Abstract 6725: AI-derived H&E biomarkers predict differential survival under immunotherapy vs chemotherapy in phase-3 metastatic NSCLC study NEPTUNE.

Abstract Background: Immunotherapy outcomes in metastatic NSCLC are variable, and existing biomarkers (PD-L1, TMB) incompletely predict benefit. Routine H&E slides are widely available and capture spatial tumor microenvironment (TME) context. We applied AI-Pathology to convert H&Es into quantitative, interpretable TME features and assessed whether they predict survival benefit with immunotherapy (IO) versus chemotherapy. Methods: We conducted a retrospective analysis of H&E slides from NEPTUNE (a phase 3, study in metastatic NSCLC comparing durvalumab+tremelimumab (D+T) vs chemotherapy (CTX)). Slides from 426 patients (D+T n=154; CTX n=272) were processed using models trained independently on external, non-NEPTUNE datasets (>6,000 pathologist-annotated H&Es). Fifteen features (immune, cancer, fibroblast densities/ratios) were extracted in cancer-associated stroma and cancer regions. Overall survival (OS) was analyzed by fitting univariate Coxph regression models with Benjamini-Hochberg FDR control. Treatment × marker interaction terms evaluated predictiveness. Routinely used markers followed cutoffs (PD-L1 ≤25%; bTMB ≤20 mut/Mb) as described in the NEPTUNE study (de Castro Jr G et al., J Thorac Oncol, 2023;18:106-119). Results: Higher stromal fibroblast density was associated with inferior OS in both arms (D+T HR=1.60 [1.14-2.25]; CTX HR=1.36 [1.05-1.76]) with no treatment interaction (nominal p_interaction=0.82), indicating a prognostic effect. Higher stromal immune density was associated with improved benefit for D+T (HR=0.64 [0.45-0.90]; median OS difference +5.8 months; FDR p=0.04) but not CTX (HR=0.84 [0.65-1.08], FDR p=0.38), with a significant treatment interaction (nominal p_interaction=0.03), supporting a predictive effect. Overall, AI-H&E features showed comparable or stronger stratification of D+T outcomes than established routine markers in this dataset, including PD-L1 (HR=0.67 [0.46-0.96], FDR p=0.08), bTMB (HR=0.78 [0.48-1.26], FDR p=0.45), and ECOG (HR=0.74 [0.52-1.07], FDR p=0.19). Conclusions: NEPTUNE did not meet its primary endpoint. The observed prognostic fibroblast signal and predictive immune signal from AI-derived H&E features suggest the potential for enriched selection or stratification to improve detection of checkpoint blockade benefit. These assay-free, interpretable biomarkers warrant prospective validation for clinical application. Citation Format: Chintan Parmar, Vladimir Roudko, Victoria Muckerson, Jorge Blando, Jennifer Daniel, Javier Díez, Sreeharsha Gunda, Alvaro Mendoza Alcala, Alex Franco, Ivan Stambolic, Cristina Benchea, Ross Stewart, Zhongwu Lai, Jorge S. Reis-Filho, Maurizio Scaltriti, Douglas Palmer. AI-derived H&E biomarkers predict differential survival under immunotherapy vs chemotherapy in phase-3 metastatic NSCLC study NEPTUNE [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6725.