Deep Learning Pathomics Platform May Predict Immunotherapy Response in Lung Cancer

A deep learning-based pathomics platform accurately predicted outcomes and response to immunotherapy in patients with metastatic non-small cell lung cancer (NSCLC), according to a study presented at the American Association for Cancer Research (AACR) Annual Meeting 2026. The biology-guided artificial intelligence model, called Pathology-driven Immunotherapy Optimization (Path-IO), was applied to routine pathology slides and outperformed PD-L1, the current standard-of-care biomarker.

The researchers developed Path-IO as a deep learning survival prediction model that analyzes pathology slide images and identifies specific features, or niches, within the tumor microenvironment. By combining these tissue patterns with imaging and clinical data, the model estimates whether a patient is at higher or lower risk of poor outcomes from immunotherapy. The study included 797 immune checkpoint inhibitor-treated NSCLC patients from UT MD Anderson Cancer Center, with external validation in 280 additional patients from Mayo Clinic, Gustave Roussy, and the phase III Lung-MAP S1400I trial. In the UT MD Anderson cohort, patients in the high-risk group had more than double the risk of death or disease progression compared with those in the low-risk group, and comparable stratification was observed in the validation datasets.

Performance was measured using the concordance index (C-index). Path-IO achieved C-indices of 0.69 for overall survival (OS) and 0.65 for progression-free survival (PFS) in the discovery cohort, and 0.63 for OS and 0.58 for PFS in the test cohort. By contrast, PD-L1 had C-indices of 0.58 (OS) and 0.57 (PFS) in the discovery cohort, declining to 0.50 and 0.51 in the test cohort. Combining pathology-based predictions with radiomics and clinical data further improved the C-index, from 0.58 to 0.70 for PFS and from 0.63 to 0.75 for OS.

Unlike prior pathomics studies, Path-IO is a biology-guided approach that grounds its predictions in tissue structures familiar to clinicians, reflecting how pathologists naturally interpret tissue. The model’s predictions correlated with immune profiling and multiplex imaging data, as higher risk scores corresponded to phenotypes less likely to be sensitive to immunotherapy. This provides biological evidence for why certain patients may have better or worse outcomes. Since the approach uses routine pathology slides, if validated as a predictive tool, it could be incorporated into existing clinical workflows without significant expense. The study represents the first deep learning-based pathomics biomarker rigorously validated across international real-world cohorts and a phase III randomized clinical trial.