 3244–3253 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
58.Pantel, K. & Alix-Panabières, C. Minimal residual disease as a target for liquid biopsy in patients with solid tumours. Nat. Rev. Clin. Oncol.22, 65–77 (2025). [DOI] [PubMed] [Google Scholar]
59.Boukouris, A. E. et al. A comprehensive overview of minimal residual disease in the management of early-stage and locally advanced non-small cell lung cancer. NPJ Precis. Oncol.9, 178 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
60.Hoang, T., Choi, M. K., Oh, J. H. & Kim, J. Utility of circulating tumor DNA to detect minimal residual disease in colorectal cancer: a systematic review and network meta-analysis. Int. J. Cancer157, 722–740 (2025). [DOI] [PubMed] [Google Scholar]
61.Negro, S. et al. Circulating tumor DNA as a real-time biomarker for minimal residual disease and recurrence prediction in stage II colorectal cancer: a systematic review and meta-analysis. Int. J. Mol. Sci.26, 2486 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
62.Geyer, P. E., Holdt, L. M., Teupser, D. & Mann, M. Revisiting biomarker discovery by plasma proteomics. Mol. Syst. Biol.13, 942 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
63.Ek, W. E., Karlsson, T., Höglund, J., Rask-Andersen, M. & Johansson, Å Causal effects of inflammatory protein biomarkers on inflammatory diseases. Sci. Adv.7, eabl4359 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
64.Moreno-Torres, M. et al. Factors that influence the quality of metabolomics data in in vitro cell toxicity studies: a systematic survey. Sci. Rep.11, 22119 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
65.Kumar, M. A. et al. Extracellular vesicles as tools and targets in therapy for diseases. Signal Transduct. Target. Ther.9, 27 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
66.Dinc, R. & Ardic, N. Artificial intelligence for exosomal biomarker discovery for cardiovascular diseases: multi-omics integration, reproducibility, and translational prospects. Cells15, 304 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
67.Welsh, J. A. et al. Minimal information for studies of extracellular vesicles (MISEV2023): from basic to advanced approaches. J. Extracell. Vesicles13, e12404 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
68.Théry, C. et al. Minimal information for studies of extracellular vesicles 2018 (MISEV2018): a position statement of the International Society for Extracellular Vesicles and update of the MISEV2014 guidelines. J. Extracell. Vesicles7, 1535750 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
69.Mateescu, B. et al. Obstacles and opportunities in the functional analysis of extracellular vesicle RNA: an ISEV position paper. J. Extracell. Vesicles6, 1286095 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
70.Darci-Maher, N. et al. Cross-tissue omics analysis discovers ten adipose genes encoding secreted proteins in obesity-related non-alcoholic fatty liver disease. EBioMedicine92, 104620 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
71.Li, H. et al. Identification and validation of core biomarkers for sepsis: a comprehensive analysis using bioinformatics and machine learning. Front. Immunol.16, 1700704 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
72.Xu, L., Li, J. & Gong, W. Applications of machine learning-assisted extracellular vesicles analysis technology in tumor diagnosis. Comput. Struct. Biotechnol. J.27, 2460–2472 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
73.Campanella, G. et al. Real-world deployment of a fine-tuned pathology foundation model for lung cancer biomarker detection. Nat. Med.31, 3002–3010 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
74.Nicke, T. et al. Tissue concepts: supervised foundation models in computational pathology. Comput. Biol. Med.186, 109621 (2025). [DOI] [PubMed] [Google Scholar]
75.Li, Z. et al. AI-enabled virtual spatial proteomics from histopathology for interpretable biomarker discovery in lung cancer. Nat. Med.32, 231–244 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
76.Pai, S. et al. Foundation model for cancer imaging biomarkers. Nat. Mach. Intell.6, 354–367 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
77.Ligero, M. et al. A whirl of radiomics-based biomarkers in cancer immunotherapy, why is large scale validation still lacking? NPJ Precis. Oncol.8, 42 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
78.Coravos, A. et al. Digital medicine: a primer on measurement. Digit. Biomark.3, 31–71 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
79.Bent, B. et al. Engineering digital biomarkers of interstitial glucose from noninvasive smartwatches. NPJ Digit. Med.4, 89 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
80.Perez, M. V. et al. Large-scale assessment of a smartwatch to identify atrial fibrillation. N. Engl. J. Med.381, 1909–1917 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
81.Zahedani, A. D. et al. Digital health application integrating wearable data and behavioral patterns improves metabolic health. NPJ Digit. Med.6, 216 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
82.Huang, X. et al. Digital biomarkers for interstitial glucose prediction in healthy individuals using wearables and machine learning. Sci. Rep.15, 30164 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
83.Macias Alonso, A. K., Hirt, J., Woelfle, T., Janiaud, P. & Hemkens, L. G. Definitions of digital biomarkers: a systematic mapping of the biomedical literature. BMJ Health Care Inf.31, e100914 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
84.Menden, M. P. et al. Machine learning prediction of cancer cell sensitivity to drugs based on genomic and chemical properties. PLoS ONE8, e61318 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
85.Ng, S., Masarone, S., Watson, D. & Barnes, M. R. The benefits and pitfalls of machine learning for biomarker discovery. Cell Tissue Res.394, 17–31 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
86.Esteva, A. et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature542, 115–118 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
87.Bera, K., Schalper, K. A., Rimm, D. L., Velcheti, V. & Madabhushi, A. Artificial intelligence in digital pathology: new tools for diagnosis and precision oncology. Nat. Rev. Clin. Oncol.16, 703–715 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
88.Richter, T., Bahrami, M., Xia, Y., Fischer, D. S. & Theis, F. J. Delineating the effective use of self-supervised learning in single-cell genomics. Nat. Mach. Intell.7, 68–78 (2025). [Google Scholar]
89.Acosta, J. N., Falcone, G. J., Rajpurkar, P. & Topol, E. J. Multimodal biomedical AI. Nat. Med.28, 1773–1784 (2022). [DOI] [PubMed] [Google Scholar]
90.Lipkova, J. et al. Artificial intelligence for multimodal data integration in oncology. Cancer Cell40, 1095–1110 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
91.Steyaert, S. et al. Multimodal data fusion for cancer biomarker discovery with deep learning. Nat. Mach. Intell.5, 351–362 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
92.Krones, F., Marikkar, U., Parsons, G., Szmul, A. & Mahdi, A. Review of multimodal machine learning approaches in healthcare. Inf. Fusion114, 102690 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
93.Chen, R. J. et al. Pan-cancer integrative histology-genomic analysis via multimodal deep learning. Cancer Cell40, 865–878.e6 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
94.Mataraso, S. J. et al. A machine learning approach to leveraging electronic health records for enhanced omics analysis. Nat. Mach. Intell.7, 293–306 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
95.Camacho, D. M., Collins, K. M., Powers, R. K., Costello, J. C. & Collins, J. J. Next-generation machine learning for biological networks. Cell173, 1581–1592 (2018). [DOI] [PubMed] [Google Scholar]
96.Lin, J. et al. Chemokine-like factor-like MARVEL transmembrane domain-containing 1 identified as a novel target protein for immune dysregulation in sepsis: a machine learning and molecular dynamics framework for diagnostic biomarkers and therapeutic exploration. Int. J. Biol. Macromol.319, 145276 (2025). [DOI] [PubMed] [Google Scholar]
97.Xie, W. et al. A novel biomarker selection method combining graph neural network and gene relationships applied to microarray data. BMC Bioinforma.23, 303 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
98.Dharanipragada, P. et al. Blocking genomic instability prevents acquired resistance to MAPK inhibitor therapy in melanoma. Cancer Discov.13, 880–909 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
99.Dalla-Torre, H. et al. Nucleotide Transformer: building and evaluating robust foundation models for human genomics. Nat. Methods22, 287–297 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
100.de Almeida, B., Lopez, M. & Pierrot, T. Generalized AI models for genomics applications. Nat. Methods22, 231–232 (2025). [DOI] [PubMed] [Google Scholar]
101.Guo, F. et al. Foundation models in bioinformatics. Natl. Sci. Rev.12, nwaf028 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
102.Topol, E. J. Learning the language of life with AI. Science387, eadv4414 (2025). [DOI] [PubMed] [Google Scholar]
103.Vorontsov, E. et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nat. Med.30, 2924–2935 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
104.Campanella, G. et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat. Med.25, 1301–1309 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
105.Ding, T. et al. A multimodal whole-slide foundation model for pathology. Nat. Med.31, 3749–3761 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
106.Xu, H. et al. A whole-slide foundation model for digital pathology from real-world data. Nature630, 181–188 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
107.Moor, M. et al. Foundation models for generalist medical artificial intelligence. Nature616, 259–265 (2023). [DOI] [PubMed] [Google Scholar]
108.Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat. Mach. Intell.1, 206–215 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
109.Budhkar, A., Song, Q., Su, J. & Zhang, X. Demystifying the black box: a survey on explainable artificial intelligence (XAI) in bioinformatics. Comput. Struct. Biotechnol. J.27, 346–359 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
110.Zhang, K., Wang, D., Lin, F., Xie, J. & Zhou, W. A comprehensive review of explainable artificial intelligence in healthcare methods, evaluation, and clinical integration. iScience29, 115026 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
111.Toussaint, P. A. et al. Explainable artificial intelligence for omics data: a systematic mapping study. Brief. Bioinform.25, bbad453 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
112.Schölkopf, B. et al. Toward causal representation learning. Proc. IEEE109, 612–634 (2021). [Google Scholar]
113.Pearl, J. An introduction to causal inference. Int. J. Biostat.6, 7 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
114.Javaid, H. et al. The impact of artificial intelligence on biomarker discovery. Emerg. Top. Life Sci.8, 89–105 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
115.Prentice, R. L. Surrogate endpoints in clinical trials: definition and operational criteria. Stat. Med.8, 431–440 (1989). [DOI] [PubMed] [Google Scholar]
116.Gao, S. et al. Modeling drug mechanism of action with large scale gene-expression profiles using GPAR, an artificial intelligence platform. BMC Bioinforma.22, 17 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
117.García-Campos, M. A., Espinal-Enríquez, J. & Hernández-Lemus, E. Pathway analysis: state of the art. Front. Physiol.6, 383 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
118.Mubeen, S. et al. The impact of pathway database choice on statistical enrichment analysis and predictive modeling. Front. Genet.10, 1203 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
119.Zhou, J. L. et al. Analysis of single-cell CRISPR perturbations indicates that enhancers predominantly act multiplicatively. Cell Genom.4, 100672 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
120.Bjerregaard, A., Prada-Luengo, I., Das, V. & Krogh, A. What do single-cell models already know about perturbations? Genes16, 1439 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
121.Egbon, O. A., Hickey, J. W. & Anchang, B. Fusion of spatiotemporal and network models to prioritize multiscale effects in single-cell perturbations. Brief. Bioinform.26, bbaf277 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
122.Kleandrova, V. V., Cordeiro, M. N. D. S. & Speck-Planche, A. Perturbation-theory machine learning for multi-target drug discovery in modern anticancer research. Curr. Issues Mol. Biol.47, 301 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
123.Davey Smith, G. & Hemani, G. Mendelian randomization: genetic anchors for causal inference in epidemiological studies. Hum. Mol. Genet.23, R89–R98 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
124.Mo, Z., Zhang, Z., Miao, Q. & Tsui, K. L. Sparsity-constrained invariant risk minimization for domain generalization with application to machinery fault diagnosis modeling. IEEE Trans. Cybern.54, 1547–1559 (2024). [DOI] [PubMed] [Google Scholar]
125.Ni, S. et al. Identifying compound-protein interactions with knowledge graph embedding of perturbation transcriptomics. Cell Genom.4, 100655 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
126.Crespi, E. et al. Resolving the rules of robustness and resilience in biology across scales. Integr. Comp. Biol.61, 2163–2179 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
127.Wood, L. B., Winslow, A. R. & Strasser, S. D. Systems biology of neurodegenerative diseases. Integr. Biol.7, 758–775 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
128.Ocana, A. Integrating artificial intelligence in drug discovery and early drug development: a transformative approach. Biomark. Res.13, 45 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
129.Arango-Argoty, G. et al. AI-driven predictive biomarker discovery with contrastive learning to improve clinical trial outcomes. Cancer Cell43, 875–890.e8 (2025). [DOI] [PubMed] [Google Scholar]
130.Collins, G. S. et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ385, e078378 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
131.Libby, P., Ridker, P. M. & Hansson, G. K. Inflammation in atherosclerosis: from pathophysiology to practice. J. Am. Coll. Cardiol.54, 2129–2138 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
132.Firestein, G. S. & McInnes, I. B. Immunopathogenesis of rheumatoid arthritis. Immunity46, 183–196 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
133.Fountzilas, E., Tsimberidou, A. M., Vo, H. H. & Kurzrock, R. Clinical trial design in the era of precision medicine. Genome Med.14, 101 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
134.Jørgensen, J. T. Companion and complementary diagnostics: clinical and regulatory perspectives. Trends Cancer2, 706–712 (2016). [DOI] [PubMed] [Google Scholar]
135.Kern, S. E. Why your new cancer biomarker may never work: recurrent patterns and remarkable diversity in biomarker failures. Cancer Res.72, 6097–6101 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
136.Ernst, S. M. et al. Utilizing ctDNA to discover mechanisms of resistance to targeted therapies in patients with metastatic NSCLC: towards more informative trials. Nat. Rev. Clin. Oncol.22, 371–378 (2025). [DOI] [PubMed] [Google Scholar]
137.Papayannopoulos, V. Neutrophil extracellular traps in immunity and disease. Nat. Rev. Immunol.18, 134–147 (2018). [DOI] [PubMed] [Google Scholar]
138.Bonaventura, A. et al. Endothelial dysfunction and immunothrombosis as key pathogenic mechanisms in COVID-19. Nat. Rev. Immunol.21, 319–329 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
139.Finlayson, S. G. et al. The clinician and dataset shift in artificial intelligence. N. Engl. J. Med.385, 283–286 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
140.Roberts, M. et al. Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and CT scans. Nat. Mach. Intell.3, 199–217 (2021). [Google Scholar]
141.Tang, S., Yuan, K. & Chen, L. Molecular biomarkers, network biomarkers, and dynamic network biomarkers for diagnosis and prediction of rare diseases. Fundam. Res.2, 894–902 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
142.Qu, N. et al. Multimodal interpretable deep learning for transcriptome-informed precision oncology and drug mechanism analysis. NPJ Digit. Med. 10.1038/s41746-026-02735-x (2026). [DOI] [PMC free article] [PubMed]
143.Subramanian, I., Verma, S., Kumar, S., Jere, A. & Anamika, K. Multi-omics data integration, interpretation, and its application. Bioinform. Biol. Insights14, 1177932219899051 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
144.Galindez, G., Sadegh, S., Baumbach, J., Kacprowski, T. & List, M. Network-based approaches for modeling disease regulation and progression. Comput. Struct. Biotechnol. J.21, 780–795 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
145.Liu, Y., Zhu, K., Peng, W., Liu, Z. & Mao, X. Multi-omics and artificial intelligence for precision drug discovery and potential clinical applications. Signal Transduct. Target. Ther.11, 210 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
146.Ben-Jaafar, A. et al. Artificial intelligence-based biomarkers for the diagnosis and treatment of neurological conditions: a narrative review. Mol. Brain19, 26 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
147.Moons, K. G. M. et al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ388, e082505 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
148.Valous, N. A. et al. Graph machine learning for integrated multi-omics analysis. Br. J. Cancer131, 205–211 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
149.Rodriguez-Mier, P. et al. Unifying multi-sample network inference from prior knowledge and omics data with CORNETO. Nat. Mach. Intell.7, 1168–1186 (2025). [Google Scholar]
150.Wang, K. et al. TG468: a text graph convolutional network for predicting clinical response to immune checkpoint inhibitor therapy. Brief. Bioinform.25, bbae017 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
151.Bonaventura, A. et al. The pathophysiological role of neutrophil extracellular traps in inflammatory diseases. Thromb. Haemost.118, 6–27 (2018). [DOI] [PubMed] [Google Scholar]
152.Marsiglia, J. Computationally guided high-throughput engineering of an anti-CRISPR protein for precise genome editing in human cells. Cell Rep. Methods4, 100882 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
153.Van Calster, B., McLernon, D. J., van Smeden, M., Wynants, L. & Steyerberg, E. W. Calibration: the Achilles heel of predictive analytics. BMC Med.17, 230 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
154.Piovani, D., Sokou, R., Tsantes, A. G., Vitello, A. S., & Bonovas, S. Optimizing clinical decision making with decision curve analysis: insights for clinical investigators. Healthcare11, 2244 (2023). [DOI] [PMC free article] [PubMed]
155.Riley, R. D. et al. Minimum sample size for developing a multivariable prediction model: PART II—binary and time-to-event outcomes. Stat. Med.38, 1276–1296 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
156.Collins, G. S. et al. Protocol for development of a reporting guideline (TRIPOD-AI) and risk of bias tool (PROBAST-AI) for diagnostic and prognostic prediction model studies based on artificial intelligence. BMJ Open11, e048008 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
157.van de Vegte, Y. J. et al. Genetic insights into resting heart rate and its role in cardiovascular disease. Nat. Commun.14, 4646 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
158.Johnson, W. E., Li, C. & Rabinovic, A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics8, 118–127 (2007). [DOI] [PubMed] [Google Scholar]
159.Leek, J. T. et al. Tackling the widespread and critical impact of batch effects in high-throughput data. Nat. Rev. Genet.11, 733–739 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
160.Hu, F. et al. DeepComBat: a statistically motivated, hyperparameter-robust, deep learning approach to harmonization of neuroimaging data. Hum. Brain Mapp.45, e26708 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
161.Fortin, J. P. et al. Harmonization of cortical thickness measurements across scanners and sites. Neuroimage167, 104–120 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
162.Wong, A. et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Intern. Med.181, 1065–1070 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
163.Dhiman, P. et al. Risk of bias of prognostic models developed using machine learning: a systematic review in oncology. Diagn. Progn. Res.6, 13 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
164.Hernández, B., Pennington, S. R. & Parnell, A. C. Bayesian methods for proteomic biomarker development. EuPA Open Proteom.9, 54–64 (2015). [Google Scholar]
165.Olsson, H. et al. Estimating diagnostic uncertainty in artificial intelligence assisted pathology using conformal prediction. Nat. Commun.13, 7761 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
166.Seabroke, S. et al. Performance of stratified and subgrouped disproportionality analyses in spontaneous databases. Drug Saf.39, 355–364 (2016). [DOI] [PubMed] [Google Scholar]
167.Russo, C. A. M. & Selvatti, A. P. Bootstrap and rogue identification tests for phylogenetic analyses. Mol. Biol. Evol.35, 2327–2333 (2018). [DOI] [PubMed] [Google Scholar]
168.Hayes, D. F. Biomarker validation and testing. Mol. Oncol.9, 960–966 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
169.Ou, F. S., Michiels, S., Shyr, Y., Adjei, A. A. & Oberg, A. L. Biomarker discovery and validation: statistical considerations. J. Thorac. Oncol.16, 537–545 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
170.Vasey, B. et al. Reporting guideline for the early stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. BMJ377, e070904 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
171.Shi, L. et al. The MicroArray Quality Control (MAQC)-II study of common practices for the development and validation of microarray-based predictive models. Nat. Biotechnol.28, 827–838 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
172.MAQC Consortium The MicroArray Quality Control (MAQC) project shows inter- and intraplatform reproducibility of gene expression measurements. Nat. Biotechnol.24, 1151–1161 (2006). [DOI] [PMC free article] [PubMed] [Google Scholar]
173.SEQC/MAQC-III Consortium A comprehensive assessment of RNA-seq accuracy, reproducibility and information content by the Sequencing Quality Control Consortium. Nat. Biotechnol.32, 903–914 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
174.Balikci, M. A., Njume, C. M. & Cakmak, A. BioMark: biomarker analysis tool. BMC Bioinform.27, 42 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
175.Jiang, L. et al. Utilizing stability criteria in choosing feature selection methods yields reproducible results in microbiome data. Biometrics78, 1155–1167 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
176.Traverso, A. et al. Powering responsible artificial intelligence with high-quality real-world data: the S-RACE platform for scalable, multi-specialty clinical research. NPJ Digit. Med.9, 6 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
177.Davatzikos, C. et al. AI-based prognostic imaging biomarkers for precision neuro-oncology: the ReSPOND consortium. Neuro. Oncol.22, 886–888 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
178.Ransohoff, D. F. Rules of evidence for cancer molecular-marker discovery and validation. Nat. Rev. Cancer4, 309–314 (2004). [DOI] [PubMed] [Google Scholar]
179.Pepe, M. S., Janes, H., Longton, G., Leisenring, W. & Newcomb, P. Limitations of the odds ratio in gauging the performance of a diagnostic, prognostic, or screening marker. Am. J. Epidemiol.159, 882–890 (2004). [DOI] [PubMed] [Google Scholar]
180.Obermeyer, Z., Powers, B., Vogeli, C. & Mullainathan, S. Dissecting racial bias in an algorithm used to manage the health of populations. Science366, 447–453 (2019). [DOI] [PubMed] [Google Scholar]
181.Mulherin, S. A. & Miller, W. C. Spectrum bias or spectrum effect? Subgroup variation in diagnostic test evaluation. Ann. Intern. Med.137, 598–602 (2002). [DOI] [PubMed] [Google Scholar]
182.Hatherley, J. A moving target in AI-assisted decision-making: dataset shift, model updating, and the problem of update opacity. Ethics Inf. Technol.27, 20 (2025). [Google Scholar]
183.Cagney, D. N. et al. The FDA NIH Biomarkers, EndpointS, and other Tools (BEST) resource in neuro-oncology. Neuro. Oncol.20, 1162–1172 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
184.Agarwal, V. et al. Learning statistical models of phenotypes using noisy labeled training data. J. Am. Med. Inform. Assoc.23, 1166–1173 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
185.Ioannidis, J. P. Why most discovered true associations are inflated. Epidemiology19, 640–648 (2008). [DOI] [PubMed] [Google Scholar]
186.Benjamens, S., Dhunnoo, P. & Meskó, B. The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database. NPJ Digit. Med.3, 118 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
187.Mattes, W. B. & Goodsaid, F. Regulatory landscapes for biomarkers and diagnostic tests: Qualification, approval, and role in clinical practice. Exp. Biol. Med.243, 256–261 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
188.Gallas, B. D. et al. FDA fosters innovative approaches in research, resources, and collaboration. Nat. Mach. Intell.4, 97–98 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
189.U.S. Food and Drug Administration. Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan. https://www.fda.gov/media/145022/download (2021).
190.U.S. Food and Drug Administration. Marketing submission recommendations for a predetermined change control plan for artificial intelligence-enabled device software functions. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence (2025).
191.Hu, Y., Zou, Y., Qiao, L. & Lin, L. Integrative proteomic and metabolomic elucidation of cardiomyopathy with in vivo and in vitro models and clinical samples. Mol. Ther.32, 3288–3312 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
192.Babu, M. & Snyder, M. Multi-omics profiling for health. Mol. Cell. Proteom.22, 100561 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
193.Oikonomou, E. K. et al. Non-invasive detection of coronary inflammation using computed tomography and prediction of residual cardiovascular risk (the CRISP CT study): a post-hoc analysis of prospective outcome data. Lancet392, 929–939 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
194.Attia, Z. I. et al. Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nat. Med.25, 70–74 (2019). [DOI] [PubMed] [Google Scholar]
195.Bai, W. et al. Automated cardiovascular magnetic resonance image analysis with fully convolutional networks. J. Cardiovasc. Magn. Reson.20, 65 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
196.Shah, A. M. et al. Large scale plasma proteomics identifies novel proteins and protein networks associated with heart failure development. Nat. Commun.15, 528 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
197.Faulkner, L. G., Howells, L. M., Pepper, C., Shaw, J. A. & Thomas, A. L. The utility of ctDNA in detecting minimal residual disease following curative surgery in colorectal cancer: a systematic review and meta-analysis. Br. J. Cancer128, 297–309 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
198.Chidharla, A. et al. Circulating tumor DNA as a minimal residual disease assessment and recurrence risk in patients undergoing curative-intent resection with or without adjuvant chemotherapy in colorectal cancer: a systematic review and meta-analysis. Int. J. Mol. Sci.24, 10230 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
199.Zheng, J., Qin, C., Wang, Q., Tian, D. & Chen, Z. Circulating tumour DNA-based molecular residual disease detection in resectable cancers: a systematic review and meta-analysis. EBioMedicine103, 105109 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
200.Zhu, L. et al. Minimal residual disease (MRD) detection in solid tumors using circulating tumor DNA: a systematic review. Front. Genet.14, 1172108 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
201.Martín-Arana, J. et al. Whole-exome tumor-agnostic ctDNA analysis enhances minimal residual disease detection and reveals relapse mechanisms in localized colon cancer. Nat. Cancer6, 1000–1016 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
202.Martínez-Castedo, B. et al. Minimal residual disease in colorectal cancer. Tumor-informed versus tumor-agnostic approaches: unraveling the optimal strategy. Ann. Oncol.36, 263–276 (2025). [DOI] [PubMed] [Google Scholar]
203.Semenkovich, N. P., Szymanski, J. J., Earland, N., Chauhan, P. S. & Chaudhuri, A. A. Genomic approaches to cancer and minimal residual disease detection using circulating tumor DNA. J. Immunother. Cancer11, e006284 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
204.Rosenlund, L. et al. ctDNA can detect minimal residual disease in curative treated non-small cell lung cancer patients using a tumor agnostic approach. Lung Cancer203, 108528 (2025). [DOI] [PubMed] [Google Scholar]
205.Eslinger, C., Wu, C. & Ahn, D. H. Minimal residual disease monitoring via ctDNA: a case report of Lynch syndrome with synchronous colorectal cancer and review of literature. J. Gastrointest. Oncol.15, 1341–1347 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
206.Kather, J. N. et al. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nat. Med.25, 1054–1056 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
207.Qian, L. et al. AI-empowered perturbation proteomics for complex biological systems. Cell Genom.4, 100691 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
208.Sekar, P. K. & Veerabathiran, R. Systems immunology meets clinical translation: multi-omic approaches to predict therapy response in cancer and autoimmune disease. Clin. Immunol. Commun.9, 22 (2025). [Google Scholar]
209.Yuen, K. C. et al. High systemic and tumor-associated IL-8 correlates with reduced clinical benefit of PD-L1 blockade. Nat. Med.26, 693–698 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
210.Kasi, P. M. et al. Impact of circulating tumor DNA-based detection of molecular residual disease on the conduct and design of clinical trials for solid tumors. JCO Precis. Oncol.6, e2100181 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
211.Bartolomucci, A. et al. Circulating tumor DNA to monitor treatment response in solid tumors and advance precision oncology. NPJ Precis. Oncol.9, 84 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
212.García de Yébenes Prous, M. J. & Carmona Ortells, L. Biomarkers: how to consolidate them in clinical practice. Reumatol. Clin.20, 386–391 (2024). [DOI] [PubMed] [Google Scholar]
213.Mattsson-Carlgren, N. et al. Longitudinal plasma p-tau217 is increased in early stages of Alzheimer’s disease. Brain143, 3234–3241 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
214.Hansson, O., Blennow, K., Zetterberg, H. & Dage, J. Blood biomarkers for Alzheimer’s disease in clinical practice and trials. Nat. Aging3, 506–519 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
215.Jack, C. R. et al. NIA-AA Research Framework: toward a biological definition of Alzheimer’s disease. Alzheimers Dement.14, 535–562 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
216.Leuzy, A. et al. Tau PET imaging in neurodegenerative tauopathies—still a challenge. Mol. Psychiatry24, 1112–1134 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
217.Mormino, E. C. et al. Tau PET imaging with 18F-PI-2620 in aging and neurodegenerative diseases. Eur. J. Nucl. Med. Mol. Imaging48, 2233–2244 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
218.Ravi, D. et al. Deep learning for health informatics. IEEE J. Biomed. Health Inform.21, 4–21 (2017). [DOI] [PubMed] [Google Scholar]
219.Zhang, W., Xiao, D., Mao, Q. & Xia, H. Role of neuroinflammation in neurodegeneration development. Signal Transduct. Target. Ther.8, 267 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
220.van der Kant, R., Goldstein, L. S. B. & Ossenkoppele, R. Amyloid-β-independent regulators of tau pathology in Alzheimer disease. Nat. Rev. Neurosci.21, 21–35 (2020). [DOI] [PubMed] [Google Scholar]
221.Serrano-Pozo, A., Frosch, M. P., Masliah, E. & Hyman, B. T. Neuropathological alterations in Alzheimer disease. Cold Spring Harb. Perspect. Med.1, a006189 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
222.Schöll, M. et al. Challenges in the practical implementation of blood biomarkers for Alzheimer’s disease. Lancet Healthy Longev.5, 100630 (2024). [DOI] [PubMed] [Google Scholar]
223.Smillie, C. S. et al. Intra- and inter-cellular rewiring of the human colon during ulcerative colitis. Cell178, 714–730.e22 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
224.Wang, L. et al. An atlas of single-cell eQTLs dissects autoimmune disease genes and identifies novel drug classes for treatment. Cell Genom.5, 100820 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
225.Smolen, J. S. et al. Treating rheumatoid arthritis to target: recommendations of an international task force. Ann. Rheum. Dis.69, 631–637 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
226.Hauser, S. L. & Cree, B. A. C. Treatment of multiple sclerosis: a review. Am. J. Med.133, 1380–1390.e2 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
227.Tilg, H., Byrne, C. D. & Targher, G. NASH drug treatment development: challenges and lessons. Lancet Gastroenterol. Hepatol.8, 943–954 (2023). [DOI] [PubMed] [Google Scholar]
228.Thiele, M. et al. Opportunities and barriers in omics-based biomarker discovery for steatotic liver diseases. J. Hepatol.81, 345–359 (2024). [DOI] [PubMed] [Google Scholar]
229.Lazarus, J. V. et al. A call for doubling the diagnostic rate of at-risk metabolic dysfunction-associated steatohepatitis. Lancet Reg. Health Eur.54, 101320 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
230.Noureddin, M. et al. Serum identification of at-risk MASH: the metabolomics-advanced steatohepatitis fibrosis score (MASEF). Hepatology79, 135–148 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
231.Rinella, M. E. & Sookoian, S. From NAFLD to MASLD: updated naming and diagnosis criteria for fatty liver disease. J. Lipid Res.65, 100485 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
232.Sergi, C. M. NAFLD (MASLD)/NASH (MASH): does it bother to label at all? A comprehensive narrative review. Int. J. Mol. Sci.25, 8462 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
233.Park, I. G. et al. Gut microbiota-based machine-learning signature for the diagnosis of alcohol-associated and metabolic dysfunction-associated steatotic liver disease. Sci. Rep.14, 16122 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
234.Sveinbjornsson, G. et al. A multi-omics study of non-alcoholic fatty liver disease. Nat. Genet.54, 1652–1663 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
235.Verschuren, L. et al. Development of a novel non-invasive biomarker panel for hepatic fibrosis in MASLD. Nat. Commun.15, 4564 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
236.Stratakis, N. et al. Multi-omic architecture, biological pathways, and prenatal determinants of childhood obesity and metabolic dysfunction. Nat. Commun.16, 654 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
237.Chakaroun, R. M. et al. Multi-omic definition of metabolic obesity through adipose tissue-microbiome interactions. Nat. Med.32, 113–125 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
238.Soták, M., Clark, M., Suur, B. E. & Börgeson, E. Inflammation and resolution in obesity. Nat. Rev. Endocrinol.21, 45–61 (2025). [DOI] [PubMed] [Google Scholar]
239.Rönn, T., Perfilyev, A., Oskolkov, N. & Ling, C. Predicting type 2 diabetes via machine learning integration of multiple omics from human pancreatic islets. Sci. Rep.14, 14637 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
240.Dai, L. et al. A deep learning system for predicting time to progression of diabetic retinopathy. Nat. Med.30, 584–594 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
241.Ratziu, V. et al. Confirmatory biomarker diagnostic studies are not needed when transitioning from NAFLD to MASLD. J. Hepatol.80, e51–e52 (2024). [DOI] [PubMed] [Google Scholar]
242.Chen, L., Zhang, X. & Shi, P. Recent advances in biomarkers for detection and diagnosis of sepsis and organ dysfunction: a comprehensive review. Eur. J. Med. Res.30, 1081 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
243.de Nooijer, A. H., Pickkers, P., Netea, M. G. & Kox, M. Inflammatory biomarkers to predict the prognosis of acute bacterial and viral infections. J. Crit. Care78, 154360 (2023). [DOI] [PubMed] [Google Scholar]
244.Xiong, W., Zhan, Y., Xiao, R. & Liu, F. Advancing sepsis diagnosis and immunotherapy: machine learning-driven identification of stable molecular biomarkers and therapeutic targets. Sci. Rep.15, 8333 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
245.Yang, J. et al. Machine learning based screening of biomarkers associated with cell death and immunosuppression of multiple life stages sepsis populations. Sci. Rep.15, 30302 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
246.Rolfo, C. et al. Liquid biopsy for advanced NSCLC: a consensus statement from the International Association for the Study of Lung Cancer. J. Thorac. Oncol.16, 1647–1662 (2021). [DOI] [PubMed] [Google Scholar]
247.Sabari, J. K. et al. A prospective study of circulating tumor DNA to guide matched targeted therapy in lung cancers. J. Natl. Cancer Inst.111, 575–583 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
248.Jiang, P. et al. Signatures of T cell dysfunction and exclusion predict cancer immunotherapy response. Nat. Med.24, 1550–1558 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
249.Auslander, N. et al. Robust prediction of response to immune checkpoint blockade therapy in metastatic melanoma. Nat. Med.24, 1545–1549 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
250.Berry, S. M., Carlin, B. P., Lee, J. J. & Muller, P. Bayesian Adaptive Methods for Clinical Trials (CRC Press, 2010).
251.Topalian, S. L., Taube, J. M., Anders, R. A. & Pardoll, D. M. Mechanism-driven biomarkers to guide immune checkpoint blockade in cancer therapy. Nat. Rev. Cancer16, 275–287 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
252.Subbiah, V. & Kurzrock, R. Challenging standard-of-care paradigms in the precision oncology era. Trends Cancer4, 101–109 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
253.Okusanya, O. O. et al. FDA-AACR strategies for optimizing dosages for oncology drug products: early-phase trials using innovative trial designs and biomarkers. Clin. Cancer Res.31, 4882–4890 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
254.Fleming, T. R. & DeMets, D. L. Surrogate end points in clinical trials: are we being misled? Ann. Intern. Med.125, 605–613 (1996). [DOI] [PubMed] [Google Scholar]
255.Fleming, T. R. & Powers, J. H. Biomarkers and surrogate endpoints in clinical trials. Stat. Med.31, 2973–2984 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
256.Moqri, M. et al. Validation of biomarkers of aging. Nat. Med.30, 360–372 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
257.Sherman, R. E. et al. Real-world evidence—what is it and what can it tell us? N. Engl. J. Med.375, 2293–2297 (2016). [DOI] [PubMed] [Google Scholar]
258.Ohtsu, H., Shimomura, A. & Sase, K. Real-world evidence in cardio-oncology: what is it and what can it tell us? JACC CardioOncol.4, 95–97 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
259.Griesinger, F., Cox, O., Sammon, C., Ramagopalan, S. V. & Popat, S. Health technology assessments and real-world evidence: tell us what you want, what you really, really want. J. Comp. Eff. Res.11, 297–299 (2022). [DOI] [PubMed] [Google Scholar]
260.Ardic, N. & Dinc, R. Artificial intelligence in healthcare: current regulatory landscape and future directions. Br. J. Hosp. Med.86, 1–21 (2025). [DOI] [PubMed] [Google Scholar]
261.U.S. Food and Drug Administration. Biomarker qualification: evidentiary framework. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/biomarker-qualification-evidentiary-framework (2018).
262.Coravos, A., Khozin, S. & Mandl, K. D. Developing and adopting safe and effective digital biomarkers to improve patient outcomes. NPJ Digit. Med.2, 14 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
263.Price, W. N. II. & Cohen, I. G. Privacy in the age of medical big data. Nat. Med.25, 37–43 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
264.Patel, A. P. et al. A multi-ancestry polygenic risk score improves risk prediction for coronary artery disease. Nat. Med.29, 1793–1803 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
265.GBD 2023 Causes of Death Collaborators Global burden of 292 causes of death in 204 countries and territories and 660 subnational locations, 1990–2023: a systematic analysis for the Global Burden of Disease Study 2023. Lancet406, 1811–1872 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
266.Alum, E. U. AI-driven biomarker discovery: enhancing precision in cancer diagnosis and prognosis. Discov. Oncol.16, 313 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
267.Singh, Y. et al. Beyond post hoc explanations: a comprehensive framework for responsible artificial intelligence in medical imaging through transparency, interpretability, and explainability. Bioengineering12, 879 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
268.Begley, C. G. & Ellis, L. M. Drug development: raise standards for preclinical cancer research. Nature483, 531–533 (2012). [DOI] [PubMed] [Google Scholar]
269.Laird, A. R. Large, open datasets for human connectomics research: considerations for reproducible and responsible data use. Neuroimage244, 118579 (2021). [DOI] [PubMed] [Google Scholar]
270.Fountzilas, E., Pearce, T., Baysal, M. A., Chakraborty, A. & Tsimberidou, A. M. Convergence of evolving artificial intelligence and machine learning techniques in precision oncology. NPJ Digit. Med.8, 75 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
271.Mendez, K. M. et al. A roadmap to precision medicine through post-genomic electronic medical records. Nat. Commun.16, 1700 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
272.Liu, Y., Dai, Y. & Wang, L. Spatial omics at the forefront: emerging technologies, analytical innovations, and clinical applications. Cancer Cell44, 24–49 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
273.Zhang, Y. et al. Robust integration and annotation of single-cell and spatial omics data using interpretable gene programs. Cell Genom.6, 101105 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
274.Van de Sande, B. et al. Applications of single-cell RNA sequencing in drug discovery and development. Nat. Rev. Drug Discov.22, 496–520 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
275.Dasaradharami Reddy, K. & Gadekallu, T. R. A comprehensive survey on federated learning techniques for healthcare informatics. Comput. Intell. Neurosci.2023, 8393990 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
276.Olawade, D. B. et al. Digital twin paradigm in diabetes prediction and management. Diab. Res. Clin. Pract.231, 113075 (2026). [DOI] [PubMed] [Google Scholar]
277.Viceconti, M. et al. In silico trials: verification, validation and uncertainty quantification of predictive models used in the regulatory evaluation of biomedical products. Methods185, 120–127 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
278.Strimbu, K. & Tavel, J. A. What are biomarkers? Curr. Opin. HIV AIDS5, 463–466 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
279.Chenchula, S., Paraskar, G., Krishna, S. & Chavan, M. Liquid biopsy in clinical practice: current evidence and future directions. In Liquid Biopsy in Cancer Management (eds Bhatt, S., Anitha, A., Chenchula, S. & Mishra, N.) 405–438 (Academic Press, 2026).
280.Ohara, S. & Suda, K. Current status and future challenges of liquid biopsy. Cells14, 2000 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
281.Abràmoff, M. D., Lavin, P. T., Birch, M., Shah, N. & Folk, J. C. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. NPJ Digit. Med.1, 39 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
282.Elijovich, L. et al. Automated emergent large vessel occlusion detection by artificial intelligence improves stroke workflow in a hub and spoke stroke system of care. J. Neurointerv. Surg.14, 704–708 (2022). [DOI] [PubMed] [Google Scholar]
283.Chilamkurthy, S. et al. Deep learning algorithms for detection of critical findings in head CT scans: a retrospective study. Lancet392, 2388–2396 (2018). [DOI] [PubMed] [Google Scholar]
284.Komorowski, M., Celi, L. A., Badawi, O., Gordon, A. C. & Faisal, A. A. The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care. Nat. Med.24, 1716–1720 (2018). [DOI] [PubMed] [Google Scholar]
285.Weikert, T. et al. Automated detection of pulmonary embolism in CT pulmonary angiograms using an AI-powered algorithm. Eur. Radiol.30, 6545–6553 (2020). [DOI] [PubMed] [Google Scholar]
286.Beam, A. L. & Kohane, I. S. Big data and machine learning in health care. JAMA319, 1317–1318 (2018). [DOI] [PubMed] [Google Scholar]
287.Esteva, A. et al. A guide to deep learning in healthcare. Nat. Med.25, 24–29 (2019). [DOI] [PubMed] [Google Scholar]
288.Willis, B. H. Spectrum bias—why clinicians need to be cautious when applying diagnostic test studies. Fam. Pract.25, 390–396 (2008). [DOI] [PubMed] [Google Scholar]
289.Hernán, M. A. & Robins, J. M. Using big data to emulate a target trial when a randomized trial is not available. Am. J. Epidemiol.183, 758–764 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
290.Poldrack, R. A. et al. Scanning the horizon: towards transparent and reproducible neuroimaging research. Nat. Rev. Neurosci.18, 115–126 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
291.Wiens, J. et al. Do no harm: a roadmap for responsible machine learning for health care. Nat. Med.25, 1337–1340 (2019). [DOI] [PubMed] [Google Scholar]
292.Kim, J. et al. Deep learning-based long-term risk evaluation of incident type 2 diabetes using electrocardiogram in a non-diabetic population: a retrospective, multicentre study. EClinicalMedicine68, 102445 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
293.Fraser, R. A., Walker, R. J., Campbell, J. A., Ekwunife, O. & Egede, L. E. Integration of artificial intelligence and wearable technology in the management of diabetes and prediabetes. NPJ Digit. Med.8, 687 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
294.Lee, H. et al. External validation of ECG artificial intelligence for emergency and cardiac assessment across a large-scale U.S. healthcare system. NPJ Digit. Med. 10.1038/s41746-026-02682-7 (2026). [DOI] [PMC free article] [PubMed]
295.FDA-NIH Biomarker Working Group. BEST (Biomarkers, EndpointS, and other Tools) Resource (Food and Drug Administration (US); FDA-NIH Biomarker Working Group, 2016). [PubMed]
296.Ciani, O. et al. Time to review the role of surrogate end points in health policy: state of the art and the way forward. Value Health20, 487–495 (2017). [DOI] [PubMed] [Google Scholar]
297.Chandrabhatla, A. S. et al. Artificial intelligence and machine learning in the diagnosis and management of stroke: a narrative review of United States Food and Drug Administration-approved technologies. J. Clin. Med.12, 3755 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
298.Carvalho, E. et al. Predetermined change control plans: guiding principles for advancing safe, effective, and high-quality AI-ML technologies. JMIR AI4, e76854 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
299.Liu, X., Cruz Rivera, S., Moher, D. & Calvert, M. J. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Lancet Digit. Health2, e537–e548 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
300.Rivera, S. C., Liu, X., Chan, A. W., Denniston, A. K. & Calvert, M. J. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI Extension. BMJ370, m3210 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
301.Aboy, M., Minssen, T. & Vayena, E. Navigating the EU AI Act: implications for regulated digital medical products. NPJ Digit. Med.7, 237 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
302.Yuan, H. Toward real-world deployment of machine learning for health care: external validation, continual monitoring, and randomized clinical trials. Health Care Sci.3, 360–364 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

All supporting data are included in the article.

[CR1] 1.Ahmad, A., Imran, M. & Ahsan, H. Biomarkers as biomedical bioindicators: approaches and techniques for the detection, analysis, and validation of novel biomarkers of diseases. Pharmaceutics15, 1630 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR2] 2.Drucker, E. & Krapfenbauer, K. Pitfalls and limitations in translation from biomarker discovery to clinical utility in predictive and personalised medicine. EPMA J.4, 7 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR3] 3.Zhao, Q. et al. Applications and challenges of biomarker-based predictive models in proactive health management. Front. Public Health13, 1633487 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR4] 4.Ioannidis, J. P. Why most published research findings are false. PLoS Med.2, e124 (2005). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR5] 5.Rifai, N., Gillette, M. A. & Carr, S. A. Protein biomarker discovery and validation: the long and uncertain path to clinical utility. Nat. Biotechnol.24, 971–983 (2006). [DOI] [PubMed] [Google Scholar]

[CR6] 6.AbdulRaheem, Y. Statistical significance versus clinical relevance: key considerations in interpretation medical research data. Indian J. Community Med.49, 791–795 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR7] 7.Parikh, C. R. & Thiessen-Philbrook, H. Key concepts and limitations of statistical methods for evaluating biomarkers of kidney disease. J. Am. Soc. Nephrol.25, 1621–1629 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR8] 8.Cook, N. R. Methods for evaluating novel biomarkers: a new paradigm. Int. J. Clin. Pract.64, 1723–1727 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR9] 9.Poste, G. Bring on the biomarkers. Nature469, 156–157 (2011). [DOI] [PubMed] [Google Scholar]

[CR10] 10.Kalluri, R. & LeBleu, V. S. The biology, function, and biomedical applications of exosomes. Science367, eaau6977 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR11] 11.Hasin, Y., Seldin, M. & Lusis, A. Multi-omics approaches to disease. Genome Biol.18, 83 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR12] 12.Gillies, R. J., Kinahan, P. E. & Hricak, H. Radiomics: images are more than pictures, they are data. Radiology278, 563–577 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR13] 13.Lambin, P. et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur. J. Cancer48, 441–446 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR14] 14.Aerts, H. J. et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat. Commun.5, 4006 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR15] 15.Fu, W., Ding, J., Gao, K., Ma, S. & Tian, L. A likelihood ratio test on temporal trends in age-period-cohort models with applications to the disparities of heart disease mortality among US populations and comparison with Japan. Stat. Med.40, 668–689 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR16] 16.Bu, K. et al. Identifying correlations driven by influential observations in large datasets. Brief. Bioinform.23, bbab482 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR17] 17.Al-Ewaidat, O. A. & Naffaa, M. M. Emerging AI- and biomarker-driven precision medicine in autoimmune rheumatic diseases: from diagnostics to therapeutic decision-making. Rheumato5, 17 (2025). [Google Scholar]

[CR18] 18.Awari, A. et al. Obesity biomarkers: exploring factors, ramification, machine learning, and AI-unveiling insights in health research. MedComm6, e70169 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR19] 19.Carletti, M. et al. Multimodal AI correlates of glucose spikes in people with normal glucose regulation, pre-diabetes and type 2 diabetes. Nat. Med.31, 3121–3127 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR20] 20.Rajpurkar, P., Chen, E., Banerjee, O. & Topol, E. J. AI in health and medicine. Nat. Med.28, 31–38 (2022). [DOI] [PubMed] [Google Scholar]

[CR21] 21.Topol, E. J. High-performance medicine: the convergence of human and artificial intelligence. Nat. Med.25, 44–56 (2019). [DOI] [PubMed] [Google Scholar]

[CR22] 22.Norgeot, B. et al. Minimum information about clinical artificial intelligence modeling: the MI-CLAIM checklist. Nat. Med.26, 1320–1324 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR23] 23.Haibe-Kains, B. et al. Transparency and reproducibility in artificial intelligence. Nature586, E14–E16 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR24] 24.Diniz, M. A. Statistical methods for validation of predictive models. J. Nucl. Cardiol.29, 3248–3255 (2022). [DOI] [PubMed] [Google Scholar]

[CR25] 25.Vickers, A. J. & Elkin, E. B. Decision curve analysis: a novel method for evaluating prediction models. Med. Decis. Mak.26, 565–574 (2006). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR26] 26.Manolio, T. A. et al. Finding the missing heritability of complex diseases. Nature461, 747–753 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR27] 27.Klarin, D. & Natarajan, P. Clinical utility of polygenic risk scores for coronary artery disease. Nat. Rev. Cardiol.19, 291–301 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR28] 28.Roadmap Epigenomics Consortium et al. Integrative analysis of 111 reference human epigenomes. Nature518, 317–330 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR29] 29.Nicholson, J. K. et al. Metabolic phenotyping in clinical and surgical environments. Nature491, 384–392 (2012). [DOI] [PubMed] [Google Scholar]

[CR30] 30.Lin, C. et al. Metabolomics for clinical biomarker discovery and therapeutic target identification. Molecules29, 2198 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR31] 31.Qiu, S. et al. Small molecule metabolites: discovery of biomarkers and therapeutic targets. Signal Transduct. Target. Ther.8, 132 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR32] 32.Ghosh, S., Zhao, X., Alim, M., Brudno, M. & Bhat, M. Artificial intelligence applied to ‘omics data in liver disease: towards a personalised approach for diagnosis, prognosis and treatment. Gut74, 295–311 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR33] 33.Cheerla, A. & Gevaert, O. Deep learning with multimodal representation for pancancer prognosis prediction. Bioinformatics35, i446–i454 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR34] 34.Ahmed, I. et al. Plasma multi-omics and machine learning reveal predictive biomarkers for type 2 diabetes and retinopathy in Qatar Biobank Cohort. J. Transl. Med.23, 1159 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR35] 35.Liu, J. et al. The role of multi-omics in biomarker discovery, diagnosis, prognosis, and therapeutic monitoring of tissue repair and regeneration processes. J. Orthop. Transl.54, 131–151 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR36] 36.Piwecka, M., Rajewsky, N. & Rybak-Wolf, A. Single-cell and spatial transcriptomics: deciphering brain complexity in health and disease. Nat. Rev. Neurol.19, 346–362 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR37] 37.Aebersold, R. & Mann, M. Mass-spectrometric exploration of proteome structure and function. Nature537, 347–355 (2016). [DOI] [PubMed] [Google Scholar]

[CR38] 38.Stuart, T. & Satija, R. Integrative single-cell analysis. Nat. Rev. Genet.20, 257–272 (2019). [DOI] [PubMed] [Google Scholar]

[CR39] 39.Boxer, E. et al. Emerging clinical applications of single-cell RNA sequencing in oncology. Nat. Rev. Clin. Oncol.22, 315–326 (2025). [DOI] [PubMed] [Google Scholar]

[CR40] 40.Tirosh, I. & Suvà, M. L. Cancer cell states: lessons from ten years of single-cell RNA-sequencing of human tumors. Cancer Cell42, 1497–1506 (2024). [DOI] [PubMed] [Google Scholar]

[CR41] 41.Kleshchevnikov, V. et al. Cell2location maps fine-grained cell types in spatial transcriptomics. Nat. Biotechnol.40, 661–671 (2022). [DOI] [PubMed] [Google Scholar]

[CR42] 42.Dezem, F. S., Arjumand, W., DuBose, H., Morosini, N. S. & Plummer, J. Spatially resolved single-cell omics: methods, challenges, and future perspectives. Annu. Rev. Biomed. Data Sci.7, 131–153 (2024). [DOI] [PubMed] [Google Scholar]

[CR43] 43.De Jonghe, J. et al. scTrends: a living review of commercial single-cell and spatial ‘omic technologies. Cell Genom.4, 100723 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR44] 44.Si, Y., Lee, J. S., Jun, G., Kang, H. M. & Lee, J. H. Spatial omics enters the microscopic arena: opportunities and challenges. Trends Genet.41, 774–787 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR45] 45.Dinc, R. & Ardic, N. Mapping the immune environment: spatiotemporal dynamics in cardiovascular events. Front. Immunol.17, 1809941 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR46] 46.Yiu, T. et al. Transformative advances in single-cell omics: a comprehensive review of foundation models, multimodal integration and computational ecosystems. J. Transl. Med.23, 1176 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR47] 47.Baião, A. R. et al. A technical review of multi-omics data integration methods: from classical statistical to deep generative approaches. Brief. Bioinform.26, bbaf355 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR48] 48.Robins, H. S. et al. Comprehensive assessment of T-cell receptor beta-chain diversity in alphabeta T cells. Blood114, 4099–4107 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR49] 49.Wigerblad, G. & Kaplan, M. J. Neutrophil extracellular traps in systemic autoimmune and autoinflammatory diseases. Nat. Rev. Immunol.23, 274–288 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR50] 50.Ardic, N. & Dinc, R. Relationship between neutrophil extracellular traps and venous thromboembolism: pathophysiological and therapeutic role. Br. J. Hosp. Med.86, 1–15 (2025). [DOI] [PubMed] [Google Scholar]

[CR51] 51.Busarello, E. et al. Cell Marker Accordion: interpretable single-cell and spatial omics annotation in health and disease. Nat. Commun.16, 5399 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR52] 52.Corcoran, R. B. & Chabner, B. A. Application of cell-free DNA analysis to cancer treatment. N. Engl. J. Med.379, 1754–1765 (2018). [DOI] [PubMed] [Google Scholar]

[CR53] 53.Wan, J. C. M. et al. Liquid biopsies come of age: towards implementation of circulating tumour DNA. Nat. Rev. Cancer17, 223–238 (2017). [DOI] [PubMed] [Google Scholar]

[CR54] 54.Crosby, D. et al. Early detection of cancer. Science375, eaay9040 (2022). [DOI] [PubMed] [Google Scholar]

[CR55] 55.Li, F. Q. & Cui, J. W. Circulating tumor DNA-minimal residual disease: an up-and-coming nova in resectable non-small-cell lung cancer. Crit. Rev. Oncol. Hematol.179, 103800 (2022). [DOI] [PubMed] [Google Scholar]

[CR56] 56.Ma, L. et al. Liquid biopsy in cancer current: status, challenges and future prospects. Signal Transduct. Target. Ther.9, 336 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR57] 57.Qureshi, Z. et al. Liquid biopsies for early detection and monitoring of cancer: advances, challenges, and future directions. Ann. Med. Surg.87, 3244–3253 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR58] 58.Pantel, K. & Alix-Panabières, C. Minimal residual disease as a target for liquid biopsy in patients with solid tumours. Nat. Rev. Clin. Oncol.22, 65–77 (2025). [DOI] [PubMed] [Google Scholar]

[CR59] 59.Boukouris, A. E. et al. A comprehensive overview of minimal residual disease in the management of early-stage and locally advanced non-small cell lung cancer. NPJ Precis. Oncol.9, 178 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR60] 60.Hoang, T., Choi, M. K., Oh, J. H. & Kim, J. Utility of circulating tumor DNA to detect minimal residual disease in colorectal cancer: a systematic review and network meta-analysis. Int. J. Cancer157, 722–740 (2025). [DOI] [PubMed] [Google Scholar]

[CR61] 61.Negro, S. et al. Circulating tumor DNA as a real-time biomarker for minimal residual disease and recurrence prediction in stage II colorectal cancer: a systematic review and meta-analysis. Int. J. Mol. Sci.26, 2486 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR62] 62.Geyer, P. E., Holdt, L. M., Teupser, D. & Mann, M. Revisiting biomarker discovery by plasma proteomics. Mol. Syst. Biol.13, 942 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR63] 63.Ek, W. E., Karlsson, T., Höglund, J., Rask-Andersen, M. & Johansson, Å Causal effects of inflammatory protein biomarkers on inflammatory diseases. Sci. Adv.7, eabl4359 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR64] 64.Moreno-Torres, M. et al. Factors that influence the quality of metabolomics data in in vitro cell toxicity studies: a systematic survey. Sci. Rep.11, 22119 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR65] 65.Kumar, M. A. et al. Extracellular vesicles as tools and targets in therapy for diseases. Signal Transduct. Target. Ther.9, 27 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR66] 66.Dinc, R. & Ardic, N. Artificial intelligence for exosomal biomarker discovery for cardiovascular diseases: multi-omics integration, reproducibility, and translational prospects. Cells15, 304 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR67] 67.Welsh, J. A. et al. Minimal information for studies of extracellular vesicles (MISEV2023): from basic to advanced approaches. J. Extracell. Vesicles13, e12404 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR68] 68.Théry, C. et al. Minimal information for studies of extracellular vesicles 2018 (MISEV2018): a position statement of the International Society for Extracellular Vesicles and update of the MISEV2014 guidelines. J. Extracell. Vesicles7, 1535750 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR69] 69.Mateescu, B. et al. Obstacles and opportunities in the functional analysis of extracellular vesicle RNA: an ISEV position paper. J. Extracell. Vesicles6, 1286095 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR70] 70.Darci-Maher, N. et al. Cross-tissue omics analysis discovers ten adipose genes encoding secreted proteins in obesity-related non-alcoholic fatty liver disease. EBioMedicine92, 104620 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR71] 71.Li, H. et al. Identification and validation of core biomarkers for sepsis: a comprehensive analysis using bioinformatics and machine learning. Front. Immunol.16, 1700704 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR72] 72.Xu, L., Li, J. & Gong, W. Applications of machine learning-assisted extracellular vesicles analysis technology in tumor diagnosis. Comput. Struct. Biotechnol. J.27, 2460–2472 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR73] 73.Campanella, G. et al. Real-world deployment of a fine-tuned pathology foundation model for lung cancer biomarker detection. Nat. Med.31, 3002–3010 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR74] 74.Nicke, T. et al. Tissue concepts: supervised foundation models in computational pathology. Comput. Biol. Med.186, 109621 (2025). [DOI] [PubMed] [Google Scholar]

[CR75] 75.Li, Z. et al. AI-enabled virtual spatial proteomics from histopathology for interpretable biomarker discovery in lung cancer. Nat. Med.32, 231–244 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR76] 76.Pai, S. et al. Foundation model for cancer imaging biomarkers. Nat. Mach. Intell.6, 354–367 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR77] 77.Ligero, M. et al. A whirl of radiomics-based biomarkers in cancer immunotherapy, why is large scale validation still lacking? NPJ Precis. Oncol.8, 42 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR78] 78.Coravos, A. et al. Digital medicine: a primer on measurement. Digit. Biomark.3, 31–71 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR79] 79.Bent, B. et al. Engineering digital biomarkers of interstitial glucose from noninvasive smartwatches. NPJ Digit. Med.4, 89 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR80] 80.Perez, M. V. et al. Large-scale assessment of a smartwatch to identify atrial fibrillation. N. Engl. J. Med.381, 1909–1917 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR81] 81.Zahedani, A. D. et al. Digital health application integrating wearable data and behavioral patterns improves metabolic health. NPJ Digit. Med.6, 216 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR82] 82.Huang, X. et al. Digital biomarkers for interstitial glucose prediction in healthy individuals using wearables and machine learning. Sci. Rep.15, 30164 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR83] 83.Macias Alonso, A. K., Hirt, J., Woelfle, T., Janiaud, P. & Hemkens, L. G. Definitions of digital biomarkers: a systematic mapping of the biomedical literature. BMJ Health Care Inf.31, e100914 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR84] 84.Menden, M. P. et al. Machine learning prediction of cancer cell sensitivity to drugs based on genomic and chemical properties. PLoS ONE8, e61318 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR85] 85.Ng, S., Masarone, S., Watson, D. & Barnes, M. R. The benefits and pitfalls of machine learning for biomarker discovery. Cell Tissue Res.394, 17–31 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR86] 86.Esteva, A. et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature542, 115–118 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR87] 87.Bera, K., Schalper, K. A., Rimm, D. L., Velcheti, V. & Madabhushi, A. Artificial intelligence in digital pathology: new tools for diagnosis and precision oncology. Nat. Rev. Clin. Oncol.16, 703–715 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR88] 88.Richter, T., Bahrami, M., Xia, Y., Fischer, D. S. & Theis, F. J. Delineating the effective use of self-supervised learning in single-cell genomics. Nat. Mach. Intell.7, 68–78 (2025). [Google Scholar]

[CR89] 89.Acosta, J. N., Falcone, G. J., Rajpurkar, P. & Topol, E. J. Multimodal biomedical AI. Nat. Med.28, 1773–1784 (2022). [DOI] [PubMed] [Google Scholar]

[CR90] 90.Lipkova, J. et al. Artificial intelligence for multimodal data integration in oncology. Cancer Cell40, 1095–1110 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR91] 91.Steyaert, S. et al. Multimodal data fusion for cancer biomarker discovery with deep learning. Nat. Mach. Intell.5, 351–362 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR92] 92.Krones, F., Marikkar, U., Parsons, G., Szmul, A. & Mahdi, A. Review of multimodal machine learning approaches in healthcare. Inf. Fusion114, 102690 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR93] 93.Chen, R. J. et al. Pan-cancer integrative histology-genomic analysis via multimodal deep learning. Cancer Cell40, 865–878.e6 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR94] 94.Mataraso, S. J. et al. A machine learning approach to leveraging electronic health records for enhanced omics analysis. Nat. Mach. Intell.7, 293–306 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR95] 95.Camacho, D. M., Collins, K. M., Powers, R. K., Costello, J. C. & Collins, J. J. Next-generation machine learning for biological networks. Cell173, 1581–1592 (2018). [DOI] [PubMed] [Google Scholar]

[CR96] 96.Lin, J. et al. Chemokine-like factor-like MARVEL transmembrane domain-containing 1 identified as a novel target protein for immune dysregulation in sepsis: a machine learning and molecular dynamics framework for diagnostic biomarkers and therapeutic exploration. Int. J. Biol. Macromol.319, 145276 (2025). [DOI] [PubMed] [Google Scholar]

[CR97] 97.Xie, W. et al. A novel biomarker selection method combining graph neural network and gene relationships applied to microarray data. BMC Bioinforma.23, 303 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR98] 98.Dharanipragada, P. et al. Blocking genomic instability prevents acquired resistance to MAPK inhibitor therapy in melanoma. Cancer Discov.13, 880–909 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR99] 99.Dalla-Torre, H. et al. Nucleotide Transformer: building and evaluating robust foundation models for human genomics. Nat. Methods22, 287–297 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR100] 100.de Almeida, B., Lopez, M. & Pierrot, T. Generalized AI models for genomics applications. Nat. Methods22, 231–232 (2025). [DOI] [PubMed] [Google Scholar]

[CR101] 101.Guo, F. et al. Foundation models in bioinformatics. Natl. Sci. Rev.12, nwaf028 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR102] 102.Topol, E. J. Learning the language of life with AI. Science387, eadv4414 (2025). [DOI] [PubMed] [Google Scholar]

[CR103] 103.Vorontsov, E. et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nat. Med.30, 2924–2935 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR104] 104.Campanella, G. et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat. Med.25, 1301–1309 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR105] 105.Ding, T. et al. A multimodal whole-slide foundation model for pathology. Nat. Med.31, 3749–3761 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR106] 106.Xu, H. et al. A whole-slide foundation model for digital pathology from real-world data. Nature630, 181–188 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR107] 107.Moor, M. et al. Foundation models for generalist medical artificial intelligence. Nature616, 259–265 (2023). [DOI] [PubMed] [Google Scholar]

[CR108] 108.Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat. Mach. Intell.1, 206–215 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR109] 109.Budhkar, A., Song, Q., Su, J. & Zhang, X. Demystifying the black box: a survey on explainable artificial intelligence (XAI) in bioinformatics. Comput. Struct. Biotechnol. J.27, 346–359 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR110] 110.Zhang, K., Wang, D., Lin, F., Xie, J. & Zhou, W. A comprehensive review of explainable artificial intelligence in healthcare methods, evaluation, and clinical integration. iScience29, 115026 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR111] 111.Toussaint, P. A. et al. Explainable artificial intelligence for omics data: a systematic mapping study. Brief. Bioinform.25, bbad453 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR112] 112.Schölkopf, B. et al. Toward causal representation learning. Proc. IEEE109, 612–634 (2021). [Google Scholar]

[CR113] 113.Pearl, J. An introduction to causal inference. Int. J. Biostat.6, 7 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR114] 114.Javaid, H. et al. The impact of artificial intelligence on biomarker discovery. Emerg. Top. Life Sci.8, 89–105 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR115] 115.Prentice, R. L. Surrogate endpoints in clinical trials: definition and operational criteria. Stat. Med.8, 431–440 (1989). [DOI] [PubMed] [Google Scholar]

[CR116] 116.Gao, S. et al. Modeling drug mechanism of action with large scale gene-expression profiles using GPAR, an artificial intelligence platform. BMC Bioinforma.22, 17 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR117] 117.García-Campos, M. A., Espinal-Enríquez, J. & Hernández-Lemus, E. Pathway analysis: state of the art. Front. Physiol.6, 383 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR118] 118.Mubeen, S. et al. The impact of pathway database choice on statistical enrichment analysis and predictive modeling. Front. Genet.10, 1203 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR119] 119.Zhou, J. L. et al. Analysis of single-cell CRISPR perturbations indicates that enhancers predominantly act multiplicatively. Cell Genom.4, 100672 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR120] 120.Bjerregaard, A., Prada-Luengo, I., Das, V. & Krogh, A. What do single-cell models already know about perturbations? Genes16, 1439 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR121] 121.Egbon, O. A., Hickey, J. W. & Anchang, B. Fusion of spatiotemporal and network models to prioritize multiscale effects in single-cell perturbations. Brief. Bioinform.26, bbaf277 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR122] 122.Kleandrova, V. V., Cordeiro, M. N. D. S. & Speck-Planche, A. Perturbation-theory machine learning for multi-target drug discovery in modern anticancer research. Curr. Issues Mol. Biol.47, 301 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR123] 123.Davey Smith, G. & Hemani, G. Mendelian randomization: genetic anchors for causal inference in epidemiological studies. Hum. Mol. Genet.23, R89–R98 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR124] 124.Mo, Z., Zhang, Z., Miao, Q. & Tsui, K. L. Sparsity-constrained invariant risk minimization for domain generalization with application to machinery fault diagnosis modeling. IEEE Trans. Cybern.54, 1547–1559 (2024). [DOI] [PubMed] [Google Scholar]

[CR125] 125.Ni, S. et al. Identifying compound-protein interactions with knowledge graph embedding of perturbation transcriptomics. Cell Genom.4, 100655 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR126] 126.Crespi, E. et al. Resolving the rules of robustness and resilience in biology across scales. Integr. Comp. Biol.61, 2163–2179 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR127] 127.Wood, L. B., Winslow, A. R. & Strasser, S. D. Systems biology of neurodegenerative diseases. Integr. Biol.7, 758–775 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR128] 128.Ocana, A. Integrating artificial intelligence in drug discovery and early drug development: a transformative approach. Biomark. Res.13, 45 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR129] 129.Arango-Argoty, G. et al. AI-driven predictive biomarker discovery with contrastive learning to improve clinical trial outcomes. Cancer Cell43, 875–890.e8 (2025). [DOI] [PubMed] [Google Scholar]

[CR130] 130.Collins, G. S. et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ385, e078378 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR131] 131.Libby, P., Ridker, P. M. & Hansson, G. K. Inflammation in atherosclerosis: from pathophysiology to practice. J. Am. Coll. Cardiol.54, 2129–2138 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR132] 132.Firestein, G. S. & McInnes, I. B. Immunopathogenesis of rheumatoid arthritis. Immunity46, 183–196 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR133] 133.Fountzilas, E., Tsimberidou, A. M., Vo, H. H. & Kurzrock, R. Clinical trial design in the era of precision medicine. Genome Med.14, 101 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR134] 134.Jørgensen, J. T. Companion and complementary diagnostics: clinical and regulatory perspectives. Trends Cancer2, 706–712 (2016). [DOI] [PubMed] [Google Scholar]

[CR135] 135.Kern, S. E. Why your new cancer biomarker may never work: recurrent patterns and remarkable diversity in biomarker failures. Cancer Res.72, 6097–6101 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR136] 136.Ernst, S. M. et al. Utilizing ctDNA to discover mechanisms of resistance to targeted therapies in patients with metastatic NSCLC: towards more informative trials. Nat. Rev. Clin. Oncol.22, 371–378 (2025). [DOI] [PubMed] [Google Scholar]

[CR137] 137.Papayannopoulos, V. Neutrophil extracellular traps in immunity and disease. Nat. Rev. Immunol.18, 134–147 (2018). [DOI] [PubMed] [Google Scholar]

[CR138] 138.Bonaventura, A. et al. Endothelial dysfunction and immunothrombosis as key pathogenic mechanisms in COVID-19. Nat. Rev. Immunol.21, 319–329 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR139] 139.Finlayson, S. G. et al. The clinician and dataset shift in artificial intelligence. N. Engl. J. Med.385, 283–286 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR140] 140.Roberts, M. et al. Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and CT scans. Nat. Mach. Intell.3, 199–217 (2021). [Google Scholar]

[CR141] 141.Tang, S., Yuan, K. & Chen, L. Molecular biomarkers, network biomarkers, and dynamic network biomarkers for diagnosis and prediction of rare diseases. Fundam. Res.2, 894–902 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR142] 142.Qu, N. et al. Multimodal interpretable deep learning for transcriptome-informed precision oncology and drug mechanism analysis. NPJ Digit. Med. 10.1038/s41746-026-02735-x (2026). [DOI] [PMC free article] [PubMed]

[CR143] 143.Subramanian, I., Verma, S., Kumar, S., Jere, A. & Anamika, K. Multi-omics data integration, interpretation, and its application. Bioinform. Biol. Insights14, 1177932219899051 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR144] 144.Galindez, G., Sadegh, S., Baumbach, J., Kacprowski, T. & List, M. Network-based approaches for modeling disease regulation and progression. Comput. Struct. Biotechnol. J.21, 780–795 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR145] 145.Liu, Y., Zhu, K., Peng, W., Liu, Z. & Mao, X. Multi-omics and artificial intelligence for precision drug discovery and potential clinical applications. Signal Transduct. Target. Ther.11, 210 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR146] 146.Ben-Jaafar, A. et al. Artificial intelligence-based biomarkers for the diagnosis and treatment of neurological conditions: a narrative review. Mol. Brain19, 26 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR147] 147.Moons, K. G. M. et al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ388, e082505 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR148] 148.Valous, N. A. et al. Graph machine learning for integrated multi-omics analysis. Br. J. Cancer131, 205–211 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR149] 149.Rodriguez-Mier, P. et al. Unifying multi-sample network inference from prior knowledge and omics data with CORNETO. Nat. Mach. Intell.7, 1168–1186 (2025). [Google Scholar]

[CR150] 150.Wang, K. et al. TG468: a text graph convolutional network for predicting clinical response to immune checkpoint inhibitor therapy. Brief. Bioinform.25, bbae017 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR151] 151.Bonaventura, A. et al. The pathophysiological role of neutrophil extracellular traps in inflammatory diseases. Thromb. Haemost.118, 6–27 (2018). [DOI] [PubMed] [Google Scholar]

[CR152] 152.Marsiglia, J. Computationally guided high-throughput engineering of an anti-CRISPR protein for precise genome editing in human cells. Cell Rep. Methods4, 100882 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR153] 153.Van Calster, B., McLernon, D. J., van Smeden, M., Wynants, L. & Steyerberg, E. W. Calibration: the Achilles heel of predictive analytics. BMC Med.17, 230 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR154] 154.Piovani, D., Sokou, R., Tsantes, A. G., Vitello, A. S., & Bonovas, S. Optimizing clinical decision making with decision curve analysis: insights for clinical investigators. Healthcare11, 2244 (2023). [DOI] [PMC free article] [PubMed]

[CR155] 155.Riley, R. D. et al. Minimum sample size for developing a multivariable prediction model: PART II—binary and time-to-event outcomes. Stat. Med.38, 1276–1296 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR156] 156.Collins, G. S. et al. Protocol for development of a reporting guideline (TRIPOD-AI) and risk of bias tool (PROBAST-AI) for diagnostic and prognostic prediction model studies based on artificial intelligence. BMJ Open11, e048008 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR157] 157.van de Vegte, Y. J. et al. Genetic insights into resting heart rate and its role in cardiovascular disease. Nat. Commun.14, 4646 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR158] 158.Johnson, W. E., Li, C. & Rabinovic, A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics8, 118–127 (2007). [DOI] [PubMed] [Google Scholar]

[CR159] 159.Leek, J. T. et al. Tackling the widespread and critical impact of batch effects in high-throughput data. Nat. Rev. Genet.11, 733–739 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR160] 160.Hu, F. et al. DeepComBat: a statistically motivated, hyperparameter-robust, deep learning approach to harmonization of neuroimaging data. Hum. Brain Mapp.45, e26708 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR161] 161.Fortin, J. P. et al. Harmonization of cortical thickness measurements across scanners and sites. Neuroimage167, 104–120 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR162] 162.Wong, A. et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Intern. Med.181, 1065–1070 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR163] 163.Dhiman, P. et al. Risk of bias of prognostic models developed using machine learning: a systematic review in oncology. Diagn. Progn. Res.6, 13 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR164] 164.Hernández, B., Pennington, S. R. & Parnell, A. C. Bayesian methods for proteomic biomarker development. EuPA Open Proteom.9, 54–64 (2015). [Google Scholar]

[CR165] 165.Olsson, H. et al. Estimating diagnostic uncertainty in artificial intelligence assisted pathology using conformal prediction. Nat. Commun.13, 7761 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR166] 166.Seabroke, S. et al. Performance of stratified and subgrouped disproportionality analyses in spontaneous databases. Drug Saf.39, 355–364 (2016). [DOI] [PubMed] [Google Scholar]

[CR167] 167.Russo, C. A. M. & Selvatti, A. P. Bootstrap and rogue identification tests for phylogenetic analyses. Mol. Biol. Evol.35, 2327–2333 (2018). [DOI] [PubMed] [Google Scholar]

[CR168] 168.Hayes, D. F. Biomarker validation and testing. Mol. Oncol.9, 960–966 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR169] 169.Ou, F. S., Michiels, S., Shyr, Y., Adjei, A. A. & Oberg, A. L. Biomarker discovery and validation: statistical considerations. J. Thorac. Oncol.16, 537–545 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR170] 170.Vasey, B. et al. Reporting guideline for the early stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. BMJ377, e070904 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR171] 171.Shi, L. et al. The MicroArray Quality Control (MAQC)-II study of common practices for the development and validation of microarray-based predictive models. Nat. Biotechnol.28, 827–838 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR172] 172.MAQC Consortium The MicroArray Quality Control (MAQC) project shows inter- and intraplatform reproducibility of gene expression measurements. Nat. Biotechnol.24, 1151–1161 (2006). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR173] 173.SEQC/MAQC-III Consortium A comprehensive assessment of RNA-seq accuracy, reproducibility and information content by the Sequencing Quality Control Consortium. Nat. Biotechnol.32, 903–914 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR174] 174.Balikci, M. A., Njume, C. M. & Cakmak, A. BioMark: biomarker analysis tool. BMC Bioinform.27, 42 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR175] 175.Jiang, L. et al. Utilizing stability criteria in choosing feature selection methods yields reproducible results in microbiome data. Biometrics78, 1155–1167 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR176] 176.Traverso, A. et al. Powering responsible artificial intelligence with high-quality real-world data: the S-RACE platform for scalable, multi-specialty clinical research. NPJ Digit. Med.9, 6 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR177] 177.Davatzikos, C. et al. AI-based prognostic imaging biomarkers for precision neuro-oncology: the ReSPOND consortium. Neuro. Oncol.22, 886–888 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR178] 178.Ransohoff, D. F. Rules of evidence for cancer molecular-marker discovery and validation. Nat. Rev. Cancer4, 309–314 (2004). [DOI] [PubMed] [Google Scholar]

[CR179] 179.Pepe, M. S., Janes, H., Longton, G., Leisenring, W. & Newcomb, P. Limitations of the odds ratio in gauging the performance of a diagnostic, prognostic, or screening marker. Am. J. Epidemiol.159, 882–890 (2004). [DOI] [PubMed] [Google Scholar]

[CR180] 180.Obermeyer, Z., Powers, B., Vogeli, C. & Mullainathan, S. Dissecting racial bias in an algorithm used to manage the health of populations. Science366, 447–453 (2019). [DOI] [PubMed] [Google Scholar]

[CR181] 181.Mulherin, S. A. & Miller, W. C. Spectrum bias or spectrum effect? Subgroup variation in diagnostic test evaluation. Ann. Intern. Med.137, 598–602 (2002). [DOI] [PubMed] [Google Scholar]

[CR182] 182.Hatherley, J. A moving target in AI-assisted decision-making: dataset shift, model updating, and the problem of update opacity. Ethics Inf. Technol.27, 20 (2025). [Google Scholar]

[CR183] 183.Cagney, D. N. et al. The FDA NIH Biomarkers, EndpointS, and other Tools (BEST) resource in neuro-oncology. Neuro. Oncol.20, 1162–1172 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR184] 184.Agarwal, V. et al. Learning statistical models of phenotypes using noisy labeled training data. J. Am. Med. Inform. Assoc.23, 1166–1173 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR185] 185.Ioannidis, J. P. Why most discovered true associations are inflated. Epidemiology19, 640–648 (2008). [DOI] [PubMed] [Google Scholar]

[CR186] 186.Benjamens, S., Dhunnoo, P. & Meskó, B. The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database. NPJ Digit. Med.3, 118 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR187] 187.Mattes, W. B. & Goodsaid, F. Regulatory landscapes for biomarkers and diagnostic tests: Qualification, approval, and role in clinical practice. Exp. Biol. Med.243, 256–261 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR188] 188.Gallas, B. D. et al. FDA fosters innovative approaches in research, resources, and collaboration. Nat. Mach. Intell.4, 97–98 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR189] 189.U.S. Food and Drug Administration. Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan. https://www.fda.gov/media/145022/download (2021).

[CR190] 190.U.S. Food and Drug Administration. Marketing submission recommendations for a predetermined change control plan for artificial intelligence-enabled device software functions. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence (2025).

[CR191] 191.Hu, Y., Zou, Y., Qiao, L. & Lin, L. Integrative proteomic and metabolomic elucidation of cardiomyopathy with in vivo and in vitro models and clinical samples. Mol. Ther.32, 3288–3312 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR192] 192.Babu, M. & Snyder, M. Multi-omics profiling for health. Mol. Cell. Proteom.22, 100561 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR193] 193.Oikonomou, E. K. et al. Non-invasive detection of coronary inflammation using computed tomography and prediction of residual cardiovascular risk (the CRISP CT study): a post-hoc analysis of prospective outcome data. Lancet392, 929–939 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR194] 194.Attia, Z. I. et al. Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nat. Med.25, 70–74 (2019). [DOI] [PubMed] [Google Scholar]

[CR195] 195.Bai, W. et al. Automated cardiovascular magnetic resonance image analysis with fully convolutional networks. J. Cardiovasc. Magn. Reson.20, 65 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR196] 196.Shah, A. M. et al. Large scale plasma proteomics identifies novel proteins and protein networks associated with heart failure development. Nat. Commun.15, 528 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR197] 197.Faulkner, L. G., Howells, L. M., Pepper, C., Shaw, J. A. & Thomas, A. L. The utility of ctDNA in detecting minimal residual disease following curative surgery in colorectal cancer: a systematic review and meta-analysis. Br. J. Cancer128, 297–309 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR198] 198.Chidharla, A. et al. Circulating tumor DNA as a minimal residual disease assessment and recurrence risk in patients undergoing curative-intent resection with or without adjuvant chemotherapy in colorectal cancer: a systematic review and meta-analysis. Int. J. Mol. Sci.24, 10230 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR199] 199.Zheng, J., Qin, C., Wang, Q., Tian, D. & Chen, Z. Circulating tumour DNA-based molecular residual disease detection in resectable cancers: a systematic review and meta-analysis. EBioMedicine103, 105109 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR200] 200.Zhu, L. et al. Minimal residual disease (MRD) detection in solid tumors using circulating tumor DNA: a systematic review. Front. Genet.14, 1172108 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR201] 201.Martín-Arana, J. et al. Whole-exome tumor-agnostic ctDNA analysis enhances minimal residual disease detection and reveals relapse mechanisms in localized colon cancer. Nat. Cancer6, 1000–1016 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR202] 202.Martínez-Castedo, B. et al. Minimal residual disease in colorectal cancer. Tumor-informed versus tumor-agnostic approaches: unraveling the optimal strategy. Ann. Oncol.36, 263–276 (2025). [DOI] [PubMed] [Google Scholar]

[CR203] 203.Semenkovich, N. P., Szymanski, J. J., Earland, N., Chauhan, P. S. & Chaudhuri, A. A. Genomic approaches to cancer and minimal residual disease detection using circulating tumor DNA. J. Immunother. Cancer11, e006284 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR204] 204.Rosenlund, L. et al. ctDNA can detect minimal residual disease in curative treated non-small cell lung cancer patients using a tumor agnostic approach. Lung Cancer203, 108528 (2025). [DOI] [PubMed] [Google Scholar]

[CR205] 205.Eslinger, C., Wu, C. & Ahn, D. H. Minimal residual disease monitoring via ctDNA: a case report of Lynch syndrome with synchronous colorectal cancer and review of literature. J. Gastrointest. Oncol.15, 1341–1347 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR206] 206.Kather, J. N. et al. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nat. Med.25, 1054–1056 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR207] 207.Qian, L. et al. AI-empowered perturbation proteomics for complex biological systems. Cell Genom.4, 100691 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR208] 208.Sekar, P. K. & Veerabathiran, R. Systems immunology meets clinical translation: multi-omic approaches to predict therapy response in cancer and autoimmune disease. Clin. Immunol. Commun.9, 22 (2025). [Google Scholar]

[CR209] 209.Yuen, K. C. et al. High systemic and tumor-associated IL-8 correlates with reduced clinical benefit of PD-L1 blockade. Nat. Med.26, 693–698 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR210] 210.Kasi, P. M. et al. Impact of circulating tumor DNA-based detection of molecular residual disease on the conduct and design of clinical trials for solid tumors. JCO Precis. Oncol.6, e2100181 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR211] 211.Bartolomucci, A. et al. Circulating tumor DNA to monitor treatment response in solid tumors and advance precision oncology. NPJ Precis. Oncol.9, 84 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR212] 212.García de Yébenes Prous, M. J. & Carmona Ortells, L. Biomarkers: how to consolidate them in clinical practice. Reumatol. Clin.20, 386–391 (2024). [DOI] [PubMed] [Google Scholar]

[CR213] 213.Mattsson-Carlgren, N. et al. Longitudinal plasma p-tau217 is increased in early stages of Alzheimer’s disease. Brain143, 3234–3241 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR214] 214.Hansson, O., Blennow, K., Zetterberg, H. & Dage, J. Blood biomarkers for Alzheimer’s disease in clinical practice and trials. Nat. Aging3, 506–519 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR215] 215.Jack, C. R. et al. NIA-AA Research Framework: toward a biological definition of Alzheimer’s disease. Alzheimers Dement.14, 535–562 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR216] 216.Leuzy, A. et al. Tau PET imaging in neurodegenerative tauopathies—still a challenge. Mol. Psychiatry24, 1112–1134 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR217] 217.Mormino, E. C. et al. Tau PET imaging with 18F-PI-2620 in aging and neurodegenerative diseases. Eur. J. Nucl. Med. Mol. Imaging48, 2233–2244 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR218] 218.Ravi, D. et al. Deep learning for health informatics. IEEE J. Biomed. Health Inform.21, 4–21 (2017). [DOI] [PubMed] [Google Scholar]

[CR219] 219.Zhang, W., Xiao, D., Mao, Q. & Xia, H. Role of neuroinflammation in neurodegeneration development. Signal Transduct. Target. Ther.8, 267 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR220] 220.van der Kant, R., Goldstein, L. S. B. & Ossenkoppele, R. Amyloid-β-independent regulators of tau pathology in Alzheimer disease. Nat. Rev. Neurosci.21, 21–35 (2020). [DOI] [PubMed] [Google Scholar]

[CR221] 221.Serrano-Pozo, A., Frosch, M. P., Masliah, E. & Hyman, B. T. Neuropathological alterations in Alzheimer disease. Cold Spring Harb. Perspect. Med.1, a006189 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR222] 222.Schöll, M. et al. Challenges in the practical implementation of blood biomarkers for Alzheimer’s disease. Lancet Healthy Longev.5, 100630 (2024). [DOI] [PubMed] [Google Scholar]

[CR223] 223.Smillie, C. S. et al. Intra- and inter-cellular rewiring of the human colon during ulcerative colitis. Cell178, 714–730.e22 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR224] 224.Wang, L. et al. An atlas of single-cell eQTLs dissects autoimmune disease genes and identifies novel drug classes for treatment. Cell Genom.5, 100820 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR225] 225.Smolen, J. S. et al. Treating rheumatoid arthritis to target: recommendations of an international task force. Ann. Rheum. Dis.69, 631–637 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR226] 226.Hauser, S. L. & Cree, B. A. C. Treatment of multiple sclerosis: a review. Am. J. Med.133, 1380–1390.e2 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR227] 227.Tilg, H., Byrne, C. D. & Targher, G. NASH drug treatment development: challenges and lessons. Lancet Gastroenterol. Hepatol.8, 943–954 (2023). [DOI] [PubMed] [Google Scholar]

[CR228] 228.Thiele, M. et al. Opportunities and barriers in omics-based biomarker discovery for steatotic liver diseases. J. Hepatol.81, 345–359 (2024). [DOI] [PubMed] [Google Scholar]

[CR229] 229.Lazarus, J. V. et al. A call for doubling the diagnostic rate of at-risk metabolic dysfunction-associated steatohepatitis. Lancet Reg. Health Eur.54, 101320 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR230] 230.Noureddin, M. et al. Serum identification of at-risk MASH: the metabolomics-advanced steatohepatitis fibrosis score (MASEF). Hepatology79, 135–148 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR231] 231.Rinella, M. E. & Sookoian, S. From NAFLD to MASLD: updated naming and diagnosis criteria for fatty liver disease. J. Lipid Res.65, 100485 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR232] 232.Sergi, C. M. NAFLD (MASLD)/NASH (MASH): does it bother to label at all? A comprehensive narrative review. Int. J. Mol. Sci.25, 8462 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR233] 233.Park, I. G. et al. Gut microbiota-based machine-learning signature for the diagnosis of alcohol-associated and metabolic dysfunction-associated steatotic liver disease. Sci. Rep.14, 16122 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR234] 234.Sveinbjornsson, G. et al. A multi-omics study of non-alcoholic fatty liver disease. Nat. Genet.54, 1652–1663 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR235] 235.Verschuren, L. et al. Development of a novel non-invasive biomarker panel for hepatic fibrosis in MASLD. Nat. Commun.15, 4564 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR236] 236.Stratakis, N. et al. Multi-omic architecture, biological pathways, and prenatal determinants of childhood obesity and metabolic dysfunction. Nat. Commun.16, 654 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR237] 237.Chakaroun, R. M. et al. Multi-omic definition of metabolic obesity through adipose tissue-microbiome interactions. Nat. Med.32, 113–125 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR238] 238.Soták, M., Clark, M., Suur, B. E. & Börgeson, E. Inflammation and resolution in obesity. Nat. Rev. Endocrinol.21, 45–61 (2025). [DOI] [PubMed] [Google Scholar]

[CR239] 239.Rönn, T., Perfilyev, A., Oskolkov, N. & Ling, C. Predicting type 2 diabetes via machine learning integration of multiple omics from human pancreatic islets. Sci. Rep.14, 14637 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR240] 240.Dai, L. et al. A deep learning system for predicting time to progression of diabetic retinopathy. Nat. Med.30, 584–594 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR241] 241.Ratziu, V. et al. Confirmatory biomarker diagnostic studies are not needed when transitioning from NAFLD to MASLD. J. Hepatol.80, e51–e52 (2024). [DOI] [PubMed] [Google Scholar]

[CR242] 242.Chen, L., Zhang, X. & Shi, P. Recent advances in biomarkers for detection and diagnosis of sepsis and organ dysfunction: a comprehensive review. Eur. J. Med. Res.30, 1081 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR243] 243.de Nooijer, A. H., Pickkers, P., Netea, M. G. & Kox, M. Inflammatory biomarkers to predict the prognosis of acute bacterial and viral infections. J. Crit. Care78, 154360 (2023). [DOI] [PubMed] [Google Scholar]

[CR244] 244.Xiong, W., Zhan, Y., Xiao, R. & Liu, F. Advancing sepsis diagnosis and immunotherapy: machine learning-driven identification of stable molecular biomarkers and therapeutic targets. Sci. Rep.15, 8333 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR245] 245.Yang, J. et al. Machine learning based screening of biomarkers associated with cell death and immunosuppression of multiple life stages sepsis populations. Sci. Rep.15, 30302 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR246] 246.Rolfo, C. et al. Liquid biopsy for advanced NSCLC: a consensus statement from the International Association for the Study of Lung Cancer. J. Thorac. Oncol.16, 1647–1662 (2021). [DOI] [PubMed] [Google Scholar]

[CR247] 247.Sabari, J. K. et al. A prospective study of circulating tumor DNA to guide matched targeted therapy in lung cancers. J. Natl. Cancer Inst.111, 575–583 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR248] 248.Jiang, P. et al. Signatures of T cell dysfunction and exclusion predict cancer immunotherapy response. Nat. Med.24, 1550–1558 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR249] 249.Auslander, N. et al. Robust prediction of response to immune checkpoint blockade therapy in metastatic melanoma. Nat. Med.24, 1545–1549 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR250] 250.Berry, S. M., Carlin, B. P., Lee, J. J. & Muller, P. Bayesian Adaptive Methods for Clinical Trials (CRC Press, 2010).

[CR251] 251.Topalian, S. L., Taube, J. M., Anders, R. A. & Pardoll, D. M. Mechanism-driven biomarkers to guide immune checkpoint blockade in cancer therapy. Nat. Rev. Cancer16, 275–287 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR252] 252.Subbiah, V. & Kurzrock, R. Challenging standard-of-care paradigms in the precision oncology era. Trends Cancer4, 101–109 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR253] 253.Okusanya, O. O. et al. FDA-AACR strategies for optimizing dosages for oncology drug products: early-phase trials using innovative trial designs and biomarkers. Clin. Cancer Res.31, 4882–4890 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR254] 254.Fleming, T. R. & DeMets, D. L. Surrogate end points in clinical trials: are we being misled? Ann. Intern. Med.125, 605–613 (1996). [DOI] [PubMed] [Google Scholar]

[CR255] 255.Fleming, T. R. & Powers, J. H. Biomarkers and surrogate endpoints in clinical trials. Stat. Med.31, 2973–2984 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR256] 256.Moqri, M. et al. Validation of biomarkers of aging. Nat. Med.30, 360–372 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR257] 257.Sherman, R. E. et al. Real-world evidence—what is it and what can it tell us? N. Engl. J. Med.375, 2293–2297 (2016). [DOI] [PubMed] [Google Scholar]

[CR258] 258.Ohtsu, H., Shimomura, A. & Sase, K. Real-world evidence in cardio-oncology: what is it and what can it tell us? JACC CardioOncol.4, 95–97 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR259] 259.Griesinger, F., Cox, O., Sammon, C., Ramagopalan, S. V. & Popat, S. Health technology assessments and real-world evidence: tell us what you want, what you really, really want. J. Comp. Eff. Res.11, 297–299 (2022). [DOI] [PubMed] [Google Scholar]

[CR260] 260.Ardic, N. & Dinc, R. Artificial intelligence in healthcare: current regulatory landscape and future directions. Br. J. Hosp. Med.86, 1–21 (2025). [DOI] [PubMed] [Google Scholar]

[CR261] 261.U.S. Food and Drug Administration. Biomarker qualification: evidentiary framework. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/biomarker-qualification-evidentiary-framework (2018).

[CR262] 262.Coravos, A., Khozin, S. & Mandl, K. D. Developing and adopting safe and effective digital biomarkers to improve patient outcomes. NPJ Digit. Med.2, 14 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR263] 263.Price, W. N. II. & Cohen, I. G. Privacy in the age of medical big data. Nat. Med.25, 37–43 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR264] 264.Patel, A. P. et al. A multi-ancestry polygenic risk score improves risk prediction for coronary artery disease. Nat. Med.29, 1793–1803 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR265] 265.GBD 2023 Causes of Death Collaborators Global burden of 292 causes of death in 204 countries and territories and 660 subnational locations, 1990–2023: a systematic analysis for the Global Burden of Disease Study 2023. Lancet406, 1811–1872 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR266] 266.Alum, E. U. AI-driven biomarker discovery: enhancing precision in cancer diagnosis and prognosis. Discov. Oncol.16, 313 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR267] 267.Singh, Y. et al. Beyond post hoc explanations: a comprehensive framework for responsible artificial intelligence in medical imaging through transparency, interpretability, and explainability. Bioengineering12, 879 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR268] 268.Begley, C. G. & Ellis, L. M. Drug development: raise standards for preclinical cancer research. Nature483, 531–533 (2012). [DOI] [PubMed] [Google Scholar]

[CR269] 269.Laird, A. R. Large, open datasets for human connectomics research: considerations for reproducible and responsible data use. Neuroimage244, 118579 (2021). [DOI] [PubMed] [Google Scholar]

[CR270] 270.Fountzilas, E., Pearce, T., Baysal, M. A., Chakraborty, A. & Tsimberidou, A. M. Convergence of evolving artificial intelligence and machine learning techniques in precision oncology. NPJ Digit. Med.8, 75 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR271] 271.Mendez, K. M. et al. A roadmap to precision medicine through post-genomic electronic medical records. Nat. Commun.16, 1700 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR272] 272.Liu, Y., Dai, Y. & Wang, L. Spatial omics at the forefront: emerging technologies, analytical innovations, and clinical applications. Cancer Cell44, 24–49 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR273] 273.Zhang, Y. et al. Robust integration and annotation of single-cell and spatial omics data using interpretable gene programs. Cell Genom.6, 101105 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR274] 274.Van de Sande, B. et al. Applications of single-cell RNA sequencing in drug discovery and development. Nat. Rev. Drug Discov.22, 496–520 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR275] 275.Dasaradharami Reddy, K. & Gadekallu, T. R. A comprehensive survey on federated learning techniques for healthcare informatics. Comput. Intell. Neurosci.2023, 8393990 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR276] 276.Olawade, D. B. et al. Digital twin paradigm in diabetes prediction and management. Diab. Res. Clin. Pract.231, 113075 (2026). [DOI] [PubMed] [Google Scholar]

[CR277] 277.Viceconti, M. et al. In silico trials: verification, validation and uncertainty quantification of predictive models used in the regulatory evaluation of biomedical products. Methods185, 120–127 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR278] 278.Strimbu, K. & Tavel, J. A. What are biomarkers? Curr. Opin. HIV AIDS5, 463–466 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR279] 279.Chenchula, S., Paraskar, G., Krishna, S. & Chavan, M. Liquid biopsy in clinical practice: current evidence and future directions. In Liquid Biopsy in Cancer Management (eds Bhatt, S., Anitha, A., Chenchula, S. & Mishra, N.) 405–438 (Academic Press, 2026).

[CR280] 280.Ohara, S. & Suda, K. Current status and future challenges of liquid biopsy. Cells14, 2000 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR281] 281.Abràmoff, M. D., Lavin, P. T., Birch, M., Shah, N. & Folk, J. C. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. NPJ Digit. Med.1, 39 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR282] 282.Elijovich, L. et al. Automated emergent large vessel occlusion detection by artificial intelligence improves stroke workflow in a hub and spoke stroke system of care. J. Neurointerv. Surg.14, 704–708 (2022). [DOI] [PubMed] [Google Scholar]

[CR283] 283.Chilamkurthy, S. et al. Deep learning algorithms for detection of critical findings in head CT scans: a retrospective study. Lancet392, 2388–2396 (2018). [DOI] [PubMed] [Google Scholar]

[CR284] 284.Komorowski, M., Celi, L. A., Badawi, O., Gordon, A. C. & Faisal, A. A. The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care. Nat. Med.24, 1716–1720 (2018). [DOI] [PubMed] [Google Scholar]

[CR285] 285.Weikert, T. et al. Automated detection of pulmonary embolism in CT pulmonary angiograms using an AI-powered algorithm. Eur. Radiol.30, 6545–6553 (2020). [DOI] [PubMed] [Google Scholar]

[CR286] 286.Beam, A. L. & Kohane, I. S. Big data and machine learning in health care. JAMA319, 1317–1318 (2018). [DOI] [PubMed] [Google Scholar]

[CR287] 287.Esteva, A. et al. A guide to deep learning in healthcare. Nat. Med.25, 24–29 (2019). [DOI] [PubMed] [Google Scholar]

[CR288] 288.Willis, B. H. Spectrum bias—why clinicians need to be cautious when applying diagnostic test studies. Fam. Pract.25, 390–396 (2008). [DOI] [PubMed] [Google Scholar]

[CR289] 289.Hernán, M. A. & Robins, J. M. Using big data to emulate a target trial when a randomized trial is not available. Am. J. Epidemiol.183, 758–764 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR290] 290.Poldrack, R. A. et al. Scanning the horizon: towards transparent and reproducible neuroimaging research. Nat. Rev. Neurosci.18, 115–126 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR291] 291.Wiens, J. et al. Do no harm: a roadmap for responsible machine learning for health care. Nat. Med.25, 1337–1340 (2019). [DOI] [PubMed] [Google Scholar]

[CR292] 292.Kim, J. et al. Deep learning-based long-term risk evaluation of incident type 2 diabetes using electrocardiogram in a non-diabetic population: a retrospective, multicentre study. EClinicalMedicine68, 102445 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR293] 293.Fraser, R. A., Walker, R. J., Campbell, J. A., Ekwunife, O. & Egede, L. E. Integration of artificial intelligence and wearable technology in the management of diabetes and prediabetes. NPJ Digit. Med.8, 687 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR294] 294.Lee, H. et al. External validation of ECG artificial intelligence for emergency and cardiac assessment across a large-scale U.S. healthcare system. NPJ Digit. Med. 10.1038/s41746-026-02682-7 (2026). [DOI] [PMC free article] [PubMed]

[CR295] 295.FDA-NIH Biomarker Working Group. BEST (Biomarkers, EndpointS, and other Tools) Resource (Food and Drug Administration (US); FDA-NIH Biomarker Working Group, 2016). [PubMed]

[CR296] 296.Ciani, O. et al. Time to review the role of surrogate end points in health policy: state of the art and the way forward. Value Health20, 487–495 (2017). [DOI] [PubMed] [Google Scholar]

[CR297] 297.Chandrabhatla, A. S. et al. Artificial intelligence and machine learning in the diagnosis and management of stroke: a narrative review of United States Food and Drug Administration-approved technologies. J. Clin. Med.12, 3755 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR298] 298.Carvalho, E. et al. Predetermined change control plans: guiding principles for advancing safe, effective, and high-quality AI-ML technologies. JMIR AI4, e76854 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR299] 299.Liu, X., Cruz Rivera, S., Moher, D. & Calvert, M. J. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Lancet Digit. Health2, e537–e548 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR300] 300.Rivera, S. C., Liu, X., Chan, A. W., Denniston, A. K. & Calvert, M. J. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI Extension. BMJ370, m3210 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR301] 301.Aboy, M., Minssen, T. & Vayena, E. Navigating the EU AI Act: implications for regulated digital medical products. NPJ Digit. Med.7, 237 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR302] 302.Yuan, H. Toward real-world deployment of machine learning for health care: external validation, continual monitoring, and randomized clinical trials. Health Care Sci.3, 360–364 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

Artificial intelligence in biomarker discovery for diseases: diagnostic and therapeutic prospects

Rasit Dinc

Nurittin Ardic

Abstract

Introduction

Fig. 1.

Biomarker models and biological principles (inter-disease relationships)

Molecular biomarkers: genomics/epigenomics, transcriptomics, proteomics/posttranslational modifications, metabolomics/lipidomics

Cellular and immune biomarkers: immune repertoires, single-cell states, cytokine programs, and NETs

Fig. 2.

Circulating biomarkers: cfDNA/cfRNA, circulating proteins/metabolites, liquid biopsy paradigms

Extracellular vesicles/exosomes as biomarker carriers

Imaging biomarkers: radiomics, functional imaging, quantitative imaging processes

Digital biomarkers: wearable devices, passive sensing, digital phenotypes and linkage to outcomes

The “Biological plausibility” framework: distinguishing causal, mechanistic, and correlational biomarkers

Table 1.

Artificial intelligence methods for biomarker discovery

Classical machine learning for high-dimensional biomarker data

Deep representation learning and self-supervised learning for omic and single-cell data

Multimodal learning in omics, imaging, and clinical data

Graph neural networks and knowledge-guided learning

Foundation and large-scale pre-training in genomics and biomedicine

Generative and simulation-assisted approaches for biomarker mechanism and target discovery

Table 2.

Explainability, uncertainty, and the risk of “false mechanistic comfort”

Biomarker type and method selection as a translational objective function

Fig. 3.

Table 3.

Mechanistic AI: mapping biomarkers to pathways, targets, and therapeutic hypotheses

From correlational signatures to mechanistic hypotheses

Knowledge-guided and pathway-constrained neural architectures

Perturbation-based learning and causal signal identification

Causal representation learning and counterfactual biomarker modeling

Linking biomarker signatures to drug response and therapeutic classification

Limitations of mechanistic AI and the role of experimental validation

Signal transduction networks, pathway-state biomarkers, and therapeutic targeting

Why pathway states outperform static biomarkers

Fig. 4.

Signal transduction complexity and biomarker instability

AI approaches for pathway-state inference and network-constrained biomarker modeling

Neutrophil extracellular traps (NETs) as a pathway-state biomarker exemplar

Biological relevance of pathway-informed biomarkers

Translational workflow for AI biomarker development

Study design and dataset generation: prospective and retrospective, endpoints, and spectrum effects

Data quality control and harmonization across platforms and centers

Feature representation, leakage prevention, and evaluation separation

Model development: calibration, uncertainty, and subgroup performance

Biomarker selection and stability: reproducibility across cohorts and resampling

Biological validation cycle: functional evidence and orthogonal confirmation

Clinical validation and benefit: from AUC to decision impact

Reporting standards, reproducibility, and auditability

Data infrastructure and multi-center reproducibility ecosystems

Levels of evidence and verification: a unified grading framework (cross diseases)

Fig. 5.

Levels of evidence for AI-generated biomarkers

Table 4.

Common failure modes in AI biomarker translation

Table 5.

Generating biomarker evidence at the regulatory level

Table 6.

Cross-disease applications (synthesis, not catalog)

Table 7.

Cardiovascular diseases

Oncology (tumor-immune ecosystems, liquid biopsy signatures, response)

Neurodegenerative disorders (early diagnosis, progression, and multimodal integration)

Autoimmune and inflammatory diseases (immune-state biomarkers, therapeutic stratification)

Metabolic disorders (multimodal metabolic phenotyping and fibrosis-risk stratification)

Infectious diseases and host-response states (dynamic biomarkers and stress testing of translation)

Fig. 6.

Translational roles of AI biomarkers in therapeutic development and clinical decision-making processes

Fig. 7.

Biomarkers for study enrichment, adaptive designs, and outcome selection

Fig. 8.

Table 8.

Accompanying diagnosis: linking AI biomarkers to treatment selection

Biomarkers as surrogate endpoints: when do they succeed and why do they fail?

Table 9.

Longitudinal monitoring and dynamic treatment adaptation

Real-world evidence (RWE) and post-implementation monitoring