From Empirical Choices to Model-informed Therapy

Conventional drug selection in oncology relies primarily on tumor type, disease stage, and a limited set of established biomarkers, such as EGFR mutations, HER2 amplification, and PD-L1 expression. While effective in defined contexts, these approaches do not fully account for the molecular heterogeneity of tumors or the complex interactions between molecular features and therapeutic outcomes. AI-based models offer a complementary approach by identifying nonlinear relationships between large-scale molecular profiles and drug response, drawing on both preclinical data sources including cell lines, patient-derived xenografts, and organoids and clinical cohorts ¹⁷.

Recent reviews have demonstrated that DL architectures, including convolutional neural networks (CNNs), graph neural networks (GNNs), and transformers, can predict drug sensitivity by combining molecular profiles with chemical descriptors or drug–target interaction networks. In several benchmarking studies, these methods have shown performance advantages over traditional ML approaches such as random forests and support vector machines, though the extent of improvement is context-dependent and not universal ^18,19,18,19.

Multi-omics Integration and Modern Frameworks

Three particularly notable developments characterize recent progress in this area. First, counterfactual treatment frameworks have been proposed as a means of generating personalized drug recommendations from multi-omics data. In one such framework separate ML models referred to as experts are trained on distinct omics modalities, including single-cell transcriptomics and proteomics, and subsequently combined into an ensemble that produces probabilistic treatment recommendations with calibrated confidence scores and feature-level explanations. Applied to ovarian cancer, this approach leverages retrospective in vitro and in vivo response data to estimate how individual patients might respond to alternative agents, offering a principled method for addressing treatment-choice bias in observational datasets ¹².

Second, flexible deep learning toolkits have been developed to facilitate multi-omics analysis in practical clinical settings. Flexynesis, a deep learning framework designed for bulk multi-omics data, supports joint prediction of drug response, cancer subtype classification, and patient survival through a multi-task learning architecture that accommodates missing data modalities and variable sample sizes limitations that are common in real-world clinical datasets ²⁰.

Third, advances in integration methodology have strengthened the analytical foundations of multi-omics modeling. Technical reviews highlight supervised frameworks such as DIABLO and sparse generalized canonical correlation analysis (sGCCA), which maximize shared variance across omics layers while improving prediction accuracy for clinical phenotypic outcomes. These statistical approaches are increasingly combined with DL architectures to enhance the stability and interpretability of resulting models ^9,21,9,21.

Performance and Clinical Relevance

Taken together, recent findings in this area point to several important themes. Multi-omics DL models have demonstrated improved predictive accuracy often higher AUCs or lower prediction errors compared to single-omics or clinical-data-only models, though performance varies across datasets and cancer types ¹⁰. Cross-domain generalization remains a recognized limitation: while some frameworks have demonstrated knowledge transfer from preclinical models to patient cohorts, performance in real-world clinical settings tends to be lower than in controlled benchmarks ¹⁸. To address interpretability concerns, feature attribution methods such as SHAP values and integrated gradients are increasingly applied to identify genes, pathways, and drug–target interactions that drive model predictions, thereby linking drug response models directly to biomarker discovery efforts ¹³.

Current Limitations & Challenges

Several limitations constrain the clinical translation of AI-based drug response models. Data quality and harmonization remain significant concerns, as multi-omics datasets generated across different platforms, laboratories, and protocols often exhibit batch effects and platform-specific biases that complicate integration. Most high-performing DL models function as black boxes, offering limited transparency into their decision-making processes, which undermines clinician trust and poses barriers to regulatory acceptance. Furthermore, while many models have shown strong performance on retrospective or preclinical datasets, prospective clinical validation and multi-center trials are still largely absent from the literature. Regulatory, ethical, and data-sharing considerations further complicate real-world implementation.

Radiotherapy Planning, Automation and Outcome Prediction

Radiation therapy (RT) is among the most widely used modalities in cancer treatment; however, the associated planning process is resource-intensive. Delineation of tumor volumes and surrounding organs at risk, dose computation, and plan optimization can collectively require several hours per patient ²². AI is being applied at multiple stages of this workflow to reduce planning burden and improve consistency.

In auto-contouring, CNN-based segmentation models have been developed to delineate tumors and organs at risk on CT and MRI images, substantially reducing the time required for manual contouring and mitigating inter-observer variability ²³. Dose prediction and plan optimization: DL methods predict 3D dose distributions based on the anatomy and prescription, which can be used as an initial solution or benchmark for planners ²⁴. At the workflow level, organizations such as ESTRO have begun systematically categorizing RT automation types, ranging from decision-support to full automation, and assessing their readiness for clinical implementation ²⁵. While these technologies show considerable promise in reducing inter-observer variability and shortening planning timelines, prospective clinical studies are still limited, and their generalizability across institutions requires further validation.

AI for Focal Vs. Radical Therapy Decisions: The Case of Unfold AI

Beyond radiotherapy, AI is also informing decisions regarding the intensity of local treatment in prostate cancer. Unfold AI (Avenda Health), developed in partnership with UCLA, is a proprietary DL-based tool that processes MRI and clinical data to generate detailed estimates of prostate tumor volume and spatial extent ²⁶. Published evaluations report that Unfold AI determines cancer margins with approximately 84% accuracy, compared to approximately 67% for clinicians using standard imaging assessments, though these figures derive from specific study contexts and should be interpreted accordingly ²⁷. By enabling more precise mapping of tumor extent, the tool supports the identification of patients who may be appropriate candidates for focal therapy rather than radical prostatectomy or whole-gland radiotherapy, with the potential to reduce treatment-related morbidity while maintaining adequate tumor control ²⁸. Table 1 summarizes selected AI applications in cancer treatment planning, illustrating the breadth of clinical roles and reported benefits across different modalities.

In Table 1 revealed that artificial intelligence is increasingly integrated into radiation therapy workflows, improving efficiency, accuracy, and clinical decision-making across multiple stages of treatment planning. In auto-contouring applications, CNN-based segmentation models significantly reduced contouring time and minimized inter-observer variability, improving consistency in tumor and organ-at-risk delineation. Similarly, deep learning–based dose prediction systems using 3D CNNs and U-Net architectures enabled rapid generation of treatment plans with dosimetric quality comparable to or better than manually created plans, while substantially accelerating the planning process. AI readiness assessments further highlighted the importance of workflow integration, providing guidance on acceptable levels of automation and supporting clinical adoption strategies. 

In prostate cancer management, the Unfold AI system demonstrated strong performance by improving tumor detection accuracy (84% vs. 67% compared to clinicians) and enabling more precise treatment stratification, with potential to significantly reduce treatment failure rates in modeled scenarios. Overall, the findings illustrate that AI enhances radiation therapy by improving planning efficiency, standardization, and diagnostic precision, while also supporting safer and more optimized treatment delivery.

Deep Learning in Genomics and Multi-omics

AI is increasingly being applied to the discovery of biomarkers associated with treatment response, adverse effects, and patient survival. DL approaches applied to genomic data have demonstrated capacity to identify key mutations, mutational signatures, and gene expression changes associated with prognosis or drug sensitivity, complementing traditional biomarker identification methods ¹⁹. Multi-omics integration combining genomics, transcriptomics, proteomics, and radiomics has shown the potential to produce more robust biomarker signatures by capturing complementary aspects of tumor biology that no single data type can fully represent ¹⁰.

Several methodological advances have supported progress in this area. GNNs have demonstrated utility in modeling gene–gene and protein–protein interaction networks, capturing relational structure that conventional methods cannot easily encode ¹³. Transformers and attention mechanisms have been applied to the joint analysis of omics layers and imaging features, enabling the identification of cross-modal associations ²⁹. Explainable AI (XAI) techniques have also been incorporated into biomarker discovery pipelines, connecting model predictions to interpretable sets of biological features and thereby supporting both the biological validation and clinical communication of candidate signatures ³⁰.

Multi-omics-derived Biomarkers in Precision Oncology

Comprehensive assessments of AI-enabled precision medicine suggest that multi-omics-based models tend to outperform single-omics biomarkers or conventional clinical risk scores in predicting survival and informing treatment selection, though performance gains vary by cancer type, data quality, and modeling approach ¹³. Illustrative examples include integrative signatures that combine mutational data, gene expression profiles, and DNA methylation patterns to predict responses to immunotherapy or targeted agents ¹⁰. As well as multi-task models such as Flexynesis that simultaneously predict cancer subtype, drug response, and survival, enabling the development of unified biomarker panels applicable across clinical objectives ²⁰. Table 2 and Figure 2 summarize key AI-enabled biomarker discovery approaches and the stages of a generalized biomarker discovery pipeline, respectively.

In Table 2 revealed that artificial intelligence–driven multi-omics integration is significantly enhancing biomarker discovery and precision oncology applications by combining diverse biological data sources to improve predictive performance. Deep learning–based multi-omics models integrating genomics, transcriptomics, and proteomics demonstrated improved risk stratification compared to traditional approaches, highlighting their ability to capture complex disease signatures more effectively. Similarly, machine learning and network-based approaches combining genomic and clinical data enabled the identification of predictive biomarkers and drug-repurposing opportunities, supporting more personalized treatment selection. Systematic reviews of precision oncology applications further illustrated that AI methods applied to multi-omics and imaging data generate a wide range of candidate biomarkers across cancer types, although they emphasize the need for standardization and prospective validation before clinical translation. In addition, methodological frameworks such as DIABLO and sGCCA provide robust integrative strategies by identifying co-varying feature modules associated with clinical outcomes, strengthening the reliability of multi-omics biomarker panels. Overall, the findings demonstrate that AI-powered integration of multi-omics data holds strong translational potential for improving prognostic accuracy, guiding therapy selection, and advancing personalized oncology, while highlighting the importance of methodological standardization for clinical adoption.

Limitations and Challenges

Biomarker discovery using AI remains subject to several important limitations. Many AI-derived biomarkers have not progressed beyond the discovery stage, as prospective validation in independent cohorts is often absent. Data heterogeneity and batch effects arising from differences in assay platforms, sample handling procedures, and population demographics remain persistent sources of variability that can compromise reproducibility. Complex multi-omics signatures may also be difficult for clinicians to interpret or act upon without clear biological grounding, limiting their practical utility. Finally, the use of multi-omics data raises ethical and regulatory concerns related to patient consent, data sharing, and cross-institutional harmonization that must be addressed before clinical adoption can be realized at scale.

Data Availability Statement

No new data were generated or analyzed in this review article.

Funding

The author declares that no funding was received for this work.

Ethics Approval and Consent to Participate

Not applicable.

Conflict of Interest

The author declares no conflict of interest.

Use of Generative AI

Generative AI tools were used only for language improvement and proofreading. The authors reviewed and approved the final content and take full responsibility for the manuscript.