A primary clinical use of AI–multi-omics is to predict which patients will benefit from immune checkpoint inhibitors (ICIs) (24, 25). Upfront identification of likely non-responders can spare patients the toxicity and cost of ineffective immunotherapy and allow timely transition to alternative modalities or trials (4). Conversely, confirming likely responders can facilitate rapid access to ICIs and, in selected contexts, support de-escalation from more toxic regimens (1). Importantly, AI models can also suggest mechanism-guided combination strategies by linking individualized tumor states to actionable vulnerabilities (24, 25).

In melanoma, an AI model integrating tumor exome and transcriptome data predicted response to anti–PD-1 therapy with validation performance around AUC ~0.7 (25). Retrospective analyses showed that patients assigned low response probabilities indeed had poorer outcomes (25). Notably, the model suggested that non-responders often retained high T-cell infiltration but exhibited an exhausted phenotype (25). This biologically nuanced stratification has informed combination strategies: a trial design emerged in which patients predicted to respond poorly despite high infiltration received anti–PD-1 plus either a LAG-3 inhibitor (targeting exhaustion biology) or added chemotherapy to immunogenically “reset” the tumor microenvironment (TME) (25, 67, 89). Early signals of improved outcomes in model-selected patients illustrate how AI prediction can be translated into rational combination regimens, rather than merely excluding patients from immunotherapy (25, 67).

In gastroesophageal cancers, where only a subset benefits from ICIs, an AI-based genomic signature has been developed to forecast survival benefit from PD-1 blockade (90, 91). The signature integrates mutation patterns, including co-occurrence such as TP53 and ARID1A alterations, with immune gene expression reflecting TME state (90). Patients designated high-risk by the signature had poor outcomes on immunotherapy alone, but the model also highlighted therapeutic options for combination (90). Specifically, many high-risk tumors displayed MAPK pathway activation, motivating addition of MEK inhibition (e.g., trametinib) to PD-1 blockade; preclinical testing supported synergy in signature-positive models (90–92). The signature further identified a subset with upregulated HSP90 stress proteins, generating a testable hypothesis that HSP90 inhibition could resensitize tumors to ICIs (90, 91). These examples underscore the clinical value of AI not only as a prognostic tool but also as a hypothesis-generating framework that connects multi-omic patterns to tractable combination strategies.

AI-informed combination approaches are also being explored in other cancers by targeting host and metabolic determinants of resistance. In non–small cell lung cancer (NSCLC), microbiome modulation through fecal microbiota transplantation (FMT) or probiotics is under investigation to enhance ICI efficacy (16, 17, 23). An AI model trained on metagenomic and clinical features identified bacterial species whose presence correlated with sustained responses; patients lacking these taxa were predicted to derive suboptimal benefit (16, 17, 23). A pilot trial subsequently tested donor FMT enriched for “favorable” bacteria such as Prevotella copri in predicted low-benefit patients receiving anti–PD-1 therapy (93). Some recipients showed increased T-cell activity and tumor shrinkage, providing early proof-of-concept for AI-guided microbiome adjuvants, while also highlighting the need for cautious donor selection and mechanistic monitoring (93).

Metabolic immunosuppression offers another translational axis. Multi-omics analyses of resistant tumors frequently show signatures of high lactate production and enriched lactylation marks, consistent with metabolic reprogramming that induces suppressive epigenetic states in immune cells (14, 15, 45). These findings have motivated trials combining checkpoint blockade with metabolic agents targeting lactate generation or export, such as lactate dehydrogenase inhibitors or monocarboxylate transporter blockade (14, 15, 45). The rationale is to reduce lactate-driven immune dysfunction, thereby converting immunologically “cold” tumors into more permissive, “hot” states (14, 45). Preclinical data suggest that lactate inhibition can restore dendritic cell function and improve anti–PD-1 efficacy, although definitive clinical results remain pending (14, 45).

Patient-specific multi-omic signatures are also being used to guide which modality to emphasize, not only whether to administer ICIs. In head and neck squamous cell carcinoma (HNSCC), an ML-refined signature (CMPIS) combining immune and stemness indicators stratified tumors into immune-hot versus immune-cold states (94). Low SPI scores reflected inflamed, immunogenic tumors, whereas high scores indicated suppressive, less differentiated TMEs (94). Retrospective analyses suggested that low-SPI patients responded well to ICIs (often with conventional chemotherapy), whereas high-SPI patients derived greater benefit from EGFR-targeted strategies than from immunotherapy (49, 94). Such signatures are proposed to inform first-line treatment allocation, aiming to maximize efficacy by matching tumor state to modality.

In hepatocellular carcinoma (HCC), where ICIs are now integrated into standard therapy (e.g., anti–PD-L1 plus anti-VEGF), AI models combining tumor genomics with blood biomarkers are being explored to identify candidates for conversion therapy, a systemic treatment intended to downstage unresectable disease into resectability (27, 28, 58). An AI radiomics model predicted which patients would achieve substantial downstaging on immunotherapy (27, 28). Predicted responders could be routed to intensified protocols combining ICIs with locoregional interventions such as transarterial chemoembolization, enabling surgical resection in a meaningful fraction; predicted non-responders could be spared ineffective ICI exposure and offered alternative trials (27, 28). This illustrates model-guided escalation and redirection based on individualized response probabilities.

Neoadjuvant immunotherapy represents another clinically relevant setting because response can be quantified at surgery. In resectable NSCLC, only a subset achieves major pathological response (MPR) or pathologic complete response (pCR) after neoadjuvant ICIs (95, 96). Multi-omics models integrating tumor genomics, baseline imaging, and peripheral immune markers have been developed to predict MPR/pCR (58, 59, 74). In a trial of neoadjuvant camrelizumab plus chemotherapy, biomarker analyses reportedly identified factors associated with a large proportion of pCR cases (95). If prospectively validated, these tools could triage patients toward neoadjuvant immunotherapy versus alternative preoperative strategies, potentially reducing overtreatment and improving surgical outcomes.

Alongside systemic decision-making, AI is being embedded into routine diagnostic pipelines to support immunotherapy decisions and monitoring. Digital pathology platforms use deep learning to quantify PD-L1 expression and tumor-infiltrating lymphocyte (TIL) densities, improving reproducibility over manual reads and enabling spatial descriptors such as TIL clustering at invasive margins (82, 97, 98). These metrics can complement molecular profiling by providing rapid, standardized immune-context measurements from routine biopsies.

Radiology is similarly adopting AI-derived imaging biomarkers. Radiomic texture and heterogeneity measures from CT, MRI, and PET can capture tumor architecture and stromal features associated with immune infiltration or exclusion (21, 59). AI analysis of baseline CT scans in lung cancer has been reported to predict response and survival after ICIs independently of clinical factors, using features such as enhancement heterogeneity, tumor shape, and tumor–tissue interface characteristics that correlate with immune stroma (99–101). In practice, pathology-derived AI immune metrics and radiomic scores could be jointly considered to guide immunotherapy versus alternative options.

Ultrasound, including contrast-enhanced ultrasound, is also being explored for serial assessment. Ultrasound imaging may quantify therapy-induced changes in vascularity and perfusion that accompany immune infiltration, potentially enabling on-treatment evaluation even when size criteria lag (102, 103). Although still experimental, such approaches could help detect immune-related inflammation patterns, distinguish true progression from pseudoprogression, and prompt timely adjustments to treatment.

Robust generalization across institutions, platforms, and patient demographics is essential for clinical use. Multicenter validation and real-world evidence are therefore central, and several examples illustrate both promise and remaining risks. An international validation of a deep-learning radiomic model for NSCLC, trained to predict anti–PD-1 response from pretreatment CT scans combined with blood inflammatory markers, maintained AUCs of ~0.82–0.86 across multiple hospitals in Europe and China (104, 105). Model outputs correlated with overall survival and are being evaluated prospectively, supporting the feasibility of cross-region transport when predictors capture durable biological correlates (104, 105).

In diffuse large B-cell lymphoma, an artificial neural network integrating mutation status, PD-L1 expression, and clinical factors achieved ~0.94 AUC in initial cohorts and retained similar performance in independent multicenter sets (61, 66). Importantly, predicted response scores correlated with spatial immune profiling: responders showed higher densities of PD-1⁺ T cells proximal to PD-L1⁺ macrophages, consistent with active immune engagement, whereas non-responders exhibited diffuse suppression (61). Such convergence between AI predictions and spatial biology strengthens mechanistic plausibility and external validity.

In gastrointestinal cancers, multi-omics signatures derived in one cohort have been tested in independent datasets and public repositories such as TCGA (106). A prognostic immunogenomic signature for microsatellite-stable colorectal cancer validated in TCGA outperformed tumor mutational burden and PD-L1 status in predicting survival on immunotherapy (91, 107). Combined DNA/RNA classifiers for gastric cancer also retained high sensitivity and specificity in independent Asian cohorts, underscoring cross-population validity amid regional genetic and environmental variation (108, 109).

Retrospective analyses further support reproducibility of some AI-derived signals. In melanoma, an “immune responsiveness” score applied to pooled transcriptomic datasets stratified benefit across studies and recapitulated expected biology: higher interferon-γ and CXCL9 expression among predicted responders and enrichment of MYC targets and WNT/β-catenin signaling among predicted non-responders (110, 111). In lung adenocarcinoma, a stemness-based classifier separating immune-hot versus immune-cold tumors was validated across seven cohorts totaling over a thousand patients; pathology review confirmed paucity of CD8⁺ T cells in high-stemness tumors by immunofluorescence (49, 112).

Prospective practice data are emerging as well. In bladder cancer, a model integrating urine tumor DNA mutations and methylation markers has been prospectively tested across centers, reducing invasive cystoscopy by flagging patients likely stable on immunotherapy versus those likely progressing (113, 114). Meta-analytic work suggests that single biomarkers often have lower generalizability in diverse populations, reinforcing the need for multi-omic composites that better capture heterogeneity across demographics and environments (115, 116). Analyses similarly show that trials incorporating AI or multi-omics immunotherapy biomarkers have increased substantially since 2020, with many specifying validation and biomarker-guided therapy endpoints (117, 118). Nonetheless, even strong models can drift with new data distributions (sequencing platforms, preprocessing pipelines, or imaging protocols), and early deployment studies indicate that performance may drop until harmonization and recalibration are applied (119).

Despite encouraging evidence, broad adoption faces major practical and ethical barriers. Data heterogeneity and standardization remain central: different centers generate omics and imaging data with distinct platforms, protocols, and recording formats, creating batch effects that compromise portability (68, 69). Harmonized pipelines and quality standards are required, and model recalibration may be necessary when deployed across scanners or sequencing workflows.

Privacy and governance constraints further limit access to the diverse cohorts needed for robust training (120). Federated learning is emerging as a solution by enabling collaborative training without moving patient-level data (121, 122). Federated efforts, such as those combining data from multiple hospitals, have achieved performance comparable to pooled-data training, illustrating feasibility while addressing privacy and heterogeneity (123, 124). Still, federated approaches require consistent annotation, secure aggregation, and rigorous governance to ensure data quality and accountability.

Prospective validation remains limited: most predictors are retrospective, and clinical translation requires evidence that acting on predictions improves outcomes (6, 125). Prospective biomarker-guided trials are underway in settings such as neoadjuvant lung cancer, but larger randomized studies are needed before routine adoption. Regulatory approval also demands demonstration of analytical validity, clinical validity, and clinical utility, as well as cybersecurity and post-market surveillance; adaptive models raise additional issues of versioning, re-approval, and drift monitoring (126, 127).

Interpretability and clinician acceptance are pivotal, as many deep learning models are perceived as black boxes. Explainable AI methods (e.g., SHAP and Grad-CAM) can clarify drivers of predictions, and visualization of multi-omic correlations can make model reasoning more tangible (128, 129). Another translation strategy is distillation into parsimonious, assayable signatures, for example, reducing complex models to small gene panels or immunohistochemistry/PCR surrogates compatible with standard workflows (58).

Equity considerations are paramount. Models trained on predominantly European-ancestry or region-limited cohorts may underperform in underrepresented groups, potentially worsening disparities (130, 131). Mitigation requires diverse cohort inclusion, fairness-aware training and calibration, subgroup-stratified reporting, and ongoing surveillance for systematic underperformance. Workflow integration is also non-trivial: outputs must be presented in clinician-friendly interfaces (e.g., dashboards showing risk spectra, confidence intervals, and key explanatory factors) and embedded within electronic records to avoid additional cognitive burden (132–134).

Finally, cost-effectiveness will shape uptake. Multi-omics profiling and advanced analytics can be expensive and unevenly available, and reimbursement will require evidence of value. Early health-economic analyses suggest that if an AI test can spare even ~20% of patients from futile therapy, it may be cost-saving, but real-world confirmation is needed (135, 136). Looking forward, translational roadmaps emphasize harmonized data handling, federated learning, explainable modeling, prospective trials, and expansion to adjunct strategies such as microbiome and metabolic modulation, with emerging concepts including wearable AI monitoring for toxicity and exploratory computational approaches (e.g., quantum computing) for complex simulations (Figure 3).

In summary, AI–multi-omics tools are increasingly positioned to personalize ICI use by predicting benefit, guiding combinations, enabling neoadjuvant triage, and supporting monitoring via pathology and imaging (Figure 4). Achieving routine clinical adoption will depend on multicenter validation, prospective evidence of utility, standardization and privacy-preserving infrastructure, interpretable and equitable model development, seamless integration into workflows, and demonstrable clinical and economic value.

A major frontier is integrating single-cell and spatial omics data with AI. Traditional bulk sequencing averages signals across millions of cells, potentially obscuring critical intratumoral and microenvironmental heterogeneity. Single-cell RNA-seq, ATAC-seq, and multimodal single-cell assays that jointly measure gene expression and cell-surface proteins can delineate the multitude of cell types and functional states within the tumor and tumor microenvironment (TME) (86, 137). Applying AI to these high-dimensional datasets can identify subtle subpopulations or gene programs associated with immune escape. In NSCLC, for example, a multi-omic single-cell study confirmed that a rare SOX2⁺ progenitor population drives lineage plasticity and ICI resistance (10). Researchers are therefore developing deep learning models that predict therapy outcome from the cellular ecosystem defined by single-cell data, learning which cellular composition patterns, such as regulatory-to-cytotoxic T-cell balance or the presence of specific myeloid suppressor subsets, track with response.

The scale of single-cell and spatial data, often thousands of cells or spatial spots per patient, demands architectures that can model relationships among cells. Graph neural networks represent cells as nodes in an interaction graph, capturing neighborhood structure and cell–cell communication (138, 139). Transformer models, which excel at long-context representation, can treat cells (or spatial spots) as elements in a sequence and infer higher-order dependencies across the TME (140). These approaches enable models to learn patterns such as adjacency of PD-L1–expressing tumor cells to exhausted T cells, or spatial clustering of suppressive myeloid states, to inform response prediction (138, 140). A forward-looking vision is a clinically integrated system that ingests a spatially resolved tumor map and outputs both a response probability and an interpretable annotation highlighting immunosuppressive niches and immune-active regions, supporting pathologists and oncologists in near real time.

Federated and distributed learning paradigms are also advancing. By enabling institutions to train models locally and share only parameter updates, federated learning can facilitate multi-center model development without pooling raw data (121, 141). These frameworks are becoming more sophisticated through secure multi-party computation and, in some proposals, blockchain-based audit trails for transparency and traceability. In immunotherapy, federated networks could allow large cancer centers to contribute to a continuously learning predictor that improves as outcomes accrue across sites (122). Supporting feasibility, the Oncology Federated Learning Network reported training a metastatic cancer survival prediction model across 17 institutions without pooling data, achieving performance within 2% of a centralized model (121, 122, 124). Extending this paradigm to ICI response could accelerate improvement and improve generalizability across populations. Parallel efforts in privacy-preserving AI, such as homomorphic encryption, which enables computation on encrypted data, may further reduce barriers to incorporating electronic health records and radiology archives into training and validation (122, 123).

Generative AI represents another expanding toolkit. GANs and diffusion models can synthesize realistic multi-omics profiles that follow statistical patterns of real cohorts, augmenting training sets when data are scarce (142, 143). This is especially relevant for rare cancers and underrepresented patient groups, where limited sample size increases risk of overfitting and biased performance. Generative models can also support mechanistic exploration by simulating multi-omics “what-if” scenarios: given a tumor’s genomic context, they can generate plausible transcriptomic and immune-cell profiles consistent with either response or resistance, suggesting molecular shifts that might convert non-responders to responders (143). In parallel, foundation-model strategies are emerging, including large language–model–like approaches for biomedical data: models pre-trained on large corpora of tumor transcriptomes or clinical narratives may be fine-tuned with limited task-specific data to predict immunotherapy outcomes more robustly (144, 145).

On the computational horizon, quantum computing is being explored for its potential to handle the combinatorial complexity inherent to multi-omics. Quantum algorithms could, in principle, accelerate feature selection or simulate molecular interactions within the TME beyond classical brute-force approaches (146, 147). While practical quantum computing for clinical medicine remains years away, quantum-inspired algorithms for optimization are already being investigated, aiming to improve parameter search and enable exploration of extremely high-dimensional interaction spaces that are otherwise computationally prohibitive.

Emerging cross-disciplinary fields further broaden the scope of prediction. Cardio-oncology exemplifies the move from efficacy-only prediction toward joint optimization for benefit and harm: AI is being used to anticipate immune-related cardiac adverse events, including ICI-associated myocarditis, by integrating cardiac biomarkers and imaging with oncologic variables (148, 149). The American Heart Association has issued statements emphasizing opportunities for AI and multi-omics to forecast long-term immunotherapy risks, including cardiotoxicity (150). Such work suggests future predictors may concurrently estimate likelihood of tumor control and probability of severe adverse events, tailoring therapy to maximize benefit while minimizing harm.

Clinical translation will require rigorous validation and refinement toward truly personalized immunotherapy, with emphasis on actionable biomarkers that drive decisions rather than merely correlating with outcomes. Accordingly, prospective trials increasingly incorporate multi-omics and AI analyses as exploratory, and sometimes primary, endpoints, creating the evidentiary pathway for clinical deployment.

Liquid biopsy is a leading avenue for personalization because it can be obtained serially, enabling longitudinal monitoring of tumor evolution and immune engagement. AI models integrating time-resolved circulating tumor DNA (ctDNA), circulating immune-cell features, and related blood-based analytes are being tested to predict outcomes and detect resistance early (29, 35, 113). In bladder cancer, a liquid-biopsy methylation signature has been used to guide therapy adaptation: patients whose blood shows a “cold tumor” methylation pattern after several weeks on ICIs may be escalated to combination therapy, whereas those with a “hot tumor” pattern may continue ICIs alone (30, 113). In melanoma, early declines in ctDNA during therapy correlate with response; models combining ctDNA dynamics with radiographic changes and on-treatment transcriptomic shifts can improve prediction beyond any single modality (25, 151). Such adaptive, biomarker-driven strategies embody the personalized medicine ideal by adjusting treatment in near real time based on each patient’s evolving molecular data.

Translation to routine care often demands simplification. Whole-exome sequencing and RNA-seq remain variably available and may not meet turnaround constraints in all clinics. A practical strategy is to develop targeted panels that capture the most informative features discovered by AI, for example, a defined set of expression markers, a curated mutation list, and selected microbiome taxa, so that a focused assay can approximate the predictive signal of broader profiling (29, 152). AI can also prioritize biomarkers likely to remain robust when implemented with simpler measurement technologies such as PCR or immunohistochemistry (113). In NPC, where EBV-driven enhancers activate CACNA2D1, an enhancer activation score based on a small set of epigenetic marks has been proposed as a biomarker that could be measured by a targeted assay and used to select patients for enhancer-blocking epigenetic strategies alongside ICIs (12, 13).

Prospective trials that explicitly test AI-guided treatment adaptation represent a critical next step. “Adaptive immunotherapy” designs envision models that continuously analyze patient data during therapy and recommend modifications, such as adding an agent, switching regimens, or adjusting intensity, when predicted response probability drops below a threshold (28, 113). Operationalizing such trials requires rapid turnaround of complex assays, real-time computation, and an integrated clinical decision engine, but successful execution could shift immunotherapy management from static, one-size-fits-all regimens to dynamic, data-informed control policies (153, 154).

Wearable and digital biomarkers are likely to expand monitoring further. Physiologic signals from wearables (activity, sleep, heart-rate variability) may provide early indicators of toxicity and possibly efficacy; for example, fatigue patterns could reflect cytokine-mediated immune activation (155, 156). AI can integrate these continuous streams with molecular and imaging features to produce a holistic patient-status estimate and alert care teams to impending immune-related adverse events for preemptive intervention (75). Proof-of-concept work has predicted severe immunotherapy toxicity by combining baseline cytokine profiles with T-cell receptor repertoire features, suggesting that periodic molecular monitoring may be complemented by continuous digital signals (157). Looking ahead, generative AI may also support “digital twin” concepts, simulating likely trajectories under alternative regimens to aid selection of safer and more effective options for individual patients.

Interdisciplinary integration remains essential. Microbiome–immune axis modeling links microbial metabolites to immune-cell states in the TME, suggesting microbiome modulation strategies to overcome resistance (158). Integration with metabolic and epigenetic research is similarly influential; discoveries such as lactylation emerged from cross-talk between metabolism and immunology, and AI is helping map where metabolic interventions might integrate with immunotherapy regimens (14, 15). Systems biology and network-science approaches, sometimes framed as immune-atlas initiatives, may further enable simulation of how perturbations (e.g., blocking TGF-β) propagate through tumor–immune dynamics, with AI serving as the computational engine for such high-dimensional simulations (159, 160).

Global research efforts are coalescing around these directions. Initiatives such as the National Cancer Institute’s Human Tumor Atlas Network and international consortia on immunotherapy biomarkers are generating rich datasets that AI can leverage. Concurrently, a growing emphasis on open data and open models is enabling external validation and community scrutiny, reducing the risk that overfitting or bias goes unnoticed. In conclusion, the coming years are poised to deliver deeper mechanistic insight and more powerful predictive tools for immunotherapy as AI advances alongside experimental technologies.