Recent advances in machine learning-enhanced extracellular vesicle omics for oncology

Abstract

Extracellular vesicles (EVs) are nanoscale, membrane-bound particles that carry nucleic acids, proteins, metabolites, and lipids. Their omics profiles can reflect tumor and microenvironmental states, making EVs a promising source of liquid biopsy biomarkers. Machine learning (ML) is well suited to EV omics because it can learn predictive signatures from high-dimensional, correlated, and sparse features and integrate complementary modalities. However, clinical translation is often hindered by EV-specific issues—including heterogeneous vesicle populations, isolation-dependent recovery of subtypes, and co-isolated particles. Few reviews comprehensively synthesize ML studies across EV transcriptomics, proteomics, metabolomics, and lipidomics in oncology. This review introduces EV fundamentals and provides a structured, in-depth evaluation of recent advances in ML-enhanced EV omics for early cancer detection, molecular subtyping, prognosis, and treatment response prediction. Key challenges—including data quality, model generalizability, algorithmic interpretability, ethical considerations, and standardization issues—are critically examined and distilled into practical recommendations for study design, validation, and reporting. Emerging directions include single-vesicle omics, higher-resolution EV profiling, interpretable multimodal fusion, and end-to-end pipelines that integrate EV multi-omics with ML. Together, these advances could help translate EV-based liquid biopsy into clinically useful tools for precision oncology.

Graphical Abstract

Introduction

Early detection, accurate risk stratification, and dynamic monitoring of therapeutic response remain critical unmet clinical needs in oncology [1]. Despite advances in imaging and tissue-based molecular profiling, many patients are still diagnosed at intermediate or advanced stages, when effective treatment options are limited and the risks of recurrence and metastasis are high [2]. Traditional serum tumor markers, such as alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), and various carbohydrate antigens, show limited sensitivity and specificity, especially in early-stage disease and in the presence of marked tumor heterogeneity [3]. These shortcomings have driven the development of minimally invasive liquid biopsy approaches, which detect tumor-derived signals in blood and other body fluids and enable repeated, non-invasive monitoring of tumor burden and biology [4]. Liquid biopsy can analyze several circulating components, including circulating tumor DNA (ctDNA), a tumor-derived fraction of cell-free DNA (cfDNA); circulating tumor cells (CTCs); tumor-educated platelets (TEPs); and extracellular vesicles (EVs) [5]. Among these, EVs—small membrane-bound vesicles released by most cell types—have attracted growing interest as versatile carriers of tumor-related information. EVs circulate stably in biofluids and transport proteins, RNAs, lipids, and metabolites that mirror the physiological and pathological states of their cells of origin [6]. Tumor-derived EVs contribute to local tumor growth, immune modulation, angiogenesis, pre-metastatic niche formation, and therapy resistance, making them promising diagnostic, prognostic, and predictive biomarkers [7]. With high-throughput assays, EV omics—including proteomics, transcriptomics, metabolomics, and lipidomics—enable comprehensive molecular profiling of EV cargo in clinical cohorts and support the discovery of clinically relevant biomarkers [8]. However, the complexity and heterogeneity of EV populations generate omics datasets that are high-dimensional, sparse, and noisy, with intricate correlation structures [9]. Variability in pre-analytical handling and analytical workflows, together with batch effects, further complicates robust identification of reliable biomarkers and predictive models using EV omics data [10].

Machine learning (ML) offers clear advantages for analyzing high-dimensional, complex datasets [11]. ML is now widely used in cancer research, particularly in precision oncology [12]. It helps assess risk, describe disease status, and track patients over time, and it has been applied across cancer types and diverse data sources, including clinical, imaging, and omics data. In EV omics, ML supports the discovery of multi-marker signatures, captures nonlinear relationships among EV features, and integrates EV-derived information with clinical and other multi-omics data [13]. Recent advances in model interpretability, calibration, and uncertainty quantification have further strengthened the translational potential of ML, helping to turn “black-box” predictions into mechanistically informative and clinically meaningful insights [14].

Despite this growing interest, most EV-focused reviews still address EV biology, isolation and enrichment techniques, or their general roles in liquid biopsy [15]. By contrast, reviews that specifically discuss artificial intelligence (AI) or ML in EV research usually emphasize EV-mediated drug delivery, surface-enhanced Raman scattering (SERS)-based or electrochemical detection platforms, or diagnostic applications in single cancer types [16]. Broader ML or liquid biopsy overviews primarily focus on bulk tumor tissue or circulating cell-free biomarkers and mention EVs only briefly [17]. As a result, comprehensive evaluations dedicated to machine learning-enhanced EV omics in oncology remain scarce. To address this gap, this review provides a structured and in-depth evaluation of recent advances in ML-enabled EV omics in oncology. First, this review introduces key concepts in EV biology and machine learning and synthesizes oncology studies that apply ML to EV-derived omics data, highlighting performance patterns and major limitations across cancer detection, molecular subtyping, prognosis, and treatment monitoring. Next, it discusses major methodological and translational challenges, including data quality, model generalizability, algorithmic explainability, ethical considerations, and standardization issues. On this basis, it proposes practical recommendations for study design, validation, and reporting. Finally, it outlines emerging directions in single-vesicle EV omics and high-resolution analysis, the development of interpretable multimodal ML methods, and the construction of end-to-end workflows that integrate EV multi-omics with ML for clinical translation.

EVs: fundamentals, omics and application overview

EVs are lipid bilayer-enclosed particles released by cells that lack self-replicative capacity [18]. They mediate the intercellular transfer of bioactive molecules (e.g., RNAs and proteins), thereby exerting important regulatory functions during tumor initiation, progression and therapeutic response [19]. The molecular cargo they carry largely reflects the molecular state of the parental cells and, being shielded by a lipid bilayer, remains relatively stable in biofluids. Consequently, EVs have emerged as a major focus of research as a novel medium for liquid biopsy and intercellular communication in oncology [20]. In this section, we briefly summarize EV biogenesis, isolation and characterization strategies, and their omics features and clinical applications in oncology. This overview lays the foundation for the subsequent discussion on the integration of EV omics and machine learning.

EV fundamentals

EVs comprise a highly heterogeneous and structurally diverse population of particles [21]. They are typically classified into three main subtypes: exosomes (originating from the endosomal system), microvesicles (generated by direct budding from the plasma membrane), and apoptotic bodies (formed during programmed cell death) [22]. Exosomes (approximately 30–150 nm in diameter) are generated as intraluminal vesicles within multivesicular bodies (MVB), which subsequently fuse with the plasma membrane to release them (Fig. 1a). In contrast, microvesicles (approximately 100–1000 nm) are formed by outward budding of the plasma membrane, a process driven by cytoskeletal remodeling and altered lipid distribution [23]. In addition to this structural diversity, EVs encapsulate a broad repertoire of bioactive molecules, principally proteins, nucleic acids, lipids and metabolites (Fig. 1b) [24]. Together, these cargo molecules recapitulate the physiological and pathological states of their cells of origin and mediate a wide range of intercellular communication processes [25]. To study these heterogeneous vesicle populations and their cargo for downstream omics analyses, robust and context-appropriate isolation strategies are essential. A range of EV isolation techniques has been established. Conventional approaches include differential ultracentrifugation [26], density gradient centrifugation [27], size-exclusion chromatography [28], ultrafiltration [29], polymer precipitation [30], and immunoaffinity-based capture [31]. In addition, methods such as microfluidics [32], tangential flow filtration [33], asymmetric flow field-flow fractionation [34], and anion-exchange chromatography have been introduced [35]. Because each method trades off purity, yield, throughput, and vesicle integrity and can shift the EV subpopulations and co-isolates recovered, isolation should be selected based on the biofluid, study aim, and the requirements of downstream EV omics assays [36]. EV characterization typically integrates assessment of both physical properties and molecular composition. Physical characterization includes particle analysis by nanoparticle tracking analysis for size and concentration measurements [37], as well as transmission electron microscopy for vesicle morphology [38]. Molecular characterization utilizes methods such as Western blotting [39], enzyme-linked immunosorbent assays (ELISA) [40], mass spectrometry [41], and real-time quantitative PCR to detect protein and nucleic acid markers associated with EVs [42]. In line with MISEV2023 and related international guidelines, EV studies are encouraged to employ multiple complementary characterization strategies and to implement rigorous quality control (QC) [18]. These practices enhance comparability across studies and improve the reproducibility of conclusions, which is particularly critical for EV omics investigations aimed at clinical translation.

Fig. 1.

Biogenesis, molecular composition, and biomedical applications of extracellular vesicles (EVs). (a) Biogenesis of exosomes through the endosomal pathway involving multivesicular bodies (MVB), and shedding of microvesicles from the plasma membrane. (b) Typical molecular composition of EVs, including membrane-associated components such as tetraspanins (CD63, CD9, and CD81), integrins, lipid rafts, and major histocompatibility complex (MHC) molecules, together with cargo molecules such as nucleic acids, proteins, and metabolites. (c) Major biomedical applications of EVs in diagnosis, therapy, and drug delivery

Overview of EV omics research

Current studies predominantly focus on EV proteomics, transcriptomics, metabolomics, and lipidomics, with some work extending into EV-related genomics and epigenomics [43]. Proteomics primarily relies on mass spectrometry, whereas transcriptomics is mainly based on high-throughput sequencing platforms. Meanwhile, metabolomics and lipidomics are often performed using targeted and untargeted mass spectrometry or chromatography-coupled techniques. Public databases such as Vesiclepedia [44], ExoCarta [45], EVmiRNA [46], exoRBase [47], and EVpedia continuously curate and update EV component annotations [48]. These resources are essential for experimental design and data comparison and foster standardization and data sharing within the field. Multi-omics integration has highlighted the pronounced molecular heterogeneity of EVs and revealed their critical roles in tumor microenvironment remodeling, signal transduction and immune regulation [49]. For instance, in malignancies such as breast and colorectal cancers, EV-mediated transfer of RNAs and proteins has been implicated in the regulation of cancer cell proliferation, invasion, metastasis, and immune evasion [50]. Some EV-related molecules have demonstrated potential as clinical diagnostic and prognostic biomarkers in various studies [51]. Compared with other omics, investigations of EV genomics are still relatively limited. This is largely due to technical challenges such as low DNA content within vesicles, susceptibility to cell-free DNA contamination, limited detection sensitivity and high background noise. Methodological innovation and standardization remain urgent needs for future progress in this area [52].

Clinical application potential of EVs in oncology

EVs are widely distributed in various body fluids—including blood, urine, and saliva—and their cargo remains relatively stable within the circulatory system [53]. As such, EVs are regarded as attractive non-invasive biomarkers and molecular carriers for liquid biopsy in oncology (Fig. 1c) [54]. Detection of EV-associated miRNAs, proteins, and metabolites has shown potential in the early diagnosis, subtyping, and recurrence monitoring of tumors [55]. For example, several studies have reported correlations between the expression levels of immune checkpoint molecules (such as PD-L1) in circulating EVs and patient responses to immunotherapy [56]. These findings suggest that EV-associated immune checkpoint markers may be useful for therapeutic stratification and efficacy assessment. Moreover, owing to their biocompatibility and tissue-targeting properties, EVs have been actively explored as natural nanoscale vehicles [57]. They are being investigated for the delivery of nucleic acid drugs, protein therapeutics and immunomodulatory agents in both preclinical and early-phase clinical studies [58]. Meanwhile, the roles of EVs in the modulation of tumor immune microenvironments and the promotion of distant metastasis are being increasingly elucidated, providing new theoretical insights for the development of novel antitumor targets and combination therapy strategies [59]. Nevertheless, significant challenges remain regarding large-scale EV production and purification, biosafety evaluation, standardization of characterization and quality control practices, and regulatory classification [60]. With ongoing advancements in single-EV analysis, high-precision omics platforms, and AI-assisted data interpretation, EV-based diagnostic and therapeutic strategies may further advance precision oncology.

Fundamentals of machine learning in cancer omics

The development and progression of tumors involve highly complex biological processes that span genomics, epigenomics, transcriptomics, proteomics, and metabolomics [61]. As high-throughput technologies have advanced, tumor research has generated increasingly large and intricate multi-omics datasets, often paired with rich clinical phenotypic information [62]. These datasets are typically high-dimensional, while sample sizes are often limited. They also show substantial heterogeneity and strong batch effects. Under such conditions, conventional statistical methods often struggle to capture key patterns and handle complex data structures [63]. ML offers powerful tools for feature representation, pattern discovery and predictive modeling [64]. It therefore provides a flexible framework for integrating diverse tumor multi-omics and clinical data to support precision oncology and individualized clinical decision-making. This section reviews the major categories of ML and representative algorithms, outlines a standard modeling workflow for tumor omics data, and summarizes key applications of ML in tumor omics and clinical research.

Machine learning paradigms and representative algorithms

Machine learning (ML), a major subfield of AI, is a key component of modern biomedical data analysis and is particularly important in tumor omics and clinical research (Fig. 2b). ML mainly includes supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning [65]. Supervised learning uses labeled data to learn mappings from features to known outcomes and is widely applied to tumor classification, molecular subtyping, survival prediction, and risk assessment [66]. Unsupervised learning identifies patterns in unlabeled data and is frequently employed for patient clustering, molecular subtyping, sample grouping, and dimensionality reduction [67]. Semi-supervised and self-supervised learning enable models to leverage limited labeled data together with large amounts of unlabeled data [68]. This paradigm is particularly useful in multi-omics and medical imaging studies where annotated samples are scarce. Reinforcement learning focuses on sequential decision-making and policy optimization [69]. It has begun to be explored for planning personalized treatment pathways and dynamically adjusting therapeutic strategies.

Fig. 2.

Overview of artificial intelligence (AI), machine learning (ML), and deep learning (DL), with representative algorithms and biomedical applications. (a) Schematic illustrations of commonly used ML algorithms, including linear regression, logistic regression, decision tree, RF, kNN, SVM, k-means clustering, dimensionality reduction methods (e.g., principal component analysis, PCA), and Naive Bayes classifiers. (b) Conceptual hierarchy showing AI as the broadest field, with ML as a subset of AI and DL as a subset of ML. (c) Major biomedical application areas of AI, ML, and DL, including biomarker discovery, diagnostics, disease monitoring, personalized therapy, and drug development

In practice, algorithm selection is guided by data characteristics, sample size, and analytical objectives. For high-dimensional tumor omics data, classical ML algorithms remain widely used. Linear and generalized linear models (such as logistic regression) are popular owing to their simplicity and interpretability and are often used to construct risk scores and prediction models in clinical settings [70]. Regularization methods such as LASSO (least absolute shrinkage and selection operator) and Elastic Net are valuable when there are many variables and relatively few samples, as they help select important features and improve model stability [71]. Tree-based models and ensemble methods, including decision trees, random forests (RF), and eXtreme Gradient Boosting (XGBoost), are well suited to handling nonlinear relationships and complex interactions [72]. Methods such as support vector machines (SVM) and k-nearest neighbors (kNN) also perform well on datasets with moderate sample sizes (Fig. 2a) [73]. For images, tissue slides, or molecular interaction networks, deep learning (DL) techniques—including convolutional neural networks (CNNs), recurrent neural networks (RNNs), transformers, and graph neural networks (GNNs)—are particularly effective at extracting hierarchical and complex features [74]. In addition, multimodal neural network architectures show promise for integrating heterogeneous data types and capturing tumor heterogeneity and longitudinal disease evolution [75].

Overview of the machine learning modeling workflow

A typical ML workflow for tumor omics analysis comprises several key steps: cohort selection and dataset partitioning, data preprocessing and quality control, feature selection and dimensionality reduction, model training and validation, and model interpretation [76]. Cohort design should be aligned with the overarching clinical study protocol. A typical setup separates a discovery cohort, a development set for training and internal validation, and an independent test set reserved for final performance assessment [77]. Data preprocessing and quality assurance are critical for obtaining reliable results. Essential steps include data cleaning, handling of outliers and missing values, harmonization of coding schemes, and feature normalization or standardization (e.g., z-score or min–max scaling) to reduce variability arising from technical factors [78]. Strong batch effects and platform differences are common in omics data and should be corrected using methods such as ComBat or Harmony [63d, 79]. Importantly, batch correction and feature selection should be implemented within the cross-validation framework to avoid information leakage.

Class imbalance frequently occurs in tumor datasets, for example when positive cases are much fewer than controls [80]. It can be mitigated using resampling strategies, class-weighted loss functions, or careful adjustment of decision thresholds to improve performance on minority classes. Feature selection and dimensionality reduction are crucial for modeling high-dimensional data [81]. Filter-based methods quickly screen and remove weakly associated variables based on univariate statistics or effect sizes [82]. Embedded methods, such as LASSO, Elastic Net, and tree-based feature importance measures, perform feature selection during model fitting, thereby reducing overfitting risk and enhancing interpretability. [82–83] Principal component analysis (PCA) is widely used for exploring global data structure, assessing batch effects, and generating reduced-dimension inputs for modeling [84]. Nonlinear techniques such as t-distributed stochastic neighbor embedding (t-SNE) and Uniform Manifold Approximation and Projection (UMAP) are mainly used for visualization and exploratory identification of sample subgroups or latent structures [85].

Model training and validation typically rely on cross-validation, including k-fold and nested cross-validation, to obtain robust performance estimates and guide hyperparameter tuning [86]. For diagnostic and classification tasks, commonly reported metrics include the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and the F1 score [87]. For prognostic modeling, the concordance index (C-index) and time-dependent ROC curves are more informative [88]. Model interpretation methods facilitate understanding of the model’s decision logic. These include examination of feature coefficients or importance scores, SHAP (Shapley additive explanations) values, and partial dependence plots [89]. They also help prioritize candidate biomarkers for further biological and clinical investigation.

Applications of machine learning in tumor clinical research

The integration of ML with tumor multi-omics and clinical data has enabled a wide range of clinical applications in oncology (Fig. 2c) [90]. In early detection, ML models that incorporate diverse biomarkers together with clinical variables can improve detection rates and risk stratification in high-risk populations compared with traditional single-marker approaches [91]. In diagnosis and subtyping, ML facilitates the fusion of heterogeneous data sources, including tissue omics, radiological imaging, and digital pathology [92]. Such integrated models support automated or semi-automated identification of tumor types, molecular subtypes, and immune microenvironment states, thereby assisting clinical decision-making. For therapy response prediction and longitudinal monitoring, ML models can integrate baseline characteristics with multi–time-point follow-up data [93]. This approach enables more consistent assessment of treatment efficacy and resistance development and can support timely treatment adjustment and monitoring. In prognosis and risk assessment, ML-based models that jointly leverage clinical, pathological, imaging, and multi-omics features can improve the precision of survival and recurrence predictions and support more individualized management plans [94]. Compared with traditional prognostic scores based on a limited number of variables, these models are better equipped to capture the underlying biological and clinical heterogeneity among patients.

Applications of machine learning in EVomics

EVs are emerging as highly promising biomarkers with substantial translational potential for cancer diagnosis and therapy [95]. The widespread implementation of high-throughput transcriptomic, proteomic, and metabolomic/lipidomic platforms has facilitated increasingly detailed characterization of EV molecular profiles [43c]. These advances open new avenues for individualized disease stratification and targeted therapeutic intervention. Nevertheless, extracting stable and reproducible biological signals from these high-dimensional datasets remains a formidable challenge. The inherent sparsity and complex intercorrelations within omics data frequently impede robust biomarker discovery and predictive model construction [96]. In response to these analytical complexities, ML provides a versatile toolkit for deciphering large-scale biomedical data and advancing precision medicine [97]. However, the choice and evaluation of ML models in EV omics are strongly shaped by the statistical properties of EV-omics datasets.

Viewed collectively, current EV-omics studies reveal several recurrent patterns in model choice and reported performance. Typical EV-omics datasets are often characterized by a “small-n/high-p” structure (i.e., a limited sample size n but a large number of measured features p), together with sparsity, missingness, and substantial feature collinearity [98]. Accordingly, current EV-omics studies often favor simpler and more robust modeling strategies, which helps explain why classical machine-learning algorithms remain dominant in this field. Regularized linear models (e.g., logistic regression with LASSO or elastic-net regularization) are frequently used to derive compact, interpretable, and potentially assay-transferable biomarker panels in p≫n settings [99]. Tree-based ensemble methods, including Random Forests and gradient boosting approaches such as XGBoost, are also widely adopted because they can capture nonlinear relationships and higher-order feature interactions while remaining effective in relatively small cohorts [100]. SVMs likewise remain competitive in high-dimensional settings, although their performance is often more sensitive to preprocessing choices and hyperparameter tuning [101]. Deep learning models, particularly CNN-based architectures, appear most advantageous for structured EV inputs, such as images, spectra, or single-EV signal maps, whereas in moderate-sized tabular omics datasets they do not consistently outperform well-tuned classical models [102]. When tree-based models are used for biomarker prioritization, permutation-based importance measures can further help reduce known biases in feature ranking [103]. Notably, some near-perfect AUCs summarized in Section 4 were reported in retrospective case–control cohorts with enriched, clear-cut cases and controls, which can overestimate diagnostic discrimination due to spectrum/selection effects [104]. Model performance may therefore attenuate in prospective, consecutively enrolled real-world screening or surveillance populations, underscoring the need for independent external validation, preferably with prospective designs, and transparent reporting [105].

This section summarizes representative application scenarios and seminal studies across distinct omics domains, delineating principal algorithms and their predictive performance. We also critically assess recurring methodological limitations to provide a balanced perspective and to support discussion of EV multi-omics integration and future clinical translation. To visualize this landscape, Fig. 3a presents an end-to-end workflow for machine learning-enhanced EV omics in oncology, spanning biofluid collection and EV isolation, multi-omics profiling, data preprocessing and feature engineering, model development and validation, and downstream clinical applications. Figure 3b then contrasts the conceptual and architectural differences among three canonical fusion paradigms—early (feature-level), late (decision-level), and hybrid (representation-level) fusion—together with EV-specific considerations (e.g., batch effects, missing-modality scenarios, and small-n/high-p overfitting risks). [90a, 106].

Fig. 3.

Workflow of ML-enhanced EV multi-omics in oncology and major EV multi-omics fusion strategies. (a) End-to-end workflow of ML-enhanced EV multi-omics, spanning biofluid collection, EV isolation and characterization, multi-omics profiling, data processing and model development, and downstream clinical and biological applications. (b) Major EV multi-omics fusion strategies. Early integration (feature-level concatenation) combines multi-omics features into a joint feature matrix for a single model. Late integration (decision-level aggregation) combines modality-specific outputs through weighted aggregation. Hybrid integration (representation-level fusion) learns modality-specific representations and merges them into a shared latent representation before final prediction. Key EV-specific considerations include small-n/high-p settings, missing modalities, batch effects, and leakage control

Machine learning in EV transcriptomics

EVs encapsulate nucleic acids—including DNA, mRNA, miRNAs, lncRNAs, and circRNAs—that reflect molecular signatures of their cells of origin and have potential for non-invasive cancer diagnosis and prognosis [107]. Most EV transcriptomics studies follow a common ML pipeline: broad RNA profiling, feature selection for dimensionality reduction, and panel reduction to compact signatures that can be evaluated across cohorts and assay platforms. In contrast, studies combining EV DNA profiling with ML remain rare. To date, only one urine EV-DNA methylation model for prostate cancer has reported a logistic-regression classifier with an AUC of 0.93 [108]. This section summarizes how EV-derived transcriptomic features have been integrated with ML for biomarker discovery and clinical prediction in oncology.

Among these applications, pancreatic ductal adenocarcinoma (PDAC) is one of the most studied settings and often serves as a benchmark for cross-cohort evaluation and clinically relevant comparators. A common strategy starts with broad miRNA or long-RNA profiling and then applies regularized or ensemble models (e.g., logistic regression, LASSO, RF) to derive small signatures that generalize to external validation. For example, one study constructed a 13-miRNA plasma panel from cell-free and exosomal miRNAs using logistic regression, reporting high diagnostic accuracy in both training and external validation cohorts for non-invasive PDAC screening (Fig. 4a) [109]. An eight-gene plasma EV long RNA (exLR) panel selected using RF and LASSO showed consistent performance across training, internal, and external cohorts, supporting EV long RNA for PDAC detection [110]. Beyond panel building, platform innovations have also been paired with ML. For example, a magnetic nanopore EV-miRNA workflow enabled three-class stage classification in PDAC mouse models (healthy vs. localized vs. metastatic), but remains preclinical without human validation [111]. More recently, a multicenter study used EV-RNA sequencing to derive an EV-lRNA index and implemented a two-stage classifier (SVM followed by RF) [112]. The model first distinguished “PDAC + chronic pancreatitis (CP) vs. healthy” and then “PDAC vs. CP,” achieving an AUC of ~ 0.97 in independent validation cohorts. This two-step design mirrors clinical practice by separating “disease vs. healthy” first and then “PDAC vs. CP,” rather than relying on simple PDAC–healthy case–control comparisons.

Fig. 4.

ML–driven EV transcriptomics across biofluids and cancer types. (a) Integration of plasma cell-free and exosomal miRNAs into an ML-derived multi-miRNA signature for non-invasive early detection of PDAC. Reproduced with permission [109]. Copyright 2022, Elsevier. (b) Identification of fecal EV miRNA signatures for CRC using ML-based feature selection and classification, enabling non-invasive stool-based screening. Reproduced with permission [115]. Copyright 2025, American Chemical Society. (c) Application of multiple ML algorithms to blood-derived exosomal RNA profiles for the development of non-invasive multi-cancer diagnostic models and tumor-of-origin classifiers. Reproduced with permission [124]. Copyright 2025, Springer Nature

With similar workflows reported across gastrointestinal cancers, EV transcriptomics–ML studies still vary widely in biofluid choice (serum, plasma, saliva, feces) and measurement platforms, while converging on a common modeling goal: compress high-dimensional RNA profiles into a small, assay-compatible panel with independent validation. For instance, a multicenter cohort identified a four-lncRNA serum EV panel by systematically comparing LASSO–logistic regression, RF, SVM, and XGBoost with cross-validation, and achieved high performance in external validation [113]. In parallel, advances in EV isolation technologies have also been paired with ML. A zeolite-amine and homo-bifunctional hydrazide-based EV isolation platform was coupled with an AI-driven deep neural network to select an optimal EV-derived biomarker panel for colorectal cancer (CRC), outperforming conventional single-marker strategies [114]. In a complementary non-invasive approach, fecal EV-miRNAs were combined with a CRISPR/Cas13a detection system and ML to support stool-based CRC screening and prognostic risk modeling (Fig. 4b) [115]. In this study, ML-based feature selection reduced an initial set of 11 miRNAs to four, and linear discriminant analysis (LDA) achieved high diagnostic accuracy while the CRISPR/Cas13a readout provided point-of-care compatibility. Furthermore, public EV RNA-seq datasets have also been mined with traditional ML and deep learning for “healthy vs. cancer” classification and stage prediction in CRC, with the best models reaching high accuracy [116]. When the clinical task shifts from diagnosis to therapy monitoring, models often need to detect subtler and time-dependent signals, and cohort sizes are frequently smaller. As a result, most studies favor conservative feature selection (e.g., LASSO, RF) to derive compact signatures that can be tested in independent datasets. In advanced gastric cancer, exosome-derived mRNAs, miRNAs, and lncRNAs were profiled and distilled into a compact RNA signature using LASSO-based modeling to identify responders to neoadjuvant chemotherapy [117]. In metastatic CRC, plasma EV-RNA signatures selected by RF and LASSO were trained to predict response to first-line chemotherapy, and then tested in independent validation cohorts [118]. For esophageal cancer, LASSO–logistic regression applied to multicenter salivary EV-miRNA profiles enabled accurate subtype diagnosis and early-stage detection. Cox-based risk scores derived from these signatures stratified prognosis in a large population, providing large-scale validation of salivary EV omics for screening and risk stratification [119]. Collectively, these gastrointestinal studies suggest that, once pre-analytical factors and EV isolation variability are controlled, careful feature selection and panel design can be as important as classifier complexity for achieving stable performance.

Beyond gastrointestinal cancers, EV transcriptomics has also been used for immunotherapy response prediction, organ-specific detection, and long-term prognosis. In multiple myeloma, IsoSeek-based plasma EV miRNA/isomiR sequencing enabled cross-validated LASSO classifiers for active-disease detection and on-treatment response assessment with prospective validation, showing consistent performance [120]. A related strategy was reported in non-small cell lung cancer (NSCLC), where a two-circRNA EV signature was derived using LASSO-based pairwise circRNA screening. This signature predicted response to immune checkpoint blockade with robust performance in both discovery and external cohorts, illustrating the potential of ML-driven EV-circRNA panels for precision immuno-oncology [121]. For early-stage lung adenocarcinoma (LUAD), multimodal fusion of serum EV long RNAs with computed tomography (CT)-derived features using XGBoost improved differential diagnosis of small nodules, with AUC around 0.95 for benign–malignant classification [122]. Saliva-based EV transcript panels have also been evaluated in oral squamous cell carcinoma, where small mRNA panels showed clear discrimination and links to tumor grade or nodal status [123]. Multi-class models extend EV transcriptomes to multi-cancer detection and tumor-of-origin prediction, while survival models reuse these signatures for prognosis. A plasma exosomal transcriptome-based model spanning eight tumor types compared multiple algorithms and found RF to perform best for simultaneous detection and tumor-of-origin prediction (Fig. 4c) [124]. In ovarian cancer, AdaBoost-based models built from plasma exosomal transcriptomes achieved strong diagnostic performance in independent test cohorts [125]. Complementary mining of public datasets has combined Cox regression with feature-selection methods such as LASSO and RF to derive exosomal lncRNA-based prognostic scores with good long-term discrimination [126]. For hepatocellular carcinoma (HCC), mining of exoRBase 2.0 combined with SVM-based feature selection yielded a diagnostic panel comprising seven mRNAs and two circRNAs. This panel showed encouraging accuracy in blind testing but still lacks prospective clinical validation [127].

Recent studies increasingly emphasize deployability—testing signatures in high-risk surveillance settings and implementing readouts on chip or PCR-based platforms—while also addressing signal dilution from bulk EV heterogeneity. In HCC, clinically oriented designs have included high-risk cohorts and chip-enabled EV capture paired with compact EV-mRNA scoring models [128]. Multi-source integration of extracellular RNAs has also produced small noncoding-RNA panels that retain sensitivity in AFP-negative and early-stage disease [129]. In breast cancer, a three-mRNA model was measured by a 4-plex droplet digital PCR platform and analyzed using RF, SVM, and neural networks. This model achieved promising diagnostic performance, illustrating the feasibility of implementing compact EV-mRNA signatures on clinically compatible platforms [130]. In endocrine oncology, circulating sEV miRNA panels modeled by logistic regression have been reported to distinguish follicular thyroid carcinoma from follicular thyroid adenoma in multicenter validation, with AUC around 0.84 [131]. To reduce signal dilution further, single-EV-level miRNA detection has been combined with deep learning (notably convolutional neural networks), enabling multicancer EV-miRNA subtyping at single-particle resolution with high performance in proof-of-concept datasets [132].

Taken together, ML-enabled EV transcriptomics in oncology has expanded from miRNA-only signatures to multi-analyte panels incorporating mRNA, lncRNA, and circRNA. It has also moved beyond plasma/serum assays to additional biofluids such as feces and saliva. Methodologically, many studies converge on the same approach: strong dimensionality reduction, compact panel design, and internal/external validation. Although high performance is often reported in case–control settings, it may not translate to real-world performance. Many datasets are retrospective and enriched for clear-cut cases, with limited evidence from high-risk populations or prospective cohorts, raising concerns about spectrum bias and performance shifts in deployment [133]. In addition, sample sizes are often modest relative to model complexity, and reporting of calibration and clinical utility remains inconsistent. Larger, prospectively designed multicenter studies with standardized pre-analytics and transparent reporting are essential to establish the clinical value of EV–ML transcriptomic signatures.

Machine learning in EV proteomics

EVs carry a diverse proteome on their surface and within their lumen, including adhesion molecules, receptors, enzymes, and signaling mediators that shape intercellular communication and microenvironmental remodeling [134]. Proteomic profiling of EVs has therefore become a key strategy for biomarker discovery and mechanistic studies across a broad spectrum of diseases [135]. Mass spectrometry (MS)-based proteomics and EV-targeted protein microarrays enable comprehensive characterization of EV protein cargo and identification of disease-associated signatures [136]. Coupled with ML, these readouts can be distilled into compact EV protein panels and predictive models that support cancer detection, subtyping, and risk stratification.

At the pan-cancer level, ML has been used in two complementary ways: (i) atlas-scale feature mining to build multicancer classifiers and (ii) physical stratification to model EV heterogeneity. A pan-cancer analysis profiled EVPs from 426 tissue and biofluid samples using high-throughput MS and applied RF to identify EV protein markers across tumor types, achieving ~ 95% sensitivity and 90% specificity for cancer detection [137]. Subsequent work showed that size heterogeneity can mask membrane protein signals in bulk measurements; a single-EV enumeration platform linking membrane proteins to particle size, together with LDA and size-based stratification, improved diagnostic accuracy across tumor types [138]. In parallel, a meta-analysis aggregating 1,083 exosomal proteomic datasets used mutual information and RF to build plasma-, serum-, and urine-based diagnostic models, with RF outperforming SVM, kNN, and Gaussian naive Bayes classifiers [139]. Together, these studies show how ML can extract compact panels from atlas-scale EV proteomics and extend them to size-defined EV subsets for liquid biopsy translation.

Building on these atlas-scale efforts, EV proteomics–ML studies in gastrointestinal malignancies often pair discovery-scale MS with targeted verification, such as parallel reaction monitoring (PRM) and ELISA. These designs also incorporate clinically meaningful comparators and covariates. In primary sclerosing cholangitis–associated cholangiocarcinoma, EV proteomic profiling combined with ML yielded a three-protein panel that distinguished primary sclerosing cholangitis–associated cholangiocarcinoma (PSC-CCA) from PSC, with performance further improved when CA19-9 was added in multicenter cohorts with external validation [140]. For early CRC diagnosis, four-dimensional data-independent acquisition (4D-DIA) proteomics and ML algorithms such as RF and LASSO have been used to define EV protein panels with robust performance in both training and test sets [141]. Chip-based multiplex assays also enable high-throughput clinical testing; for example, work focusing on exosomal membrane proteins integrated a high-throughput CRC-EVArray chip with multicenter cohorts and evaluated multiple ML algorithms to obtain a three-protein signature for early detection and progression prediction [142]. Extending this workflow, a multicenter CRC study integrated tumor tissue and plasma EV proteomes profiled by 4D-DIA and used RF-related ML methods for feature selection and modeling. Candidates were then validated by PRM and ELISA to derive a three-protein signature and build the ColonTrack model. This model showed strong discrimination in training and test sets and maintained consistent accuracy in both internal and external validation cohorts (Fig. 5a) [143]. In upper gastrointestinal cancer, a microfluidic barcode chip for profiling membrane and intravesicular proteins in serum small EVs was combined with an RF classifier to build a nine-protein model for early esophageal squamous cell carcinoma (ESCC). Starting from 14 candidates identified by 4D-DIA, the final model achieved high diagnostic accuracy and outperformed the SCC serum marker in multicenter cohorts (Fig. 5b) [144].

Fig. 5.

Representative ML-enabled EV proteomics strategies in oncology. (a) ColonTrack integrates EV proteomics, ML-based feature selection, and targeted validation to derive a protein panel for CRC detection. Reproduced with permission [143]. Copyright 2025, Elsevier. (b) A microfluidic barcode biochip combined with RF modeling enables serum small-EV protein profiling for early ESCC diagnosis. Reproduced with permission [144]. Copyright 2025, Wiley-VCH GmbH. (c) Serum EV proteomics coupled with ML identifies TALDO1 as a marker associated with breast cancer metastasis. Reproduced with permission [147]. Copyright 2024, American Association for Cancer Research. (d) A five-EV membrane protein panel identified by single-particle nanochip profiling and ML supports early-stage lung cancer diagnosis. Reproduced with permission [148]. Copyright 2025, Wiley-VCH GmbH. (e) High-throughput EV proteomics with ML-based analysis identifies GDF15 as a plasma EV biomarker for prostate cancer diagnosis and grading. Reproduced with permission [151]. Copyright 2025, American Chemical Society

Beyond gastrointestinal tumors, EV proteomics–ML approaches have supported early detection and classification in breast and lung cancer. In breast cancer, reverse-phase protein arrays (RPPA) profiling of plasma small EV proteins combined with kNN, elastic net, and RF produced a seven-protein signature that distinguished patients from healthy controls with higher accuracy than conventional serum markers. The same signature also supported staging, molecular subtype classification, and recurrence risk modeling [145]. Capture platforms, such as microfluidic tumor-EV isolation chips, have been combined with hybrid ML frameworks (e.g., boosting, CNNs, and SVM) to select four-protein panels that performed well in both overall breast cancer and TNBC classification [146]. Another breast cancer study used 4D-DIA MS-based proteomics with XGBoost and related algorithms to build a seven-protein diagnostic model that was validated across five independent cohorts and identified TALDO1 in serum EVs as a marker of distant metastasis (Fig. 5c) [147]. In lung cancer, a high-throughput single-particle nanochip platform and ML were used to screen 52 EV membrane proteins and derive a five-protein panel for early-stage disease. This panel outperformed CEA and showed stable performance across pathological subtypes, smoking status, sex, and age (Fig. 5d) [148]. Similarly, in NSCLC, high-throughput proteomics applied to multicenter plasma samples and a multi-step pipeline combining LASSO, RF, and Boruta produced a seven-protein panel from CD155-immunocaptured EVs with excellent diagnostic performance in validation cohorts [149].

In urologic oncology, urinary and plasma EV proteomes provide information that complements prostate-specific antigen (PSA), and ML is primarily used to select minimal marker sets that improve grading and differential diagnosis. A chemical glycoproteomics strategy for high-throughput profiling of urinary EV surface proteins identified a four-marker panel with strong diagnostic and grading value, and SVM-based models built on this panel yielded high diagnostic accuracy [150]. In a second study, a TiO₂ affinity capture method coupled with 4D-DIA MS and ML (PLS-DA and RF) pinpointed GDF15 as an optimal marker for prostate cancer diagnosis and grading, with additional support from ELISA and immunohistochemistry (Fig. 5e) [151]. Building on urinary EVs, a high-throughput proteomic platform for large EVs combined with XGBoost and RF, together with PSA, produced a four-protein panel that markedly improved differential diagnosis between prostate cancer and prostatitis, especially in the PSA gray zone [152].

Overall, EV proteomics–ML studies have progressed from feasibility demonstrations to proof-of-concept multicancer diagnostic models. Across gastrointestinal, breast, lung, and urological cancers, diverse EV proteomic platforms combined with ML have generated compact protein panels that often outperform conventional serum biomarkers and support early detection, tumor subtyping, and risk stratification. At the same time, most studies still rely on retrospective designs, often from one or a few centers, and show substantial heterogeneity in EV isolation, proteomic workflows, and analysis pipelines. These factors limit comparability and clinical translation and highlight the need for standardized, prospectively evaluated EV–ML studies in larger and more diverse cohorts.

Machine learning in EV metabolomics and lipidomics

Altered metabolic states and metabolic reprogramming are hallmarks of tumor progression [153]. EV metabolomics and lipidomics extend these concepts to a vesicle-protected pool of small molecules that can be profiled from biofluids, offering a complementary biomarker space to EV RNAs and proteins [154]. Compared with transcriptomics or proteomics, metabolite and lipid signals can be more sensitive to pre-analytical variation and batch effects, so ML is frequently used for multivariate denoising, feature selection, and integration with clinical variables or other omics layers [155]. The resulting models typically reduce high-dimensional spectral features to a small set of metabolites or lipids that can distinguish cancer subtypes, disease states, and treatment outcomes.

In gastrointestinal malignancies, EV metabolomics–ML studies have explored multiple biofluids and analytical platforms. A common approach is to combine untargeted profiling with supervised learning to identify a small diagnostic metabolite set. For example, in early gastric cancer, urinary exosomes were captured using a CD63-aptamer–modified polymorphic carbon material and profiled by laser desorption/ionization mass spectrometry (LDI-MS). Orthogonal partial least squares discriminant analysis (OPLS-DA) models separated early gastric cancer from controls in both discovery and blinded cohorts, and three metabolites served as biomarkers that tracked disease progression [156]. In CRC, fecal microbiota–derived EVs from patients and healthy volunteers were characterized by untargeted metabolomics. SVM, PLS-DA, and RF models were used to screen diagnostic markers. Five dysregulated metabolites were selected, and an SVM model based on these features achieved high diagnostic performance [157]. In another CRC cohort, 16S rDNA sequencing was combined with metabolomics to assess changes in the microbiome and small-molecule metabolites. Binary logistic regression models based on metabolomics, microbiome, and combined features all showed strong diagnostic performance, with the integrated model performing best [158]. Beyond case–control diagnosis, EV metabolomics has also been applied to outcome prediction when longitudinal endpoints are available. For ESCC recurrence prediction, the EXODUS platform was used to isolate high-purity plasma exosomes and to profile 196 metabolites, including fatty acids, amino acids, and sugars, by LC-MS. RF feature selection identified four key metabolites that formed a recurrence prediction panel with strong performance (Fig. 6a) [159]. In HCC, magnetic core–shell nanoparticles enabled efficient serum exosome isolation and metabolomic fingerprinting. Unsupervised clustering highlighted between-group differences, and supervised models (OPLS-DA, RF, SVM, kNN, and neural networks) were then applied. Six key exosomal metabolites were selected, and the final model showed high accuracy for HCC detection (Fig. 6b) [160].

Fig. 6.

ML applications in EV metabolomics and lipidomics. (a) Exosome metabolic profiling using a CD63-aptamer–coupled polymorphic carbon substrate and ML-based modeling enables detection of early gastric cancer. Reproduced with permission [156]. Copyright 2022, American Chemical Society. (b) Engineered magnetic core–shell nanoparticles combined with LC–MS metabolomics and ML classifiers support non-invasive diagnosis of HCC. Reproduced with permission [160]. Copyright 2024, American Chemical Society. (c) A versatile exosomal metabolite assay platform based on a gold nanoparticle–coated magnetic covalent organic framework with aptamer modification, together with ML-based feature selection and classification, improves diagnostic accuracy and biomarker discovery in endometrial cancer. Reproduced with permission [163]. Copyright 2024, American Chemical Society

EV metabolomics and lipidomics can also be strengthened through multimodal modeling that combines vesicle-associated features with matched non-EV fractions or other omics layers. In prostate cancer, metabolomic profiling of serum small EVs and EV-depleted serum was integrated with conventional PSA. Features were selected by OPLS-DA and integrated into RF and related models. The resulting multimodal EV-based model, built from 11 EV-derived and 2 EV-free metabolites, outperformed PSA or serum alone [161]. In melanoma, proteins and metabolites in plasma-derived EVs were measured in parallel. kNN, LASSO-logistic regression, and RF models based on the combined protein–metabolite signature achieved good classification accuracy and highlighted markers associated with clinical stage. These results support the value of EV proteomics–metabolomics integration for risk stratification and patient-level modeling [162]. Similar workflows have been applied to other solid tumors, with emphasis on translating untargeted fingerprints into a minimal, interpretable marker set. In endometrial cancer, a composite-material platform enabled precise exosome isolation and metabolic fingerprinting. SVM and related algorithms selected a small set of discriminative features that achieved high AUC and accuracy in independent testing, and SVM-based feature weighting highlighted four candidate metabolic biomarkers (Fig. 6c) [163]. Finally, at a pan-cancer level, EV lipidomics has enabled the discovery of tumor-specific lipid markers. Using PCA and OPLS-DA, phosphatidylserine (PS) was identified as a tumor-specific EV lipid. This finding led to the development of the PSEV-MultiCancer flow-cytometry blood test for minimally invasive screening in surgically resectable cancers. The test achieved 74.7% sensitivity for stage I/II disease and an AUC of 0.932, demonstrating the translational promise of EV lipidomics-based biomarkers [164].

Viewed together, EV-encapsulated metabolites and lipids carry informative signatures across many types of cancer and in pan-cancer settings. Compact biomarker panels derived from these datasets often show strong discrimination for early detection, recurrence prediction, and risk stratification. Multi-omics strategies that link EV metabolites or lipids with proteomics or microbiome data can further improve diagnostic performance and provide mechanistic insight. However, most current studies are based on small, retrospective, single-center cohorts with heterogeneous EV processing protocols and limited external validation, which raise concerns about overfitting and generalizability. To establish clinical utility, future work will need standardized pre-analytical protocols, harmonized analytical pipelines, and large, prospectively designed multicenter studies with rigorous external validation to support regulatory approval.

Methodological trends and translationalgaps

Taken together, Sections 4.1–4.3 show that ML is increasingly integrated into EV-based transcriptomic, proteomic, metabolomic, and lipidomic profiling for oncology applications. Across studies, EV-omics features have been used for cancer detection (including early detection and multi-cancer settings), cancer-type classification (and tumor-of-origin inference in multi-class designs), prognosis and recurrence-risk prediction, and treatment-response monitoring. EV-omics features are typically distilled into compact, assay-transferable panels through rigorous feature selection and validation. Building on the algorithm considerations summarized earlier in Section 4 , this section focuses on methodological trends and translational gaps that most strongly affect robustness, comparability, and real-world deployment. Because cohort spectrum, sampling strategy, and validation design vary widely across studies, headline discrimination metrics (e.g., AUC) should be interpreted as proof-of-concept and are not directly comparable without considering these design factors [105, 165].

Despite this progress, current EV–ML studies share several methodological limitations that restrict their readiness for clinical translation. Most studies employ small, single-center case–control cohorts with limited phenotypic diversity and inadequate representation of high-risk or screening populations. These designs increase the risk of overfitting and spectrum bias and tend to inflate performance estimates, particularly in the absence of independent external or prospective validation [166]. Accordingly, many near-perfect AUCs should be interpreted as early proof-of-concept rather than deployable performance, and claims of clinical readiness should be supported by independent external validation in representative screening/surveillance populations. [104–105] Analytical rigor in model development and evaluation is also frequently suboptimal. Many studies use simple train–test splits without nested cross-validation and do not clearly separate feature selection from model training, allowing information leakage and overly optimistic results. This is particularly problematic in high-dimensional omics, where feature selection and hyperparameter tuning must be performed within an inner resampling loop (i.e., nested cross-validation) to avoid biased error estimates [167]. Systematic assessment of calibration, decision-analytic utility, and performance across clinically relevant subgroups is uncommon, and formal analyses of model interpretability remain the exception rather than the rule. Given that calibration is critical for risk-stratification use cases, reporting calibration curves/slope/intercept and, where appropriate, decision-curve analysis can better reflect clinical utility beyond AUC [168]. To improve transparency and reduce avoidable bias, future EV–ML studies should report and validate models in line with TRIPOD+AI and assess risk of bias using PROBAST [105, 165].

A further gap lies in the limited implementation of true multi-omics integration. Most studies focus on single omics layers (transcriptomics, proteomics, metabolomics, or lipidomics) rather than jointly modeling complementary information across modalities. Bollard et al.‘s combined proteomics-metabolomics analysis in melanoma represents a notable exception, demonstrating that multi-omics integration substantially enhances diagnostic accuracy and biological insight. As high-throughput profiling technologies mature, integrating EV transcriptomics, proteomics, metabolomics, and lipidomics within rigorous ML frameworks will be essential to maximize biomarker robustness and illuminate tumor-microenvironment interactions [169]. Realizing this potential will require multicenter clinical cohorts that include diverse and high-risk populations, together with standardized pre-analytical handling, EV isolation, and analytical pipelines. Equally important are rigorous model development and validation procedures, including nested cross-validation, clear separation of training and evaluation sets, independent external validation using an external cohort, and prospective multicenter validation where feasible, with transparent reporting aligned with emerging ML guidelines. Representative studies across omics modalities, cancer types, ML algorithms, and performance metrics are summarized in Table 1, with cohort size and validation type provided to better contextualize reported performance. Notably, most EV-omics ML models remain retrospective and rely mainly on internal validation, underscoring the need for adequately powered, prospective, multicenter external validation before clinical deployment.

Table 1.

Representative machine learning applications for EV transcriptomic, proteomic, and metabolomic/lipidomic biomarker studies

Omics	Tumor types	Algorithms	Model performance	Cohort (n) & validation	Sample type	Ref.
Transcriptomics	Pancreatic Ductal Adenocarcinoma	Logistic regression	AUC: 0.93–0.98	Discovery: (n = 101); Training: (n = 96); Validation: internal (n = 95); Total sample: (n = 292)	Exosomal/Cell-free miRNAs	[109]
	Pancreatic Ductal Adenocarcinoma	LASSO and RF	AUC: up to 0.96	Discovery: NR; Training: (n = 188); Validation: internal (n = 135), external (n = 178); Total sample: (n = 501)	Plasma EV long RNA	[110]
	Gastric Cancer	LASSO/RF/SVM/XGBoost	AUC: 0.942–0.959	Discovery: (n = 105); Training: (n = 978; 70%/30% internal split for training/testing); Validation: external (n = 227), prediction cohort (n = 285); Total sample: (n = 1595)	Serum exosome non-coding RNA	[113]
	Colorectal Cancer	DNN	AUC: 0.9722–1.0	Discovery: NR; Training: (n = 70); Validation: internal (n = 30); Total sample: (n = 100)	Plasma EV miRNA	[114]
	Colorectal Cancer	LDA	AUC: up to 0.9864	Discovery: (n = 46); Training: (n = 57); Validation: external (n = 26); Total sample: (n = 129)	Fecal EV miRNA	[115]
	Gastric Cancer	LASSO Cox regression	AUC: up to 1	Discovery: NR; Training: (n = 63); Validation: external (n = 43); Total sample: (n = 106)	Plasma EV RNAs	[117]
	Colorectal Cancer	LASSO-logistic	AUC: 0.986	Discovery: NR; Training: (n = 80); Validation: internal (n = 62) + external, prospective (n = 48); Total sample: (n = 190)	Plasma EV RNAs	[118]
	Esophageal Cancer	LASSO	AUC: 0.968	Discovery: (n = 126); Training: (n = 342); Validation: internal (n = 207), external (n = 226); Total sample: (n = 901)	Salivary EVP miRNA	[119]
	Lung Adenocarcinoma	XGBoost	AUC: 0.948	Discovery: NR; Training: (n = 146); Validation: internal (5-fold CV, n = 146); Total sample: (n = 146)	Serum EV long RNA (evlRNA) + CT attributes (radiomics)	[122]
	Multi-cancer types	RF	AUC: up to 0.98	Discovery: (n = 818, public/previous datasets); Training: (n = 1109); Validation: internal (n = 276); Total sample: (n = 2203)	Plasma EV RNA	[124]
Proteomics	Cholangiocarcinoma	Binary logistic regression	AUC: 0.947	Validation: internal (n = 57, split from Discovery), external (n = NR); Total sample for this model: (n = 187)	Serum EVs	[140]
	Colorectal Cancer	RF/Rpart/Logistic regression/kNN/SVM	AUC: 0.963	Discovery: (n = 37); Training: (n = 338); Validation: internal (n = 328), external (n = 246); Total sample: (n = 949)	Serum EVs	[141]
	Esophageal Squamous Cell Carcinoma	RF	Accuracy: 90.8%	Discovery: (NR; proteomics screening n = 30 for target selection); Training: (n = 103); Validation: internal (n = 43), external (n = 127); Total sample: (n = 273)	Serum EVs	[144]
	Breast Cancer	Logistic regression/kNN/RF	Accuracy: 88%; Sensitivity: 94%	Discovery: (n = 74); Training: (CV within Discovery n = 74); Validation: internal (CV within Discovery n = 74), external (n = 24); Total sample: (n = 98)	Plasma EVs	[145]
	Triple-Negative Breast Cancer	CNN-SVM hybrid	AUC: 0.973	Discovery: (n = 130); Training: (n = 91, 70% split from Discovery); Validation: internal (n = 39, 30% split from Discovery), external (n = 40); Total sample: (n = 170)	Plasma EVs	[146]
	Lung Cancer	Logistic regression/Deep learning	AUC: 0.926–0.931	Discovery: (n = 947); Training: (n = 473, 1:1 split from Discovery); Validation: internal (n = 474, 1:1 split from Discovery), external (n = NR); Total sample: (n = 947)	Plasma EVs	[148]
	Non-Small Cell Lung Cancer	LASSO/RF/Boruta	AUC: up to 1.0	Discovery: (n = 28); Training: (n = 71); Validation: internal (n = 71, within-Training), external (NR); Total sample: (n = 99)	Plasma EVs	[149]
	Prostate Cancer	SVM	AUC: 0.8636	Discovery: (n = 172); Training: (NR; 80% split within the independent cohort n = 196); Validation: internal (NR; 20% hold-out test within the same cohort), external (n = 196); Total sample: (n = 368)	Urine EVs	[150]
	Prostate Cancer	RF/PLS-DA	AUC: 0.908	Discovery: (n = 80); Training: (n = 167); Validation: internal (n = 72), external (n = 218); Total sample: (n = 537)	Plasma EVs	[151]
Metabolomics/ lipidomics	Gastric Cancer	OPLS-DA	Accuracy: >90%	Discovery: (n = 39); Training: (n = 39); Validation: internal (n = 10, blind test set), external (NR); Total sample: (n = 49)	Urine EVs	[156]
	Colorectal Cancer	SVM/PLS-DA/RF	AUC: 0.985	Discovery: (n = 76, public database); Training: (CV within Discovery n = 76); Validation: internal (CV within Discovery n = 76), external (NR); Total sample: (n = 76)	Fecal Microbial EVs	[157]
	Esophageal Squamous Cell Carcinoma	RF	AUC: 0.98	Discovery: (n = 91); Training: (n = 50, split within patient subset n = 71); Validation: internal (n = 21, split within patient subset n = 71), external (NR); Total sample: (n = 91)	Plasma EVs	[159]
	Hepatocellular Carcinoma	LR/RF/SVM/NN/kNN/OPLS-DA	AUC: 0.90–1	Discovery: (n = 146); Training: (NR; 80% split within Discovery); Validation: internal (NR; 20% split within Discovery), external (NR); Total sample: (n = 146)	Serum EVs	[160]
	Prostate Cancer	OPLS-DA	AUC: 0.85	Discovery: (n = 120); Training: (n = 84, 7:3 split from Discovery); Validation: internal (n = 36, 7:3 split from Discovery), external (NR); Total sample: (n = 120)	Serum EVs	[161]
	Endometrial Cancer	SVM	AUC: 0.924	Discovery: (n = 105); Training: (n = 83); Validation: internal (n = 22), external (NR); Total sample: (n = 105)	Plasma EVs	[163]

Internal validation includes methods such as cross-validation or split-sample validation; external validation refers to validation in an independent cohort. NR, not reported; CV, cross-validation; CT, computed tomography

Challenges

Current evidence shows that ML-driven EV omics has generated many candidate biomarker panels for cancer across diverse clinical contexts. However, most signatures remain at the discovery or early validation stage, and only a small subset has progressed toward regulatory-grade evaluation or clinical implementation. Globally, over 470 clinical trials involving EVs have been registered across more than 200 diseases, but the vast majority remain at exploratory or early validation stages, with only a handful progressing to clinical use or regulatory review [170]. Flagship products such as ExoDx Prostate IntelliScore (EPI) [171], EvoLiver™ [172], and MisyU™ [173] illustrate the feasibility of EV-based diagnostics and the potential added value of AI integration. At the same time, these examples underscore persistent, field-wide obstacles. Key challenges include data governance, algorithmic bias and generalizability, ethics and regulatory compliance, model interpretability, and technical standardization and scalability (Fig. 7).

Fig. 7.

ML-driven EV multi-omics lifecycle, major challenges, and future priorities. The central schematic depicts EV-derived multi-omics data entering an iterative ML pipeline comprising four stages: data, modeling, validation, and deployment. The left side summarizes key challenge domains, including data governance, bias and generalizability, ethics and regulation, model interpretability, and standardization and translation. The right side highlights corresponding priorities for future work, including secure and compliant data use, fair and generalizable models, ethical and regulatory alignment, transparent and interpretable models, and standardized workflows to support clinical adoption

Challenges in data governance

Data are the foundation of ML, and high-quality predictive models require large, well-curated, and well-annotated datasets [174]. In EV research, the International Society for Extracellular Vesicles (ISEV) and its updated MISEV2023 guidelines emphasize the importance of systematic and standardized pre-analytical handling [18]. This includes the timing and method of blood collection, processing delays, and freeze-thaw cycles. Equally important is complete and consistent metadata recording, including key EV-processing steps (e.g., isolation, characterization, and quality control), batch identifiers, and traceable workflows. However, clinical and experimental information is often incomplete. This creates two serious consequences: first, it impedes cross-institutional data integration and limits model generalizability in external cohorts; second, it increases the risk that models learn center-specific or batch-specific signals rather than true disease signals, thereby undermining model validity in real-world applications [175].

Integrating multi-omics and clinical data represents another key barrier. EV-based proteomic, transcriptomic, metabolomic, and lipidomic data are typically generated on different platforms and in different batches, resulting in pronounced batch effects and technical noise [43c, 176]. Missing data within and across omics layers are common, particularly when multi-omics profiles are incomplete or imbalanced across technologies [98c]. In parallel, clinical data deficiencies are equally prominent. Key covariates such as electronic health records (EHRs), imaging data, genomic data, and lifestyle factors are often incomplete, fragmented, or inconsistently recorded in real-world settings [177]. Together, these technical and organizational issues create a fragmented landscape of institutional “data silos”, in which EV omics and clinical data are locked in isolated, non-interoperable systems with limited possibilities for sharing and joint analysis. For ML-enabled EV omics, such silos reduce the effective sample size and the diversity available for model training [178]. They also limit the ability to detect true associations and complicate multicenter external validation and fair benchmarking of models. Furthermore, when high-dimensional omics data are tightly linked to longitudinal clinical information, their identifiability and sensitivity increase substantially, bringing them under strict privacy and data protection frameworks such as the General Data Protection Regulation (GDPR) [179]. This creates major legal and ethical barriers to large-scale cross-institutional data pooling and reuse, hindering the construction of truly large, multimodal, multicenter training and validation cohorts [180].

Therefore, data governance for clinically oriented ML–EV research cannot be addressed by “data cleaning and imputation” alone but requires a coordinated, system-level governance framework. This should include detailed metadata architectures aligned with MISEV2023 and related guidelines, covering pre-analytical handling, EV processing metadata, batch information, quality-control monitoring, and traceable workflows. Such frameworks could be operationalized through an EV-specific reporting and registration platform analogous to EV-TRACK, linking structured protocol reporting with public dataset deposition [181]. In addition, shared data warehouses and analysis environments built on FAIR principles (findable, accessible, interoperable, reusable) should incorporate standardized procedures for batch correction, handling of missing data, outlier detection, and quality control [182]. By promoting interoperability and controlled sharing, such infrastructure can help break down institutional data silos while maintaining local control over sensitive data. Alongside this infrastructure, robust informed-consent language and clear data-use agreements are essential. Data access committees and tiered permission systems can guide responsible data sharing. Federated learning and related multicenter collaboration schemes can then support cross-institutional model training and evaluation without centralizing raw data [183]. For example, ProCanFDL trains deep learning models locally at each site and aggregates only model parameters, enabling joint analysis of multicenter proteomics data while achieving cancer subtype classification performance close to centralized learning [184]. These approaches protect patient privacy while enabling reproducible data integration for ML-enabled EV omics.

Challenges of bias and generalizability

ML models for EV multi-omics face a fundamental challenge: many models use high-dimensional feature spaces but are trained on relatively small, context-specific datasets [185]. Under these conditions, it is difficult to maintain robust performance in multicenter, multiethnic, real-world settings. High-capacity models require adequate sample sizes for stable parameter estimation. When samples are insufficient, models tend to overfit [186]. This leads to unstable decision boundaries and poor performance during external validation. Sensitivity and specificity decline sharply when the model encounters new data. Methodological reviews of oncology ML prediction models show that sample sizes are frequently below recommended thresholds for regression-based models [187]. Formal justifications for sample size are rarely provided. Truly independent external validations are uncommon. Consequently, performance drops, calibration degrades, and generalizability to new populations is limited. Real-world examples illustrate this gap: many cancer prediction models show good discrimination during development, but when evaluated in larger, pragmatic, or multicenter cohorts, they show reduced discrimination and clinical sensitivity/specificity [188]. This highlights the gap between proof-of-concept results and clinical performance.

Algorithmic bias in ML extends beyond simple overfitting. It also includes fairness-related bias and spectrum effects. Conceptual and empirical work in oncology AI shows that biased sampling, non-representative cohorts, and unmeasured structural inequities can cause ML models to disadvantage specific patient groups, even when overall accuracy appears high [189]. A landmark study by Obermeyer et al. provides a compelling example [190]. A widely used commercial algorithm predicted health risk based on past healthcare spending, treating spending as a proxy for illness burden. However, structural inequities mean that Black patients often have lower healthcare expenditures despite greater health needs. As a result, the algorithm systematically underestimated Black patients’ risk and reduced the proportion of Black patients flagged for additional care. This illustrates how target choice can drive bias even when overall accuracy appears high. The spectrum effect represents another source of bias and generalization failure. Diagnostic and risk-prediction tools often perform differently when applied to populations with different disease prevalence, severity patterns, or case-control composition compared to the original development cohort [133d]. By analogy, ML–EV models trained mainly on single-region cohorts, narrow disease causes, or under-representative high-risk groups may not generalize to other etiologies or demographic contexts [185].

Regulatory agencies now recognize these challenges. The FDA explicitly requires medical AI systems to be trained and validated on representative datasets [191]. They also call for multicenter and multiethnic evaluation and for performance to be reported across key subgroups, with ongoing monitoring for bias and performance drift. Addressing bias and generalizability in ML–EV models therefore rests on three elements. First, large, multicenter, and standardized datasets are needed as a foundation. Second, relevant subgroups and classes should be adequately represented to avoid systematic gaps in the data. Third, rigorous external validation should be built into model development and testing, together with policy and technical measures that clearly describe model scope, intended use, and limitations.

Ethical and regulatory challenges

EV omics combined with ML raises distinct ethical and regulatory issues. When high-dimensional omics data are linked with longitudinal clinical information, re-identification risks increase substantially [183b]. Such integrated datasets can directly reveal sensitive attributes including disease risk, genetic susceptibility, and treatment response. This heightens requirements in informed consent, especially for future secondary use, and tightens expectations for de-identification, anonymization, and the design of data security and governance frameworks [179c, 192]. Clinical prediction models require access to sensitive variables such as sex, age, and comorbidities. These features are essential for model accuracy and subgroup analysis, but they also increase privacy risks and the potential for misuse [179d, 193]. Data-protection regimes such as the GDPR stress data minimization, purpose limitation, strict access control, and traceability across the full data life cycle. [179c] For EV-AI projects, this requires clear data boundaries at the design stage, de-identification and pseudonymization where feasible, and encrypted transfer, tiered storage, and auditable access controls [179c, 194]. In closed institutional environments, limited use of data that are not fully anonymized may be acceptable after rigorous ethical and compliance review, provided that access is logged, accountability is defined, and incident-response procedures are in place to reduce harm to patient privacy [183b, 192].

When EV–ML models are embedded in clinical workflows, errors and adverse outcomes raise questions about responsibility and liability. It is often unclear whether responsibility should rest mainly with the clinician, the developer or manufacturer, or the deploying institution [195]. Clarifying these responsibility boundaries is therefore a central ethical and governance challenge for EV-based AI tools. In practice, this requires regulatory frameworks and institutional policies that specify who is responsible for pre-deployment validation, for monitoring model performance and safety in routine care, and for detecting and correcting errors or adverse impacts over time [195]. In addition, minimum requirements for transparency and understandable explanations of model outputs are also necessary. Clinicians and patients should be able to judge when and how model recommendations should influence clinical decisions [195].

To address these challenges, international organizations and regulators are rapidly developing ethical and governance frameworks for medical AI. WHO guidance and UNESCO recommendations on AI ethics establish principles around transparency, accountability, data protection, human rights, and equity [196]. FDA and MHRA guidance on Good Machine Learning Practice emphasize representative datasets, clearly defined intended use, rigorous validation, change management, and post-market monitoring [194b]. Safe, compliant, and sustainable EV-AI deployment requires technical innovation together with robust, transparent data governance and ethical oversight. Patient protections should be embedded throughout design, validation, deployment, and ongoing revision.

Challenges in model interpretability

High-dimensional ML models built on EV proteomic, transcriptomic, and metabolomic or lipidomic data are often treated as “black boxes” [197]. This concern is particularly acute in multi-omics settings, where extremely large feature spaces combine with small sample sizes, pronounced batch effects, and complex model architectures [198]. Such models can raise concerns among clinicians about reliability and among regulators about safety, control, and compliance. Trustworthy medical AI should remain transparent throughout model development, validation, and deployment [199]. This enables bias detection, clearer definition of intended use and scope, and communication of key risks in an understandable way. Although multiple guidelines have outlined requirements for transparency and explainability in medical AI, the level of public reporting for real-world products remains limited [200]. In a survey of commercial medical AI tools, Fehr et al. reported transparency scores between 6.4% and 60.9%, with a median of 29.1% [201]. Major gaps involved training data composition, ethical/legal considerations, and deployment constraints, with little reporting on informed consent, safety monitoring, or GDPR compliance.

These gaps make it difficult to judge software quality and scope of use for regulators, hospital administrators, and researchers. Front-line clinicians need clear evidence on clinical performance, safety, and risks, because they may carry legal responsibility when they use these tools [202]. Patients and the public also depend on transparency to make informed choices and to decide whether an algorithm is appropriate for their own context. Lack of interpretability and transparency is therefore not just a technical issue. It can also weaken the physician–patient relationship, which relies on trust and shared decision-making [203]. When patients cannot understand the basic reasoning and limits behind AI-assisted decisions, they are more likely to question fairness and reliability [204]. This may reduce acceptance of and adherence to medical advice and undermine prevention, early detection, and timely treatment.

Explainable AI methods can help address these problems. Techniques such as feature importance, SHAP, partial dependence plots, and counterfactual explanations enable examination of decision logic and identification of stable, biologically meaningful features [205]. Beyond transparency, explainability can also support hypothesis generation for AI-driven biological discovery in EV nanobiotechnology. Specifically, SHAP and local interpretable model-agnostic explanations (LIME) can decompose predictions into feature-level contributions at both cohort and individual-sample levels [206]. This supports prioritization of EV cargo candidates (proteins/miRNAs/metabolites) and interactions that consistently drive the model, rather than relying on AUC alone. Because local/global attributions may be sensitive to feature collinearity and dataset shift, stability checks across resampling, folds, and clinically relevant subgroups are recommended before drawing mechanistic conclusions [207].

A practical reverse-engineering workflow involves three steps: (i) translating top-ranked features into pathway programs and tissue/cell-type sources by integrating EV-cargo annotations with deconvolution frameworks (e.g., EV-origin) and tumor single-cell atlases; [208] (ii) mapping these candidate features onto EV biology (biogenesis, selective cargo loading) and measurable surface chemistry/biophysics (e.g., glycan patterns, surface charge, size distribution); [18, 209] and (iii) linking them to EV–recipient interactions (uptake routes, tropism, downstream signaling) to formulate testable hypotheses [210]. This linkage is biologically plausible because EV uptake occurs through multiple endocytic routes and is influenced by vesicle/target-cell surface proteins and glycoproteins. For example, tumor exosomal integrins have been implicated in organotropic uptake and pre-metastatic niche education [211], engineering EV surface glycans alters biodistribution and cellular uptake in vivo [212], and EV surface charge (zeta potential) relates to colloidal stability and membrane interactions that can modulate bioavailability [213]. Finally, to demonstrate mechanistic credibility rather than simple pattern recognition, model-identified markers should be supported by orthogonal evidence (e.g., targeted MS/ELISA, single-EV phenotyping, and blocking/perturbation experiments) and reported with EV-specific characterization consistent with MISEV2023 recommendations [18].

To make this reverse-engineering workflow more actionable, Table 2 summarizes practical explainable-AI strategies for translating common XAI outputs (e.g., SHAP/LIME, interaction analyses, and effect-shape plots) into mechanistic hypotheses about EV cargo function in the tumor microenvironment. Organized from left to right, the table shows how these outputs can be linked to pathway programs, tissue or cell-type sources, recipient-cell interactions, and measurable EV physicochemical properties. In this way, it bridges black-box feature attribution with biologically testable EV mechanisms and helps move the analysis beyond simple pattern recognition toward AI-guided biological discovery. Beyond this framework, explainability in EV multi-omics models has at least three core values. First, it helps assess whether model-selected EV marker panels align with plausible pathways and cellular sources [197b, 214]. Second, it can clarify which features drive decisions, where overfitting may occur, and how performance varies across subgroups, supporting regulatory review and quality control [196b, 199a, 215]. This information helps regulators and internal review teams define target populations and intended use scenarios. Third, it helps clinicians interpret individual risk estimates and integrate model outputs with clinical judgement, imaging, and laboratory data [216]. This can make decisions easier to explain and justify and may also improve patient understanding and acceptance. Incorporating interpretability constraints at the model design stage and systematically reporting related information in studies and product documentation can help reduce trust and regulatory barriers posed by “black-box” models [217]. This approach links technical performance to clinical utility while supporting patient-centered decision-making.

Table 2.

How explainable AI outputs can be translated into EV-cargo mechanisms in the tumor microenvironment and anchored to EV biology

XAI output	What it highlights in EV-omics	How to generate a mechanism hypothesis	How to anchor it in EV biology	Ref.
Global attribution (SHAP, permutation importance, grouped Shapley)	Stable cohort-level drivers across EV proteins, miRNAs, metabolites/lipids, and pathway scores	Map top features to pathways and tissue/cell-of-origin signals	Distinguish surface-accessible from luminal cargo to support plausible mechanisms	[208a, 218]
Patient-level explanation (per-sample SHAP/LIME + ALE)	Patient-specific driver cargo and effect-shape patterns	Explain why a sample is high-risk and identify directionality or threshold-like effects	Link feature effects to uptake-related traits or effector cargo programs	[206b, 218d, 219]
Interaction attribution (SHAP interaction values)	Co-acting cargo pairs or modules across omics	Identify non-additive interactions beyond single markers	Support module-level mechanisms, such as uptake determinants paired with effector programs	[220]
TME role inference with cell–cell communication models	Prioritized EV ligands/proteins plus recipient-cell programs from scRNA-seq	Map candidate cargo to ligand–receptor axes and downstream programs in specific TME cell types	Link ML features to recipient-cell signaling mechanisms	[209b, 221]
Uptake/trafficking and EV physicochemical anchoring	Surface proteins, glycans, integrins, zeta potential, and related surface chemistry	Interpret top features as determinants of uptake, tropism, biodistribution, and niche education	Anchor biomarkers to measurable EV properties and cell-interaction routes	[210, 212b, 213]

ALE, accumulated local effects; TME, tumor microenvironment; XAI, explainable artificial intelligence; scRNA-seq, single-cell RNA sequencing

Standardization and translational challenges

Technical standardization and large-scale clinical translation remain major bottlenecks for bringing ML–EV multi-omics from the laboratory into real-world practice. These bottlenecks arise at both the algorithmic level and the EV experimental workflow level.

From the algorithmic perspective, ML-based prediction models have long been developed and reported heterogeneously [105, 222]. Several systematic reviews show that ML-based clinical prediction studies often report incomplete information on data splitting, feature selection, hyperparameter tuning, and final model specifications [222b, 223]. Key details—such as tree depth in RF, penalty/kernel parameters in SVM, or learning rates and subsampling rules in gradient boosting—are frequently missing. These gaps make models hard to reproduce and hinder fair comparison across studies. In response, updated reporting guidelines such as TRIPOD + AI explicitly require clear descriptions of model development workflows, hyperparameter spaces and tuning strategies, as well as internal and external validation schemes [105]. However, the implementation of these standards in current ML–EV studies remains limited. Insufficient standardization directly weakens model reproducibility and transportability and complicates clinical evaluation and regulatory review.

From the EV experimental perspective, heterogeneity between laboratories in sample collection, pre-analytical handling, EV isolation and enrichment, quantification, and reporting is a major barrier to cross-center deployment of ML–EV models. Pre-analytical variables can reshape EV readouts, including anticoagulant choice, time-to-processing, and the initial centrifugation and filtration steps [224]. Storage conditions and freeze–thaw history can cause particle loss and purity changes [225]. These effects can propagate into downstream proteomic readouts and shift ML feature distributions. The MISEV guidelines define minimum requirements for sample sources, pre-processing variables, separation methods, characterization markers, and negative controls to improve transparency and comparability [18, 226]. Nevertheless, systematic reviews indicate that adherence to these key items is still suboptimal in a large proportion of EV publications, with important methodological details frequently missing [227]. Comparative studies further show that different isolation methods can produce systematic differences in particle recovery, protein contamination, EV subpopulation composition, and downstream omics profiles [228]. Without shared standard operating procedures (SOPs) and quality-control metrics, models cannot maintain stable performance across laboratories.

Importantly, EV isolation intrinsically shapes the recovered EV-enriched fraction, because different methods enrich distinct EV subpopulations while introducing varying degrees of non-vesicular background (e.g., lipoproteins, abundant plasma proteins, and protein aggregates) [18, 229]. For example, ultracentrifugation-based workflows, particularly density-gradient ultracentrifugation, can co-isolate proteinaceous particles and lipoproteins such as high-density lipoprotein (HDL) [229a, 230]. In contrast, immunoaffinity capture, often used on microfluidic devices, enriches EVs through predefined surface markers. Although this can improve subtype specificity, it may also under-represent marker-negative vesicles and their cargo [231]. As a result, the same biological condition can yield different EV-omics feature spaces across isolation principles (e.g., miRNA profiles) [232]. This may reduce model transferability across studies using different methods. In addition, workflow variability—such as rotor and centrifugation settings, SEC fraction-window definitions, column/kit/chip lots, operator handling, and device-to-device variability—can act as batch effects that reshape measured EV-omics profiles [229a, 233]. This is highly relevant to ML: if case/control labels, centers, or timepoints are linked to workflow or batch, classifiers may appear to perform well by learning technical or contaminant-related patterns rather than EV-derived oncogenic signals [105, 166b, 233]. Accordingly, Table 3 summarizes the major isolation-specific noise sources, batch-effect drivers, their potential effects on ML bias, and practical mitigation strategies.

Table 3.

Isolation-dependent sources of background noise and batch effects that may bias ML models in EV-omics studies

EV isolation method	Main background noise	Major batch-effect drivers	Potential ML confounding risk	QC and mitigation	Ref.
Differential ultracentrifugation	Protein aggregates; lipoproteins	Rotor setting; pellet handling	Contaminant-driven features; poor transfer	Report key settings; batch-aware split; external validation	[237]
Density-gradient ultracentrifugation	Lipoprotein overlap; fraction-window sensitivity	Gradient lot; fraction boundaries; density calibration	Fraction/batch effects mistaken for biology	Fixed density; predefined pooling; batch-aware split	[238]
Size-exclusion chromatography	Lipoprotein co-elution; column carryover	Column lot; flow rate; fraction definition	Column/fraction patterns become predictive	Fixed columns/fractions; pooled QC; external validation	[239]
Polymer precipitation	Soluble proteins; lipoproteins; polymer residue	Reagent lot; incubation conditions	Non-EV signals dominate features; weak reproducibility	Avoid for discovery omics when possible; cleanup; reagent blanks	[240]
Ultrafiltration or TFF	Co-retained similarly sized particles; membrane bias	Membrane lot/MWCO; fouling; washing efficiency	Membrane-related patterns act as batch effects	Standardized membrane/washing; recovery metrics; orthogonal cleanup	[241]
Immunoaffinity capture	Subpopulation bias; non-specific binding	Antibody/bead lot; epitope accessibility; elution conditions	Overfitting to selected EV subtypes	Define target subtype; track capture efficiency; cross-workflow validation	[242]
Microfluidics	Platform-dependent selection; adsorption artifacts	Chip fabrication; flow control; device drift	Device/run patterns dominate features	Calibration standards; reference sample per run; external validation	[243]

TFF, tangential flow filtration; MWCO, molecular weight cut-off; lot, manufacturing batch of a reagent or consumable

To move from research-grade omics to scalable clinical testing, standardization needs to advance in three linked dimensions. First, at the EV workflow level, SOPs for isolation and characterization should be harmonized across large cohorts and multiple centers, with quantitative QC metrics (e.g., particle recovery, protein-to-particle ratios, marker profiles, and between-batch consistency). Interlaboratory ring trials and external quality assessment programs should calibrate these procedures over time [234]. When multicenter or multi-batch data are involved, validation should reflect real-world deployment (e.g., batch-, center-, and method-aware splits rather than purely random splits) to reduce optimistic bias driven by technical confounding [235]. Second, at the algorithm level, TRIPOD+AI should be implemented consistently, with clear reporting of data partitioning, feature engineering, hyperparameter tuning, and model selection [105]. Robust strategies such as nested cross-validation and independent external validation should be embedded from the outset. Third, at the diagnostic platform level, discovery signatures should be distilled into smaller, technically feasible, and cost-effective targeted panels (e.g., multiplex immunoassays or targeted MS assays) that can be run in accredited laboratories and linked to electronic health-record systems [236]. Collectively, EV-intrinsic heterogeneity and workflow-related variability shape the stability and translational readiness of ML-driven EV omics models and should be addressed in study design and evaluation. With coherent standards across algorithm development, EV workflows, and diagnostic platforms, ML–EV models are more likely to move beyond proof-of-concept and deliver reproducible value in clinical practice.

Practical recommendations for ML–EV omics

Current ML–EV omics research faces linked challenges in data governance, bias and generalizability, ethics and regulation, model interpretability, and technical standardization and translation [196, 199a, 244]. Heterogeneous pre-analytical and analytical workflows, small and non-representative cohorts, incomplete reporting, and limited external validation increase the risk of overfitting and spectrum bias. They also restrict the transportability of published models across centers and populations [136, 222b, 245]. At the same time, evolving data-protection and AI regulations require stronger privacy safeguards, clearer accountability, and higher levels of transparency and interpretability than many current ML–EV studies provide [215b, 216a]. To orient future work toward clinically robust, ethically acceptable, and translationally viable applications, practical recommendations can be drawn from existing methodological and regulatory guidance. Key sources include MISEV2023, TRIPOD+AI, SPIRIT-AI, CONSORT-AI, and Good Machine Learning Practice (GMLP) principles from agencies such as the FDA and MHRA [18, 105, 194b, 246]. These recommendations are summarized in Box 1.

Box 1. Key considerations for designing and reporting ML–EV omics studies.

Study design.

• State a focused clinical question and use case, such as early detection in a defined high-risk group, molecular subtyping, prognosis, or prediction of treatment response.

• Prefer multicenter cohorts with adequate sample size, prespecified inclusion and exclusion criteria, and clear definitions of outcomes and follow-up.

• Reserve an independent external cohort, and if possible a prospective cohort, for final validation. Report performance by key subgroups (for example, age, sex, ethnicity, and stage).

EV workflows and data governance.

• Align pre-analytical handling, EV isolation, and characterization with MISEV2018/2023. Record critical variables, batch identifiers, and quality-control metrics in a structured way.

• Use standardized pipelines for batch correction, handling of missing data, outlier detection, and quality control. Store EV and omics metadata in machine-readable formats to support reuse.

• Ensure compliance with data-protection rules such as GDPR, including clear consent for secondary use, robust de-identification or pseudonymization, and role-based access control with audit trails.

EV omics integration and feature engineering.

• Describe omics platforms, preprocessing steps, and normalization procedures for each data type in sufficient detail to allow replication.

• Apply principled feature-selection and dimensionality-reduction methods, such as regularized regression, tree-based importance measures, or biologically informed filters. Embed these steps within cross-validation to avoid information leakage.

• When several EV omics layers are available, consider integration schemes that use prior knowledge, such as pathways, molecular networks, or cell-of-origin annotations, and state clearly how modalities are combined.

Model development, validation, and reporting.

• Predefine data-partitioning schemes. Use nested cross-validation or similar approaches for hyperparameter tuning, and keep external test sets untouched until final evaluation.

• Report full model specifications, including algorithms, hyperparameters, training procedures, and performance metrics for discrimination, calibration, and clinical utility, following TRIPOD+AI and related standards.

• Examine robustness across relevant subgroups and perform sensitivity analyses for key design choices, such as feature sets, thresholds, batch-correction methods, and alternative algorithms. Share code, model objects, and de-identified data in suitable repositories when governance permits.

Interpretability, ethics, and clinical integration.

• Use interpretable ML tools, such as feature importance, SHAP, partial dependence, or counterfactual explanations, to link EV features to plausible pathways and EV cell-of-origin hypotheses.

• Report model limitations, likely sources of bias, and intended use populations in a transparent way, and align development with ethical and governance frameworks from WHO, UNESCO, and regulators.

• Design ML–EV tools with actual clinical workflows in mind, including decision thresholds, integration with EHR systems, and processes for monitoring performance, updating models, and handling performance drift or adverse events.

Future perspective and conclusion

Over the past decade, rapid advances in both EV research and AI have begun to reshape our understanding of tumor biology and drive progress in clinical oncology. EV-derived multi-omics combined with ML now supports non-invasive early detection, refined risk stratification, therapy selection, and longitudinal monitoring of treatment response and minimal residual disease. For clinical translation, the momentum behind “multi-omics plus AI” is not driven by data volume alone, but also by the limits of single-omics approaches, which increasingly struggle to support actionable conclusions in complex diseases [247]. By integrating EV omics with ML, liquid biopsy is moving toward higher-resolution readouts and more robust prediction models. As single-vesicle analysis technologies and advanced AI models continue to develop, EV-based multi-omics platforms are likely to become increasingly important in precision oncology. This section discusses emerging directions from the perspectives of technology, algorithms, and clinical system development.

Single-vesicle EV omics and high-resolution analysis

Most current ML-driven EV omics studies in oncology rely on bulk vesicle preparations [115, 148, 163]. In these settings, molecular readouts are averaged across large and heterogeneous EV populations. This approach has enabled robust biomarker discovery and predictive modeling, but it inevitably masks rare yet clinically important EV subpopulations [248]. These include tumor-derived or immunomodulatory vesicles that may carry the most informative signals for early detection and treatment stratification. New technologies now make it possible to profile EVs at the single-particle level. Microfluidic platforms [249], nano-flow cytometry [250], super-resolution microscopy [251], and surface-enhanced Raman scattering (SERS) [252] can all be adapted for single-EV analysis. Single-EV transcriptomic and proteomic assays have also been reported [253]. These platforms can measure several biophysical and molecular features per particle in a single experiment. They generate high-dimensional imaging, cytometric, or spectral readouts. High-resolution molecular assays are also improving for variant profiling of EV nucleic acids. One example is SCOPE, which couples Cas13-based target recognition with T7-driven RNA replication and signal amplification in a single-pot workflow [254]. It achieves sub-attomolar sensitivity with single-nucleotide resolution and can detect variant allele fractions as low as 0.01% in spike-in tests [254]. These advances support high-resolution EV omics and may help uncover EV-mediated mechanisms and candidate targets [255].

Integration of single-EV analysis with AI is beginning to emerge. Convolutional neural networks have been combined with single-EV SERS data to detect molecular changes linked to glioblastoma progression [256]. Deep-learning methods have also been applied to single-EV imaging to characterize EV heterogeneity [257]. By contrast, truly omics-scale integration of single-EV datasets with AI remains relatively underexplored. One notable example is a proximity barcoding assay that enables high-throughput single-vesicle proteomic profiling of urinary EV surface proteins [253b]. By combining four ML algorithms, this study identified key single-EV protein markers for early diagnosis and risk stratification of sepsis-associated acute kidney injury. Resources such as the Single Vesicle Atlas (SVAtlas) are starting to integrate single-vesicle datasets across diseases and platforms [258]. These databases support large-scale analysis and biomarker discovery. They can also serve as testbeds for developing and benchmarking ML models for single-EV data, including representation learning, EV subpopulation clustering, and clinical phenotype prediction.

Beyond conventional machine learning, single-EV omics also creates a natural entry point for foundation models and generative AI. In single-cell omics, foundation models are pretrained on very large atlases to learn transferable representations that can be adapted to downstream tasks with limited labels [259]. Geneformer and scGPT (single-cell Generative Pretrained Transformer) are representative examples of such pretrained models [260]. Generative models in the scVI (single-cell variational inference)/scvi-tools (single-cell variational inference tools) family provide a concrete blueprint for handling sparsity, technical noise, batch effects, and missing modalities—issues that are also central to single-EV measurements [261]. Transfer learning for reference mapping, exemplified by scArches (single-cell architecture surgery; reference-atlas mapping via transfer learning), further shows how new datasets can be projected onto atlas representations despite protocol differences [261d]. Domain adaptation in high-dimensional cytometry, exemplified by DeepCyTOF (deep learning–enabled cytometry using domain adaptation), provides additional precedent for learning models that generalize across instruments and batches [262]. In practice, EV-specific foundation or generative models will require atlas-scale single-vesicle data, standardized metadata, and strict safeguards against learning technical artifacts [259, 263]. If these requirements are met, this paradigm may become a key milestone for decoding vesicle-subpopulation heterogeneity and translating rare, clinically informative EV signals into actionable biomarkers.

AI-driven bioinformatics workflows have also been developed to infer small EV secretion from droplet-based single-cell transcriptomics and to link EV programs to tumor progression and treatment response [264]. Looking ahead, single-EV platforms combined with atlas-scale single-vesicle datasets are likely to enable foundation-style pretraining, deep generative modeling, and graph- or probabilistic-learning approaches to better capture EV heterogeneity and cell-of-origin relationships [257b, 258–263, 265]. Critical barriers remain for clinical implementation, including signal dropout, batch effects, compositional bias, throughput limitations, cost, and usability. Addressing these challenges is essential for integrating single-vesicle omics data into robust ML-driven EV multi-omics pipelines.

Interpretable multimodal ML algorithms for EV multi-omics

Looking ahead, EV multi-omics will increasingly rely on interpretable multimodal models and privacy-preserving collaboration. Integrating EV multi-omics with other clinical and molecular modalities may further improve diagnostic accuracy. Explainable AI (XAI) frameworks can reduce “black-box” concerns and strengthen clinical trust [266]. Large language models may support clinical decision-making and communication [267]. These directions require models that are both multimodal and interpretable. For complex multi-omics data, interpretable AI methods such as SHAP, feature ranking, and partial dependence plots are widely used to clarify model decisions and help identify molecular features related to disease and prognosis [204b, 205]. In EV studies, combining these tools with pathway enrichment, cell-of-origin annotation, and longitudinal or treatment-response data helps move from generic “black-box” importance scores toward biologically plausible and experimentally testable hypotheses. For example, they can be used to prioritize EV-borne transcripts or proteins within specific signaling pathways or cellular compartments [204b, 205, 268]. However, current explainable AI methods primarily reveal correlations rather than causality. Their stability and reproducibility across cohorts, platforms, and cancer types remain incompletely characterized. Therefore, they cannot substitute for rigorous causal inference, experimental validation, or mechanistic modeling [204b, 205, 268]. In clinical settings, large language models may help translate model outputs into clearer reports by summarizing results and highlighting uncertainty [269].

These interpretability frameworks can be extended to multimodal designs in which EV multi-omics is treated as an additional molecular layer alongside tissue and blood biomarkers.

Recently reported multi-omics integration methods such as MultiGATE, DeepKEGG, and MWENA embed pathway information and molecular network structure directly into model architectures [270]. Related pathway-informed Transformer models (e.g., Pathformer) further support pathway-aware integration and may be useful when extending multimodal frameworks to EV multi-omics [271]. These approaches can improve both predictive performance and interpretability, and they provide templates for EV-centered, pathway-aware multimodal models. In this setting, EV multi-omics can add vesicle-derived signals that complement conventional biomarkers and help refine patient-level molecular profiles. When data sharing is limited, federated and other privacy-preserving learning strategies enable multicenter model training without pooling raw data. This can enhance robustness and generalizability while maintaining privacy and regulatory standards [272]. Applying these approaches within collaborative EV multi-omics networks may also help address data governance and algorithmic bias. While quantum ML has been proposed for analyzing complex, high-dimensional data, its role remains largely theoretical and unproven in biomedical practice [273]. Until stronger evidence emerges, classical and hybrid ML approaches should remain the standard for EV multi-omics. All such methods should be benchmarked and externally validated, with transparent reporting, before clinical deployment.

EV multi-omics–ML ecosystems for clinical translation

Widespread clinical adoption of ML-based EV multi-omics will depend on a robust ecosystem of standards, infrastructure, and governance rather than on individual models alone. MISEV2018 and MISEV2023 provide core recommendations for EV nomenclature, sample handling, isolation, characterization and functional analysis, helping to reduce pre-analytical and analytical variability [18, 226]. However, ongoing challenges such as technical heterogeneity, incomplete reporting, and analytical inconsistencies still limit clinical translation [244, 274]. Addressing this requires consistent adherence to EV guidelines and harmonized protocols for sample processing, multi-omics generation, quality control, and normalization to limit technical variability. In addition, transparent model development and reporting aligned with TRIPOD+AI, SPIRIT-AI, and CONSORT-AI provide a methodological framework for clinically credible ML–EV studies [105, 246b, 246c].

Large, well-annotated multicenter biobanks that collect EV samples alongside imaging, pathology, genomics, and detailed clinical variables are crucial for training robust models and quantifying center effects [258, 275]. They also enable independent external validation and, when linked to single-EV resources such as SVAtlas, can expand detectable phenotypes and cell-of-origin signatures. In addition, public multi-omics liquid biopsy portals such as cfOmics provide a useful template for EV-focused resources by enabling standardized data access, cross-study benchmarking, and reproducible analysis workflows [276]. A practical priority is to develop end-to-end EV multi-omics–ML platforms that connect standardized pre-analytical workflows, multi-omics data processing, feature engineering, and model training and evaluation within a single, auditable pipeline [245, 277]. These platforms should interface with clinician-facing decision support tools and be interoperable with electronic health records and institutional information technology (IT) systems. Interdisciplinary collaboration among EV biologists, clinicians, data scientists, engineers, and regulatory experts is central to designing such pipelines and ensuring appropriate data governance.

Existing EV-based diagnostic tests, such as the ExoDx Prostate (IntelliScore) assay and EV protein panels for ovarian cancer, illustrate a realistic translational pathway from assay development to clinical use [171a, 278]. This pathway typically involves analytical validation, retrospective and prospective evaluation, demonstration of clinical and economic utility, and eventual incorporation into guidelines and reimbursement frameworks. Future ML-based EV multi-omics tests are likely to follow similar but more complex trajectories. They will require explicit procedures for algorithm versioning, rigorous external validation, assessment of robustness and calibration, and ongoing monitoring for data drift in real-world practice. Within a mature EV multi-omics–ML ecosystem, these activities should be supported by shared standards and interoperable platforms so that models can be updated, revalidated, and retired in a controlled and transparent manner.

In conclusion, ML-enhanced EV omics provides a powerful framework for developing liquid biopsy tools for cancer screening, early detection, risk stratification, and therapy monitoring. Translating these advances into clinical practice will require systematic progress across five interconnected domains. These include (i) robust data governance and standardized EV/omics protocols; (ii) rigorous bias mitigation and external validation in large, diverse cohorts; (iii) transparent, interpretable models that support clinical decision-making; (iv) formal regulatory pathways and ethical oversight; and (v) collaborative frameworks aligning researchers, clinicians, regulators, industry, and funders. Particular emphasis should be placed on unified technical standards and biomarker validation frameworks that bridge the translational gap between discovery studies and real-world implementation. By addressing these priorities in concert, ML-driven EV omics can move beyond proof-of-concept to deliver reliable, deployable tools that meaningfully improve cancer prevention, diagnosis, and treatment in routine clinical care.

Author contributions

ZW drafted the manuscript and led the revisions. LW prepared Figs. 1, 2, 3 and 7. ZG, HL, and ZY contributed to manuscript revision. ZW and SM contributed to manuscript writing. ZH and XS contributed to manuscript editing. YX contributed to manuscript revision and provided overall guidance on the manuscript framework. JWPY supervised the project, contributed to manuscript revision, and provided overall conceptual guidance and direction. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (82573164, 82472743, 82300921); the Guangxi Science and Technology Program (AD25069077); the Natural Science Foundation of Heilongjiang Province (LH2023H043); the Open Funds of the State Key Laboratory of Oncology in South China (HN2025-02); the Open Funds of the Shanghai Key Laboratory of Cancer Systems Regulation and Clinical Translation (CSRCT-KF0002); the Open Research Project from the Key Laboratory of Clinical Laboratory Medicine of the Guangxi Department of Education (GXGXLCJYZDX2025001); the Open Research Fund of the Key Laboratory of Gastrointestinal Cancer (Fujian Medical University), Ministry of Education (FMUGIC-202501); the Open Funds of the Key Laboratory of Clinical Laboratory Technology for Precision Medicine (FKLCLT-202504); the Open Research Fund of the Anhui Province Key Laboratory of Non-coding RNA Basic and Clinical Transformation (NcRNA202508); the Open Project of the Anhui Provincial Key Laboratory of Tumor Evolution and Intelligent Diagnosis and Treatment (KFKT-202505); the Thematic Research Support Scheme of the State Key Laboratory of Liver Research, The University of Hong Kong (SKLLR/TRSS/2025/08), and the Research Grants Council Theme-Based Research Scheme (Project No. T12-716/22-R).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Competing interests

The authors declare no competing interests.