Artificial intelligence for precision oncology from phenotyping and drug discovery to clinical translation

Abstract

Artificial intelligence (AI) is reshaping oncology by extracting clinically actionable signals from complex cancer data and accelerating drug development. In this narrative review, we summarize how machine learning and deep learning models support cancer phenotyping, including tumor detection, molecular subtyping, prognosis, and treatment response prediction across histopathology, radiology, and multi-omics data. We then discuss AI-enabled virtual screening, drug repurposing, generative molecular design, and hybrid computational–experimental pipelines that streamline oncology drug discovery and optimization. Cross-cutting limitations are examined, including data quality and representativeness, class imbalance, bias and fairness, model interpretability, and ethical, privacy, and regulatory challenges in clinical deployment. Finally, we highlight emerging directions such as multimodal foundation models, federated learning, AI stewardship, and patient-specific digital twins, and outline a roadmap for integrating trustworthy AI into precision oncology. Realizing the full potential of AI will require rigorous validation, transparent reporting, and close collaboration between clinicians, data scientists, regulators, and patients to ensure equitable, patient-centred benefit.

Graphical Abstract

Introduction

Cancer remains a global health challenge, causing about 10 million deaths in 2020, and incidence is expected to increase substantially by 2050 [1]. Therapy failure reflects profound biological complexity. Tumors display marked intra- and intertumoral heterogeneity, with multiple subclones and molecular profiles coexisting within a single patient, undermining uniform diagnostics and standardized treatments [2, 3]. Conventional pathology and single-gene biomarkers often fail to capture this complexity, contributing to suboptimal treatment selection and acquired resistance [4]. There is a need for advanced approaches to characterize tumors, predict outcomes, and guide therapy decisions at the individual-patient level (Fig. 1).

Fig. 1.

The graphic abstract summarizes how machine learning and deep learning models extract phenotypic and molecular features from histopathology, radiology, and multi-omics data to support cancer detection, subtyping, prognosis, and treatment selection. AI-enabled virtual screening, drug repurposing, de novo molecular design, and closed-loop experimental pipelines are shown accelerating oncology drug discovery. The right panel illustrates clinical deployment, including decision support, trial matching, and governance frameworks addressing bias, privacy, and regulation, highlighting feedback loops between research and real-world practice

Over the past two decades, artificial intelligence (AI), particularly machine learning (ML), has emerged as a core strategy to meet this need by extracting structure from high-dimensional biomedical data beyond human cognitive capacity [4, 5]. Early applications using classical ML on genomic data showed that neural networks and support vector machines applied to gene-expression profiles can identify cancer subtypes and prognostic signatures overlooked by traditional analyses, but these models frequently overfit small cohorts and lacked generalizability [6, 7]. The rise of deep learning (DL) subsequently provided enhanced capacity to model complex, non-linear relationships in multimodal cancer data. Convolutional and graph neural networks integrate histopathology, radiology, multi-omics, and clinical variables to detect and classify tumors, stratify risk, and predict response, often outperforming single-modality models [8, 9]. These successes demonstrate that data-driven approaches can reveal subtle diagnostic and predictive signals, while exposing persistent challenges, including the need for large, diverse training datasets and improved interpretability of “black-box” models in clinical practice [8].

In parallel, AI is reshaping oncology drug discovery and development [10]. Key applications across virtual screening, drug repurposing, generative molecular design and AI-integrated experimental pipelines are summarized in Fig. 2; Table 1. Conventional pipelines are slow, costly, and characterized by high attrition [11, 12]. AI can increase efficiency by prioritizing promising therapeutic hypotheses. Ligand-based models, including neural-network quantitative structure–activity relationship (QSAR) approaches, predict anticancer activity from chemical structure to enable virtual screening of large libraries [13]. Network-based algorithms exploit known relationships among drugs, genes, and diseases to propose new drug–target or drug–disease links and repurposing opportunities [14]. Building on these foundations, modern DL methods support more powerful virtual screening and de novo design: deep neural networks refine molecular docking predictions and integrate pharmacogenomic screens to nominate repurposing candidates for specific molecular subtypes, while generative models design novel molecules optimized for potency, selectivity, and drug-like properties [15–18]. Collectively, these AI-driven strategies expand searchable chemical space and accelerate lead identification.

Fig. 2.

Artificial intelligence in oncology drug discovery and development. The workflow illustrates stages from target identification and hit discovery to lead optimisation and early safety profiling. Virtual screening combines docking with machine-learning scorers to enrich active compounds, while network-based models and transcriptomic signatures support drug repurposing and rational combinations. Generative models design novel molecules under multi-objective constraints on potency, selectivity, and ADMET properties. AI-guided active learning loops connect in silico prediction with cell-based, organoid, and animal experiments to iteratively refine promising candidates

Table 1.

Representative AI applications across the precision oncology pipeline

Pipeline stage	Task/clinical scenario	Data modalities	Typical AI methods	Representative studies	Key contributions & main limitations
Cancer phenotyping	Automated tumour detection and grading in digital pathology	H&E whole-slide images	Weakly supervised CNNs, multiple-instance learning, attention-based models	Campanella et al. [32] (clinical-grade WSI classification); recent hybrid attention models for colon cancer histopathology.	Contributions: Approach or match expert pathologists for prostate, breast and colorectal cancer detection; enable high-throughput slide triage and consistent grading. Limitations: Performance drops across institutions due to domain shift (scanner, staining); limited interpretability; need regulatory-grade prospective validation.
Cancer phenotyping	Lesion detection, segmentation and staging on CT/MRI/PET	Radiology images (CT, MRI, PET/CT)	U-Net and variants, 3D CNNs, hybrid Inception-U-Net with optimization or attention modules	UIGO Inception-U-Net with gravitational optimization for liver tumour segmentation and feature selection.	Contributions: Improve sensitivity/specificity for lung and liver tumour detection, delineate volumes for radiotherapy planning, support ultra-low-dose protocols. Limitations: Dependence on labelled volumes; generalizability issues across scanners/protocols; overfitting on small or imbalanced datasets.
Cancer phenotyping	Molecular subtyping and survival/risk stratification	Bulk and single-cell omics, spatial transcriptomics, clinical variables	Classical ML (SVM, RF) and modern multimodal deep learning (transformers, GNNs)	Early microarray classifiers (Golub 1999; Alizadeh 2000) and recent pan-cancer multimodal survival models integrating histology and genomics.	Contributions: Reveal novel molecular subtypes and prognostic signatures; integrate spatial and molecular context of the tumour microenvironment; enable pan-cancer risk models. Limitations: Batch effects, small cohorts and class imbalance; models trained on TCGA-like data underperform in under-represented populations; mechanistic interpretability remains limited.
Drug response prediction (preclinical/translational)	Predicting anticancer drug sensitivity in cell lines, PDX models and patients	Multi-omics of cancer cell lines or tumours (RNA-seq, mutations, copy-number), drug descriptors/structures	Hybrid ML/DL (GNNs, attention networks, residual architectures; CGAN-assisted feature learning)	Singh et al. 2025 comprehensive review of ML/DL models for anticancer drug response using CCLE and GDSC.	Contributions: Prioritise effective drugs or combinations based on tumour molecular profiles; identify predictive biomarkers of response; enable virtual “in silico screens” before wet-lab testing. Limitations: Label noise and assay variability; difficulty transferring models from cell lines to patients; limited coverage of rare drugs and rare cancer subtypes.
Virtual screening & docking	Enrichment of hits against oncology targets (e.g., kinases, epigenetic enzymes)	Protein structures, large virtual small-molecule libraries	Deep-learning-based docking scorers, post-docking ML classifiers, GNNs for binding affinity	Machine-learning filters on docking to identify PI3K/tankyrase inhibitors with improved hit rates over docking alone.	Contributions: Reduce the number of compounds that must be synthesised/tested; improve early enrichment and recover non-obvious chemotypes; support target-focused and polypharmacology screens. Limitations: Susceptible to artifacts and dataset bias; limited robustness to novel chemotypes and new targets; docking scores and in vitro potency do not always translate to in vivo efficacy.
De novo molecular design	Designing novel anticancer agents with desired potency, selectivity and ADMET	Chemical structures (SMILES/graphs), bioactivity and ADMET datasets	Variational autoencoders, GANs and diffusion models; reinforcement learning; scaffold-aware generative models	Genotype-conditioned diffusion models that generate small molecules tailored to tumour mutation profiles; Multi-objective scaffold-aware VAEs optimising potency and drug-likeness.	Contributions: Explore vast chemical space beyond existing libraries; co-optimise potency, selectivity and pharmacokinetics; suggest isoform-selective or multi-target agents. Limitations: Synthesizability and retrospective “benchmark overfitting”; generated molecules still require multiple medicinal-chemistry and experimental refinement cycles; uncertain generalisation beyond training chemistries.
AI-integrated experimental pipelines	Closed-loop hit finding, lead optimisation and early safety profiling	High-throughput screening data, omics readouts, toxicity assays	Active learning, Bayesian optimisation, self-driving lab orchestration	AI-guided screens that reduce the number of tested compounds by orders of magnitude while discovering potent DDR1 or IDH1 inhibitors; GNN-based ADMET predictors integrated as toxicity filters.	Contributions: Dramatically increase efficiency of hit identification; prioritise safer chemical matter early; enable iterative “design-test-learn” cycles tightly linking computation and experiment. Limitations: Integration with lab automation infrastructure is complex; performance can drift as assay conditions change; models require continuous retraining with new experimental data.
Clinical decision support & trial matching	Matching patients to targeted therapies and precision oncology trials	Structured EHR data, tumour sequencing reports, clinical trial registries	Rule-based eligibility engines, ML ranking models, NLP for unstructured eligibility criteria	MatchMiner and MatchMiner-AI platforms for computational trial matching in precision oncology programmes	Contributions: Increase identification of actionable alterations and trial opportunities; standardise interpretation of genomic profiles; reduce manual screening burden for clinicians. Limitations: Dependent on up-to-date trial registries and harmonised variant annotation; potential bias if models are tuned on trials available only at large centres; limited evidence yet that algorithmic matching improves survival.
Clinical deployment, ethics and governance	Safe, equitable implementation of AI tools in oncology practice	De-identified and federated clinical data; operational EHR and imaging streams	Federated learning, uncertainty-aware models, explainable AI (saliency maps, SHAP), AI stewardship frameworks	Federated learning for rare cancer boundary detection and multi-centre pathology; FDA-approved AI pathology systems as “second readers”.	Contributions: Address data-sharing constraints while leveraging multi-centre cohorts; improve transparency and clinician trust through XAI and model-facts labels; promote continuous post-deployment monitoring. Limitations: Regulatory pathways for adaptive models remain immature; genomic privacy and informed consent are unresolved in many jurisdictions; risks of algorithmic bias and under-performance in under-represented groups if data diversity is not ensured.

H&E: hematoxylin and eosin; WSI: whole-slide image; CT: computed tomography; MRI: magnetic resonance imaging; PET: positron emission tomography; PET/CT: positron emission tomography–computed tomography; CNN: convolutional neural network; GNN: graph neural network; SVM: support vector machine; RF: random forest; PDX: patient-derived xenograft; CGAN: conditional generative adversarial network; CCLE: Cancer Cell Line Encyclopedia; GDSC: Genomics of Drug Sensitivity in Cancer; ADMET: absorption: distribution: metabolism: excretion and toxicity; DDR1: discoidin domain receptor 1; IDH1: isocitrate dehydrogenase 1; EHR: electronic health record; NLP: natural language processing; FDA: Food and Drug Administration

A central frontier is translating AI-derived insights into clinical benefit by integrating computation with experimental biology and clinical workflows [5]. Hybrid pipelines couple virtual screening or risk-prediction models with iterative wet-lab validation and AI-based ADMET prediction to filter unsafe compounds early [11]. In precision oncology, AI systems are being deployed to match patients to targeted therapies or clinical trials based on tumor molecular profiles; AI-enabled pathology platforms have already achieved regulatory clearance [19, 20]. Nonetheless, substantial barriers remain, including data quality and sharing constraints, bias and fairness concerns, limited model interpretability, and regulatory and ethical hurdles [21].

In the following sections, we review state-of-the-art AI applications across the oncology pipeline, from cancer phenotyping to drug discovery and clinical integration, analyze cross-cutting limitations, and outline future directions such as multimodal foundation models, federated learning, and AI-driven digital twins aimed at realizing the full potential of AI in cancer care.

AI for cancer phenotyping and patient stratification

Accurate cancer phenotyping – identifying tumor type, molecular subtype, and likely disease course – is central to treatment selection [22]. An overview of AI-enabled cancer phenotyping and patient stratification is summarized in Fig. 3. AI methods increasingly support this process by detecting complex patterns in tumor data that correlate with outcome [23]. The evolution of AI in cancer phenotyping can be framed in two eras: an early phase of classical machine learning on limited datasets and a modern phase dominated by deep learning on multimodal big data, both aiming to surpass human assessment and simple biomarkers [22, 23].

Fig. 3.

Artificial intelligence for cancer phenotyping and patient stratification. The schematic depicts multimodal data sources, including whole-slide histopathology images, radiology scans, bulk and single-cell omics, spatial transcriptomics, and clinical variables. Modality-specific encoders such as convolutional, graph, and transformer networks feed a fusion layer that outputs tumour detection, grading, molecular subtype, risk scores, and predicted treatment response. Examples of weakly supervised pathology models, radiology segmentation networks, and survival predictors are indicated. Icons representing saliency maps and feature-attribution plots illustrate how explainable AI supports clinician trust

Early machine learning approaches in the 2000 s provided proof-of-concept. Algorithms such as neural networks, decision trees, and support vector machines were applied to gene-expression microarrays, revealing previously unrecognized tumor subtypes and risk groups [22, 24]. Autoassociative neural networks distinguished leukemia subtypes from latent gene-expression features, and clustering combined with SVMs improved classification from noisy microarray data [22]. These models often achieved cross-validated sensitivities and specificities above 80%, showing that relatively simple AI could outperform single-marker analyses [22]. However, they were frequently overfitted and poorly generalizable: performance often dropped on external cohorts because thousands of features were modeled on small samples, yielding spurious associations [25, 26]. Extensive feature selection and expert-driven preprocessing further limited scalability, and many models yielded little mechanistic insight [25]. This era nonetheless established rigorous cross-validation and external validation in cancer genomics and underscored the need for larger, more diverse datasets and more expressive models, motivating the transition to deep learning [22, 25].

Deep learning for multimodal data (mid-2010s onward) has greatly expanded AI’s role in cancer phenotyping. Enabled by large cohorts from international consortia, researchers now train deep networks that automatically learn representations from raw inputs [23, 27]. Rather than analyzing one modality in isolation, many systems integrate whole-slide pathology images, radiology scans, DNA/RNA sequencing, and clinical parameters into unified models [23]. Modality-specific encoders (e.g. CNNs for images, transformers or graph networks for genomic or spatial data) feed into an integration layer that predicts endpoints such as subtype, risk score, or therapy response [23, 28]. By linking cross-modal patterns – associating a radiographic feature with an RNA signature, for instance – multimodal networks often exceed the performance of single-modality models and can reveal “imaging-genomic” subtypes with distinct prognoses [23, 27]. In lung cancer, combined models that fuse CT imaging and mutation data improve prediction of histologic subtype and survival [29, 30]. Graph neural networks applied to spatial transcriptomics, treating cells as nodes connected by physical proximity, accurately classify spatial neighborhoods in tumors, illustrating AI’s ability to quantify tissue architecture beyond conventional pathology [28, 31].

Deep learning has also transformed purely image-based phenotyping. In digital pathology, weakly supervised CNNs trained on large collections of whole-slide images match expert pathologists in detecting tumor regions in cancers such as prostate and skin [32–34]. In radiology, DL models analyzing CT or MRI scans both detect lesions and compute radiomic features that correlate with grade or aggressiveness [35–37]. U-Net–based architectures segment tumors with greater speed and consistency than manual contouring and can identify subtle textural or morphological patterns imperceptible to human observers, enabling earlier or more refined detection [38–40].

Despite these advances, deep learning in oncology faces challenges. Robust training requires large, balanced datasets; otherwise, highly parameterized networks may overfit or encode site-specific bias [23, 41]. Models trained to near-perfect accuracy on one institution’s data often generalize poorly to others because of differences in patient populations, imaging protocols, or slide preparation [33]. Data-augmentation strategies can partially mitigate this, but more structural solutions are emerging [33]. Federated learning allows institutions to collaboratively train models without sharing raw data, enabling learning from broad patient populations while preserving privacy[42, 43]. Early studies indicate that such models can approach the performance of centrally pooled training and yield more generalizable tumor classifiers than single-site models [42].

Interpretability remains a major limitation. Deep networks often behave as “black boxes,” offering limited explanation for their outputs [44]. In clinical decision-making, oncologists must understand why an AI system labels a tumor as high-risk before modifying therapy [23, 44]. This has driven a focus on explainable AI (XAI) in oncology. Techniques such as Grad-CAM generate heatmaps on medical images indicating regions that most influenced a malignancy prediction, and feature-attribution methods such as SHAP quantify the contribution of genes, radiomic features, or clinical variables to a risk score [44, 45]. In breast cancer detection, Grad-CAM overlays on mammograms have helped radiologists verify that a CNN is focusing on the tumor mass rather than artifacts, increasing trust and facilitating adoption [44]. In many settings, a slightly less accurate but interpretable model may be preferable to an opaque model with marginally higher performance.

AI systems for cancer phenotyping are moving from retrospective analysis to prospective clinical deployment. AI-assisted digital pathology platforms have obtained regulatory authorization for primary diagnosis, acting as a “second reader” that flags suspicious regions on slides to improve efficiency and consistency [33, 41]. In radiology, algorithms for lung-nodule detection on CT are under evaluation as triage tools to help manage rising imaging workloads [35]. Beyond diagnosis, AI-based prognostic models are being tested to stratify patients in clinical trials, for example distinguishing high-risk early-stage lung cancer patients who may benefit from adjuvant therapy from low-risk patients who can avoid overtreatment [23, 27]. Major cancer centers and precision oncology companies now use AI to integrate genomic and clinical data at scale to recommend targeted therapies and match patients to clinical trials, with pilots reporting improved identification of actionable alterations and trial opportunities compared with standard workflows [46, 47]. These developments underscore that AI-driven phenotyping is beginning to influence real-world care, but its ultimate impact will depend on rigorous prospective validation, equitable performance across populations, transparent outputs, and sustained clinician trust.

AI in oncology drug discovery and development

Virtual screening and drug repurposing with AI

One of AI’s earliest and most immediate contributions is improving compound screening for anticancer activity [48]. Instead of experimentally testing hundreds of thousands of molecules, researchers now use ML models to virtually evaluate compounds and prioritize a smaller, enriched subset for assays [49]. Neural network–based quantitative structure–activity relationship (QSAR) models ingest large molecular libraries and predict which molecules are likely to inhibit a cancer cell line or target protein [50]. In an anti-metastatic setting, a neural QSAR model achieved ~ 70% accuracy in classifying new compounds as active or inactive – a substantial enrichment over random selection in a chemical space where actives are rare and screening is expensive [50].

AI has also enhanced structure-based virtual screening through integration with molecular docking [51]. Docking simulations fit ligands into a protein’s 3D structure and rank poses with heuristic scoring functions, but these scores are noisy and yield many false positives [16]. Deep learning models trained on docking and activity datasets learn which pose and ligand features correlate with binding [16, 52]. Post-docking neural classifiers have reached high discriminative performance (AUC around 0.9 for kinase targets) and helped identify dual PI3K/tankyrase inhibitors for colorectal cancer, with several predicted molecules validated experimentally [53, 54]. Acting as smart filters on top of docking, such models reduce false positives and focus attention on a shortlist enriched in genuine binders [55, 56].

Ranking-based approaches further extend virtual screening and repurposing. Instead of binary active/inactive labels, learning-to-rank models order compounds by predicted efficacy or signature-reversal strength [57]. This is especially useful for drug repurposing, where the goal is to prioritize existing drugs for a specific molecular context [58]. In transcriptomics-driven studies, ML models have ranked marketed drugs by their ability to reverse cancer gene-expression signatures; top-ranked agents show higher hit rates in cell-based assays than randomly chosen drugs [57, 58]. A notable example in triple-negative breast cancer (TNBC) used tumor expression signatures to nominate lapatinib – a HER2-targeted drug not traditionally used in TNBC – as a high-priority candidate, and subsequent experiments confirmed its activity in TNBC models [59, 60]. Such examples illustrate how AI can expose non-obvious drug–disease relationships and accelerate repositioning of approved drugs.

Network-based models and polypharmacology

Many tumors rapidly evolve resistance to single-target therapies, motivating interest in polypharmacology and rational drug combinations. Network-based AI models construct graphs linking genes, proteins, pathways, drugs, and diseases and then use graph analytics or graph neural networks to identify hub nodes and vulnerable subnetworks [61, 62]. These analyses can highlight targets or combinations likely to exert synergistic effects and can suggest repurposing candidates for specific tumor types [63].

In practice, such models often rediscover known drug–target relationships – an important sanity check – while also generating novel, mechanistically plausible hypotheses [61]. Some network studies have identified metastasis hubs and proposed natural compounds predicted to inhibit them; others have mined gene–drug–disease networks in lung and other cancers to nominate repurposing candidates that converge on shared pathways [64, 65]. Although only a subset of predictions has progressed to clinical evaluation, early crossovers are encouraging. AI-driven analysis implicated the anti-inflammatory drug celecoxib as a potential inhibitor of the epigenetic enzyme OGT; subsequent biochemical and cellular assays confirmed OGT inhibition and anticancer activity, and celecoxib-derived scaffolds are now being explored further [66, 67]. These examples highlight how network-oriented AI can build a richer hypothesis space than intuition or single-pathway analyses alone.

Generative models and multi-objective molecular design

While the approaches above primarily select or rank existing molecules, generative AI introduces a paradigm in which models design novel compounds de novo [68]. Variational autoencoders, generative adversarial networks, and diffusion models have been trained on large collections of chemical structures to learn a latent “chemical language.” [68, 69]. Conditioned on design objectives, these models can output new molecules one atom or bond at a time [70]. A key strength of generative methods is their ability to support multi-objective optimization: beyond predicted potency against a target, they can incorporate constraints on selectivity, physicochemical properties, and ADMET profiles [71].

Genotype-conditioned diffusion frameworks have been used to propose inhibitors predicted to be particularly effective in cells harboring specific oncogenic mutations, yielding diverse molecules that satisfy potency and selectivity criteria better than earlier algorithms [72, 73]. Autoencoder-based pipelines have designed isoform-selective PI3Kα inhibitors, with several AI-generated compounds showing strong biochemical selectivity over other family members, illustrating the potential to minimize off-target toxicity [74, 75]. Generative frameworks have also been applied to multi-target design, for example creating compounds that modulate multiple angiogenesis pathways simultaneously [76].

Generative design loops often integrate predictive models for ADMET to enable in silico “fail fast” strategies [77, 78]. A typical workflow generates candidate molecules, predicts binding, solubility, permeability, metabolic stability, and organ-specific toxicities, and then discards structures with unfavorable profiles before synthesis [79, 80]. This allows medicinal chemists to concentrate resources on a small set of high-value candidates. Nonetheless, AI-generated molecules almost always require iterative medicinal chemistry refinement and additional modeling rounds before emerging as viable leads, emphasizing that generative AI augments rather than replaces human expertise [81, 82].

AI-integrated experimental pipelines

Regardless of predictive accuracy, any proposed drug must be validated experimentally and clinically. The field is therefore converging on hybrid AI–experimental pipelines that couple in silico predictions with iterative wet-lab feedback [83]. In such workflows, AI models first screen chemical libraries or multi-omic cancer datasets to generate hypotheses – top-ranked targets or compounds [83]. These hypotheses are then tested using biochemical, cell-based, or organoid assays, and the resulting data feed back into the models for retraining or refining the search space, a process often formalized as active learning [84, 85]. The loop repeats until sufficiently potent, selective, and safe leads are identified.

These integrated strategies can yield substantial efficiency gains. Instead of screening 100,000 compounds, an AI-guided campaign may test a few hundred, learn from the outcomes, and converge on high-quality hits in a fraction of the time and cost [83, 86]. After a docking-based virtual screen for kinase inhibitors, for example, an SVM post-filter achieved high accuracy in distinguishing true actives from inactives in follow-up assays, markedly improving hit rates compared with unfiltered docking [53]. Similar campaigns using graph convolutional neural networks to prioritize compounds for epigenetic targets have shown high concordance between predicted and measured inhibition, further validating the “AI triage then experiment” strategy [87].

AI is also being incorporated into early safety and pharmacokinetic evaluation. ML models for toxicity – including graph neural networks trained to predict liver, cardiac, or hematologic toxicities – now achieve high discriminative performance in benchmark datasets [88, 89]. In hybrid pipelines, these models triage AI-generated or screened compounds, flagging high-risk structures before animal studies and prompting redesign or termination of problematic chemotypes [90]. Such AI-based ADMET filters can substantially reduce the need for certain animal studies and help ensure that only comparatively “clean” candidates progress [88, 90]. Collectively, AI is now embedded across the oncology drug-development continuum. It broadens the search space in the ideation phase by scanning vast chemical and biological landscapes, and then narrows this space by steering experimental validation toward the most promising targets and molecules [83]. Synergy between AI and human expertise has already yielded tangible successes: celecoxib’s repurposing as an OGT inhibitor and lapatinib’s non-intuitive application in TNBC exemplify AI-enabled discoveries subsequently confirmed in cell-based models [66]. At the same time, real-world deployment has surfaced practical challenges, including integration of computational tools with laboratory automation, maintenance of model calibration as new data accumulate, and assurance that predictions remain robust across cancer subtypes and patient populations [85, 91]. Ongoing efforts to build cloud-based infrastructures linking AI platforms directly to robotic experimentation may ultimately enable largely automated, closed-loop AI-driven discovery.

Challenges and limitations in translating AI to oncology

Data quality, bias, and generalizability

AI algorithms are only as good as the data used to train them. Cross-cutting challenges and potential solutions for trustworthy clinical translation are depicted in Fig. 4. In oncology, obtaining large, high-quality, and representative datasets remains a core limitation [23]. Many early studies reporting high accuracies were later shown to be overfit: models memorized quirks of the training cohorts and failed on new patients [25]. In early cancer microarray work, predictors that achieved ~ 95% accuracy in one cohort often dropped to near chance-level on external data due to batch effects and population differences [22]. Batch effects—systematic technical discrepancies between datasets—are pervasive in genomics and difficult to correct fully [92]. Equally problematic is lack of representativeness. Large resources such as The Cancer Genome Atlas (TCGA) overrepresent certain ethnicities and common tumor types, so models trained on these data may perform poorly in African or South Asian populations or in rare subtypes absent from the training set [47, 93]. This raises concerns about health disparities: without safeguards, oncology AI systems risk preferentially benefiting groups already advantaged in healthcare [94]. Similar failures have been reported outside oncology, where dermatology AIs underperform on darker skin tones because training images were predominantly light-skinned [95]. Comparable biases are likely in cancer imaging and pathology if training data are geographically or demographically narrow [95].

Fig. 4.

Cross-cutting challenges and future directions for AI in precision oncology. Central panels summarize key limitations, including data quality, batch effects, class imbalance, demographic under-representation, model interpretability, dataset shift, and regulatory and ethical constraints in real-world deployment. Surrounding elements highlight enabling solutions such as federated learning, uncertainty-aware and explainable models, privacy-preserving analytics, and bias auditing. A forward-looking section depicts multimodal foundation models, digital twins, and AI stewardship programmes integrating continuous monitoring, clinician education, and governance to promote robust, equitable clinical translation worldwide

Class imbalance is another major challenge. In drug discovery, active compounds may be outnumbered by inactives, and in clinical prediction, adverse events such as recurrence are rarer than non-events [96]. In such settings, a classifier can appear highly accurate by always predicting the majority class [97]. Appropriate metrics (e.g. precision–recall curves) and techniques such as minority oversampling and algorithmic reweighting are therefore essential [97]. Methods like SMOTE for synthetic minority sampling and data augmentation for images can partly mitigate imbalance, but there is no substitute for more real, diverse data [98]. Consequently, there is growing emphasis on multi-institutional cohorts and prospective data collection in underrepresented populations, and several initiatives now mandate demographic diversity and bias auditing in clinical AI development.

Collaborative training paradigms aim to improve generalizability without compromising privacy. Federated learning keeps data local by sending models to participating sites and aggregating parameter updates [99]. Federated models for cancer imaging and genomics have matched centralized training while learning from broader, heterogeneous populations [99]. Multi-site MRI tumor segmentation models, for example, generalize better across scanners and demographics than single-center models [39]. Nevertheless, federation introduces challenges, including variable data quality and inter-site distribution shifts [99]. Other partial remedies include de-identified open datasets, data trusts, and carefully validated synthetic data [100].

Overall, the next generation of oncology AI depends on data quality and representativeness. The field is increasingly incorporating fairness metrics and bias-mitigation strategies such as adversarial training to reduce reliance on sensitive attributes. The overarching goal is robust, equitable performance so that the benefits of precision oncology are shared across all patient populations rather than concentrated in a privileged subset.

Interpretability and transparency of AI models

Lack of interpretability is a major barrier to deploying powerful AI models in oncology. Many high-performing systems, especially deep neural networks with millions of parameters, act as “black boxes” that output predictions (e.g. “90% chance of recurrence in 5 years”) without a clear rationale [101, 102]. This opacity creates problems. Clinicians are reluctant to rely on AI for high-stakes decisions such as therapy selection if they cannot interrogate the reasoning, because medical decisions typically require justification—pathologists, for instance, point to specific morphologic features when calling a tumor malignant or benign [103]. If an AI system cannot provide analogous evidence, its recommendations are likely to be discounted. Scientifically, opaque models also limit discovery: interpretable systems that highlight key genes, pathways, or image patterns can reveal mechanisms of drug response or resistance and suggest new biomarkers, whereas black-box models offer little direct biological insight [45, 99].

These concerns have driven growth of explainable AI (XAI) in oncology. For image-based models in pathology and radiology, saliency mapping is widely used [104]. Methods such as Grad-CAM (Gradient-weighted Class Activation Mapping) generate heatmaps showing which regions of an image most influenced a prediction [105]. In mammographic breast cancer screening and digital pathology, Grad-CAM overlays can demonstrate that a convolutional network focuses on suspicious lesions or tumor regions rather than artifacts, giving clinicians confidence that the network is “looking” at relevant areas [105, 106].

For non-image inputs such as genomics or clinical variables, feature-attribution methods like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) assign importance scores to each feature for an individual prediction, quantifying how much each factor pushes the output toward one class or another [45, 107]. If a model predicts a favorable response to a drug, SHAP analysis might reveal that high expression of Gene X and a particular mutation are dominant contributors, whereas age or comorbidities are minor [92]. This supports clinical trust—physicians can recognize known predictive factors and identify plausible novel ones—and guides research, as highlighted features can be investigated mechanistically [45]. In drug-sensitivity models, such analyses have pinpointed molecular substructures and genomic alterations that drive predicted efficacy, suggesting specific drug–target interactions for experimental follow-up [108].

A complementary strategy is to design models that are inherently interpretable. Not all tasks require large networks; in some settings, simpler models such as decision trees or rule lists provide transparent decision paths, albeit with possible loss of accuracy [109, 110]. More advanced designs aim to combine expressiveness with interpretability, for example attention-based multimodal architectures or hybrid causal models [23]. An attention-based model for breast cancer therapy response outputs both a prediction and attention weights over modalities (gene expression, imaging, clinical data), indicating which contributed most for a given patient [111]. Such transparency aligns with individualized care. In many clinical scenarios, a modest reduction in predictive accuracy may be acceptable if it yields greater interpretability and higher likelihood of clinician adoption [93, 112].

Interpretability also intersects with regulation and ethics. Regulatory agencies such as the U.S. FDA emphasize understanding algorithm behavior when evaluating AI-enabled medical devices [113]. Transparent models support human oversight, allowing clinicians to audit, contextualize, and, when necessary, override AI outputs [99]. Ultimately, improved explainability is not just a technical goal but a prerequisite for safe integration of AI into oncology practice.

Ethical, privacy, and regulatory challenges in deployment

Bringing AI from the research lab into real-world oncology practice requires navigating stringent ethical, legal, and regulatory constraints, because decisions are life-critical, genomic data are uniquely personal, and clinical standards demand robust evidence before changing practice [99, 114].

Regulatory approval is a major barrier. Agencies such as the FDA classify many AI/ML tools as Software as a Medical Device (SaMD) and require not only high accuracy but also reliability, safety, and generalizability demonstrated in prospective studies [113, 115]. Models that perform impressively on retrospective datasets often degrade in real-time deployment because of distribution shift, for example changes in imaging hardware, protocols, or patient mix [115]. To date, few oncology tools have obtained clearance, mostly in diagnostic support, and successful examples underwent multi-center validation and were released as “locked” algorithms [113]. Regulators typically expect the approved model to remain stable; substantial updates may trigger new review unless a predefined change-control plan exists [113]. This creates tension, because continual learning from new data is attractive yet unchecked evolution could harm patients. Regulatory science is developing frameworks for adaptive algorithms, but practical pathways remain unsettled [116, 117].

Understanding and managing failure modes is another priority. AI systems can fail in unanticipated ways—for example, a vision model misled by artifacts on a slide or a survival model overconfident for patients different from those in the training distribution [118]. Regulators increasingly expect stress testing on out-of-distribution and adversarial-like cases and safeguards such as automatic flagging of low-confidence predictions for human review [117, 118].

Ethical and privacy considerations are equally central. Oncology data frequently include germline and somatic genomic sequences that can reveal identity, family relationships, and disease predisposition [119]. While AI research benefits from multi-institutional cohorts, legal frameworks such as HIPAA and GDPR impose strict constraints [114]. Standard de-identification is insufficient when genomic profiles are quasi-identifiers [114]. Emerging solutions include differential privacy, which constrains training so that individual records cannot be reverse-engineered, and federated learning, which keeps data behind institutional firewalls while sharing only model updates [99]. These approaches allow models to learn from distributed data while reducing privacy risk [99].

Patient consent raises further questions. When biopsies or genomic profiles are used to develop AI tools that may later be commercialized or widely deployed, it is unclear whether current broad research consents are adequate [120]. Patients may need to be explicitly informed if their data will train algorithms that can influence treatment and if downstream uses, including commercialization, are envisioned [121, 122]. Transparent, patient-friendly consent processes will be essential for maintaining trust.

Algorithmic bias is both a technical and moral concern. If an AI system underperforms for certain subgroups—for example, by underestimating risk in a demographic group or misclassifying tumors in underrepresented populations—its deployment could deepen existing disparities [123, 124]. Professional societies and regulators are beginning to recommend bias audits as part of pre-deployment evaluation, requiring performance to be reported across relevant demographic and clinical subgroups [123]. Public reporting can drive model improvement and guide clinicians on where tools should be used with particular caution [123].

Even after regulatory and ethical hurdles are cleared, real-world deployment poses practical challenges. Hospitals must integrate AI tools with electronic health records and imaging systems, ensure adequate compute infrastructure, and maintain cybersecurity [125, 126]. Many advanced models depend on GPUs or cloud services, which may be difficult for smaller centers to support [125]. At the same time, AI systems introduce new attack surfaces, so robust security, logging, and monitoring are integral to responsible deployment [126].

Clinician training and workflow integration are equally critical. A high-performing model that produces long, opaque reports in a separate interface may be ignored if it disrupts clinical workflow [125]. Human-centered design is required so that AI outputs are concise, interpretable, and surfaced at the right moment in existing tools [125]. For example, a recurrence-risk score integrated into the oncologist’s EHR view, accompanied by a simple visual cue and brief explanation, is more likely to be used than a standalone dashboard [23, 125].

Uncertainty quantification (UQ) can further support safe use. If models can signal when they are uncertain or encountering out-of-distribution inputs by flagging low-confidence cases or unusual patterns, clinicians can adjust their reliance accordingly [127, 128]. Pathology AIs that highlight low-confidence slides or rare morphologies encourage experts to scrutinize those cases more carefully, reducing the risk of over-reliance while maintaining algorithmic support in routine scenarios[128].

Resource disparities pose a systemic challenge. State-of-the-art AI development often requires specialized hardware and expertise, raising the risk that only large academic or well-funded centers will benefit [119, 129]. To mitigate this, there is growing emphasis on efficient models that run on commodity hardware and on open-source tools and shared infrastructures [99]. Initiatives such as public imaging repositories and open model hubs allow broad access to datasets and pre-trained models, supporting more equitable adoption [99].

Overall, translating AI into routine oncology care is as much a policy and systems problem as a technical one. Progress depends on aligning algorithms with regulatory standards for safety and effectiveness, adhering to ethical principles of privacy, equity, and transparency, and embedding tools into workflows in a usable and sustainable way. Stakeholders are collaborating on guidelines and reporting standards, which extend clinical trial frameworks to AI interventions and aim to foster rigorous, transparent evaluation and, ultimately, trustworthy AI-enabled cancer care.

Future outlook and conclusion

The landscape of AI in oncology is rapidly evolving, with advances that could substantially change cancer research and patient care. We highlight key emerging paradigms and how they may shape this future, while offering a concluding perspective on the trajectory toward AI-augmented precision oncology.

A prominent trend is the rise of foundation models in biomedicine—large Transformer-based systems pre-trained on vast datasets and then adapted to specific tasks [23, 68]. Analogous to GPT-style models in language processing, oncology is beginning to see multimodal foundation models trained on medical text, images, and molecular data [27]. Models trained on millions of pathology and radiology images can be fine-tuned to classify cancer types with state-of-the-art accuracy using relatively little task-specific data [32]. Other efforts pre-train on clinical notes and oncology literature, enabling “medical AI assistants” that interpret records, summarize evidence, and suggest candidate treatments [111]. Domain-specialized large language models illustrate that such systems can reach expert-level performance on medical exam questions [130]. While text-only models do not diagnose cancer directly, they already support decision-making by synthesizing literature for rare mutations, checking guidelines, or generating structured summaries of complex cases [59].

The next step is tighter integration of language models with dedicated vision, genomics, and structured-data networks into an AI “oncology copilot.” In this vision, a patient’s imaging, pathology, genomics, and clinical history are analyzed by specialized encoders whose outputs are coordinated by a foundation model that also reasons over biomedical knowledge [23, 27]. Early prototypes show AI agents that combine imaging features with automated literature queries to propose individualized treatment options and clinical trial matches [19, 20]. Although still experimental, these systems foreshadow a future in which oncologists work alongside AI partners that continuously ingest new data and publications, offering context-aware, evidence-linked recommendations at the point of care [131].

Generative AI for drug discovery is also advancing. Current diffusion, reinforcement-learning, and other generative frameworks already design molecules with multiple desired attributes; forthcoming models are likely to embed biological context more deeply, for example generating compounds tailored to a tumor’s genomic profile or designed to circumvent resistance mechanisms [72, 73]. Coupled with improved virtual screening and mechanistic simulation, this could move the field toward AI-suggested personalized therapeutics. A speculative scenario would sequence a patient’s tumor, identify key vulnerabilities, and have an AI system design or repurpose a compound for that profile, then simulate its effects on a patient-specific “digital twin” representing tumor and normal-tissue biology [72, 119]. While such personalization remains aspirational, its building blocks are emerging, including organ-on-chip systems, virtual trials that complement real ones, and AI models that assist adaptive trial design by updating dose levels or enrollment criteria in response to interim data [83, 86]. Tools that match patients to niche trials based on complex genomic and clinical eligibility are an early manifestation of this trajectory [19].

Realizing these opportunities will depend less on further gains in predictive accuracy and more on robustness, trustworthiness, and integration into health-care systems. Interdisciplinary collaboration is essential. Cancer biologists, clinicians, data scientists, ethicists, and software engineers must co-design tools that address real clinical needs and constraints [5, 94]. Federated consortia and public–private partnerships are building shared infrastructures for data and models [99]. Attention is also shifting to the post-deployment life cycle of AI. Unlike static diagnostic assays, models may drift as patient populations, practice patterns, and hardware evolve [115]. Continuous performance monitoring, periodic re-validation on new cohorts, and anomaly detection are therefore becoming core elements of “AI stewardship” programs within hospitals [127].

Methodological and reporting standards are another pillar. Guidelines such as CONSORT-AI and SPIRIT-AI provide checklists for designing and reporting clinical trials that involve AI systems, requiring documentation of model versions, workflow integration, update procedures, and how clinicians used algorithmic outputs [113, 132]. Such transparency supports reproducibility, regulatory review, and critical appraisal, and helps ensure that negative or neutral results are also reported. Systematic analysis of these failures is crucial for identifying common pitfalls, such as poor workflow fit, limited generalizability, or lack of clinician engagement [113].

Looking ahead, AI is likely to become a foundational component of precision oncology. It already shows value in extracting clinically meaningful signals from heterogeneous data, identifying high-risk patients, and suggesting new therapeutic hypotheses [23, 27]. The central challenge for the next decade is embedding these capabilities into routine practice in ways that are equitable, transparent, and demonstrably beneficial. This will require validation across diverse populations, mitigation of bias, strong privacy protections, and user-centered interface design so that AI outputs are understandable and actionable at the bedside.

Ultimately, the most compelling vision is one of synergy: oncologists provide nuanced judgment, empathy, and contextual understanding, while AI systems supply scalable pattern recognition, rapid evidence synthesis, and hypothesis generation. Achieving this vision will require sustained collaboration, rigorous science, and vigilant governance, but if realized, AI-augmented oncology could accelerate discovery, enable more precisely tailored therapies, and help reduce the global burden of cancer.

Author contributions

XW and QW conceived and designed the review. SC, XW, DX and BD performed the literature search and drafted the initial manuscript. SC, YH, JH, GD and YT contributed to data extraction, figure and table preparation, and critical interpretation of the literature. QW supervised the project, provided clinical and methodological guidance, and extensively revised the manuscript. All authors contributed to manuscript revision and approved the final version of the manuscript.

Funding

None.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.