AI models predict cancer immunotherapy response better than PD-L1, but few are validated in patients
A narrative review finds multimodal AI beats single biomarkers in retrospective work, yet prospective validation is scarce and performance often drops externally.
Summary
Immunotherapy helps some cancer patients for years while others get little benefit, and standard tests such as PD-L1, tumor mutational burden and microsatellite instability predict the difference only partly. This narrative review surveys artificial intelligence models that combine genetic data, CT/MRI/PET imaging, digital pathology slides, blood markers and clinical details into composite predictors. Several retrospective studies report AUCs of 0.8 or higher, and AI is also being applied to CAR-T patient selection and toxicity prediction. The authors stress that prospective, multicenter validation is rare, external testing often shows worse performance, and many models are opaque. For now these tools are research aids, not proven decision tools. Key hurdles are data standardization, bias, transparency and regulation.
Detailed Summary
Immune checkpoint inhibitors and cell therapies have changed oncology, yet clinicians still cannot reliably tell who will respond. This narrative review, drawing on PubMed, Google Scholar and trial registries through early 2025, asks whether artificial intelligence can close that gap by predicting response and personalizing treatment across solid and blood cancers.
The authors first describe why current markers fall short. PD-L1 and tumor mutational burden vary across cancer types, assays and cutoffs differ between laboratories, and a single biopsy cannot capture heterogeneity within a tumor or changes over time. Microsatellite instability predicts well but is found in only about 5% of tumors. None of these markers reliably predicts immune-related toxicity. The tumor microenvironment, with its suppressive immune cells, dense matrix and abnormal blood vessels, adds further complexity.
The core argument is that combining data types outperforms any single marker. The review covers genomics and transcriptomics (including neural networks that predict neoantigen binding), radiomics from CT, MRI and PET, digital pathology on whole-slide images, spatial and single-cell omics, and circulating tumor DNA. Reported performance includes an AUC of about 0.80 for the multi-input deepAFM model, 0.83 for predicting PD-L1 status from NSCLC imaging, about 0.92 for PD-L1 expression from breast cancer pathology, and 0.838 for NSCLC spatial multi-omics. A gastric cancer radiopathomics signature reached AUCs of 0.978 to 0.822 across validation cohorts and was linked to memory B-cell infiltration.
Tumor-specific applications include neoantigen discovery for vaccines in melanoma, machine learning on real-world NSCLC data using neutrophil-to-lymphocyte ratio, performance status and PD-L1, and an imaging tool from Onc.AI that received FDA breakthrough device designation for risk stratification. AI is also being used to select CAR-T patients, optimize manufacturing and anticipate cytokine release syndrome and neurotoxicity. Emerging approaches include federated learning, foundation models, digital twins and wearable monitoring.
The evidence level matters. These are mostly retrospective, single-center model-development results, and the authors state that external validation often shows deterioration. Gastrointestinal models are often not externally validated, and datasets are heterogeneous. Data standardization, bias, limited reproducibility, poor longitudinal validation and a lack of explainability remain barriers to clinician trust and regulatory acceptance.
The implication is that multimodal AI is a promising mechanism-level and prediction advance, but it has not been shown to improve patient outcomes. The authors call for prospective multicenter trials, transparent reporting, explainable models and stronger regulatory frameworks before AI moves from a retrospective research tool to routine clinical decision support. Because it is a narrative review without systematic quality appraisal, the cited performance figures should be read as illustrative rather than pooled estimates.
Key Findings
- Multimodal deep learning (deepAFM) combining histopathology, genomics, radiomics and clinical data reported AUC up to about 0.80 for immunotherapy response (retrospective).
- Gastric cancer radiopathomics signature (H&E slides plus CT, seven ML models) reached AUCs of 0.978 to 0.822 across validation cohorts and tracked memory B-cell infiltration.
- Imaging-based AI predicted PD-L1 status in NSCLC with AUC 0.83; pathology AI predicted PD-L1 expression in breast cancer with AUC about 0.92.
- NSCLC spatial multi-omics model reached AUC 0.838, but the review notes external validation often shows performance deterioration.
- Conventional markers are limited: MSI-H is a strong pan-cancer predictor but prevalence is only about 5%, while PD-L1 and TMB lack standardized cutoffs and are partial predictors.
- Onc.AI's Serial CTRS imaging tool received FDA breakthrough device designation for NSCLC risk stratification, which is a designation, not proof of outcome benefit.
- Prospective, multicenter validation is scarce, and many GI cancer models lack external validation; no outcome-improvement trial results are reported.
Methodology
This is a narrative review, not a systematic review or meta-analysis. Literature was searched in PubMed, Google Scholar and trial registries through early 2025 using AI, immunotherapy, biomarker, tumor microenvironment, CAR-T and radiomics terms. Prospective or externally validated studies and landmark reviews were prioritized, but no formal quality assessment, risk-of-bias scoring or pooled analysis is described. Performance figures (mostly AUCs) are taken from individual cited studies.
Study Limitations
The authors note that most studies are retrospective, validation is rarely prospective or multicenter, and performance often declines externally. Data heterogeneity, bias, limited reproducibility and model opacity persist. As a narrative review it may be subject to selection bias, and no conflicts of interest were reported in the supplied text, which is truncated before the breast cancer and CAR-T sections.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe:
