Abstract 2458: Prediction of atezolizumab benefit in NSCLC by computational pathology: Retrospective validation in a phase III trial

Abstract Background: Anti-PD-(L)1 therapies are widely used to treat non-small cell lung cancer (NSCLC). However, suboptimal disease response motivates development of biomarkers to identify patients likely to benefit from these therapies. Here, we demonstrate an artificial intelligence (AI) powered model that analyzes the tumor microenvironment (TME) of hematoxylin and eosin (H&E) images to predict patient benefit from atezolizumab (anti-PD-L1) in NSCLC. Methods: 1, 243 digitized H&E slides were collected from two trials of atezolizumab versus docetaxel in second-line advanced NSCLC. POPLAR, a phase II trial, was used for training (n=162) and OAK, its phase III follow-on, was used as validation (n=1, 081). Lunit Scope IO, a previously developed deep learning model specialized for TME analysis, was deployed on each image to label tissue and cells. Then, the densities of tumor cells, lymphocytes, fibroblasts, macrophages, endothelial cells, and mitotic figures were calculated in the tumor epithelium and intra-tumoral stroma. An elastic-net penalized Cox regression model using these density values was fit on the atezolizumab-treated patients of the training set to predict overall survival (OS) and applied to the entire dataset. To divide patients into model-low/high groups, the risk score threshold maximizing between-arm median OS difference in model-high patients was identified on the training set. Model performance was evaluated by concordance index (c-index) in each treatment arm, by hazard ratio (HR) between arms in the model-low/high groups, and by a Cox regression model with treatment arm, model-low/high category, and their interaction in the validation set. Results: In the tumor epithelium, lymphocytes and tumor cell densities were positively correlated with OS in atezolizumab-treated patients, while mitotic cell and macrophage densities were negatively correlated. In the intra-tumoral stroma, densities of fibroblasts, macrophages, lymphocytes, and endothelial cells were positively correlated with OS. Model risk score was more strongly associated with OS in the atezolizumab arm (c-index=0.58 95% confidence interval: [0.55, 0.61]) than in the docetaxel arm (c-index=0.54 [0.51, 0.57]). Among the n=464 model-high validation set patients, atezolizumab-treated patients had significantly longer OS (HR=0.65 [0.53, 0.80]) with no significant difference between arms in model-low patients (n=617, HR=0.92 [0.78, 1.10]). Model low/high category had a significant interaction (p=0.02) with treatment in survival prediction. Conclusions: We applied AI-driven analysis of histology to large randomized trials to quantify how TME cellular composition correlates with benefit from immunotherapy. Our results suggest that digital pathology-based approaches in combination with AI analysis could yield biomarkers for patient stratification and treatment selection strategies. Citation Format: Patrick Leo, Niha Beig, Barzin Y. Nabet, Chang Ho Ahn, Soohyun Hwang, Sanghoon Song, Chan-Young Ock, Daniel Ruderman. Prediction of atezolizumab benefit in NSCLC by computational pathology: Retrospective validation in a phase III trial [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2458.