AI models fuse multi-omics data to predict immunotherapy response in cancer

Immune checkpoint inhibitors have transformed the treatment of many cancers, but durable benefit still reaches only a subset of patients. A narrative review published in Cancer Cell International argues that the field's standard biomarkers, including PD-L1 expression and tumor mutational burden, provide only partial predictive value, and that artificial intelligence applied to multi-omics data could offer a more complete predictive framework for identifying which patients will benefit from these drugs. The review, led by Junjie Wang and colleagues, explains why single-marker predictions fall short.

Response to checkpoint inhibitors emerges from the interplay of tumor antigenicity, antigen presentation, immune infiltration, the spatial organization of the tumor microenvironment, and host systemic factors. Because no single measurement reflects all of these dimensions, the authors make the case for integrating genomic, transcriptomic, epigenomic, proteomic, metabolomic, microbiome, imaging, and longitudinal liquid-biopsy data into unified models of treatment benefit and resistance. Artificial intelligence methods are increasingly used to fuse these heterogeneous data types and to model nonlinear interactions relevant to treatment benefit and resistance.

The review organizes the major integration paradigms into several families: early, late, and intermediate fusion strategies, along with hierarchical, graph-based, and latent-factor approaches. Each paradigm carries distinct strengths, limitations, and potential clinical use cases, and the authors examine them in turn rather than endorsing a single method. A substantial portion of the review is devoted to recurring methodological pitfalls that have affected AI biomarker studies.

The authors highlight endpoint heterogeneity across studies, missing data modalities, information leakage between training and test sets, domain shift when models are deployed on data from different centers, inadequate external validation, and poor calibration of predicted probabilities. The review presents these issues as obstacles to reliable clinical translation, and the authors argue that attention to them is necessary for predictive models to become clinically usable. To address these weaknesses, the review proposes practical standards for model evaluation and reporting intended to support clinical translation.

The authors emphasize that progress in the field will depend not only on increasingly sophisticated algorithms but also on rigorous study design, harmonized outcome definitions, and multicenter validation. As a narrative review, the article synthesizes existing literature rather than presenting new clinical trial data, and its recommendations are framed as steps toward more reliable and clinically usable predictive models for immunotherapy.

Subject of Research: Artificial intelligence for predicting immune checkpoint inhibitor outcomes through multi-omics integration

Article Title: Artificial intelligence for predicting immune checkpoint inhibitor outcomes through multi-omics integration

Article References: Wang, J., Wang, X., Wang, Q., Xiong, D., Cui, S., Duan, B., Ding, G., Huang, Y., & Feng, Y. (2026). Artificial intelligence for predicting immune checkpoint inhibitor outcomes through multi-omics integration. Cancer Cell International. https://doi.org/10.1186/s12935-026-04490-1

Image Credits: AI Generated

DOI: 10.1186/s12935-026-04490-1

Keywords: immunotherapy, immune checkpoint inhibitors, artificial intelligence, multi-omics, biomarkers, cancer research, machine learning, tumor microenvironment, predictive modeling, Cancer Cell International, Artificial, intelligence