Digital pathology–based AI spatial biomarker to predict outcomes for immune checkpoint inhibitors in advanced non-small cell lung cancer.

Abstract

8569 Background: Accurate prediction of outcomes with anti-PD-1/PD-L1 immune checkpoint inhibition (ICI) remains a significant challenge in non-small cell lung cancer (NSCLC). In this study, we develop an artificial intelligence (AI) approach for single-cell analysis of H&E-stained whole-slide images (WSIs) to predict objective response and clinical benefit of ICI in two independent cohorts of NSCLC patients. Methods: For biomarker discovery, we analyzed WSIs and clinical data from 118 advanced lung cancer patients at Stanford University. Of these, 46 patients (39%) were treated with ICI monotherapy and 72 (61%) patients were treated with ICI and concurrent chemotherapy. For external validation, 233 advanced lung cancer patients treated with ICI monotherapy at MSKCC were used. Deep learning models were deployed for automated tumor area detection and segmentation of cell nuclei in WSIs. We developed a fully automated cell annotation approach that leverages multiplex immunofluorescence and trained a deep learning model to classify nuclei into 10 cell types on H&E images, including tumor cells and major immune and stromal cells such as T cells, B cells, neutrophils, macrophages, fibroblasts, and endothelial cells. A total of 331 features were computed to quantify cell composition and cell-cell spatial interactions in the tumor microenvironment. Treatment outcomes were assessed using progression-free survival (PFS) and best objective response per the Response Evaluation Criteria in Solid Tumors (v1.1), with statistical significance reported at the 95% confidence level. Results: Five spatial features were included in the prediction model that characterize the cell-cell interactions between tumor cells, fibroblasts, T cells, and neutrophils. In the validation cohort, the spatial biomarker had a strong association with PFS (hazard ratio=5.46, P<0.0001), while the association of PD-L1 expression with PFS was modest (hazard ratio=1.67, P=0.002). For patients with high PD-L1 expression (TPS>50%), the spatial biomarker significantly stratified patients for PFS (hazard ratio=5.21, 95% CI 3.21-8.48, P<0.0001). For predicting objective response, a multivariate model consisting of the spatial features achieved AUROC=0.76 compared to AUROC=0.66 for PD-L1 TPS, while combining the spatial features with PD-L1 expression led to AUROC=0.78. Conclusions: A single-cell computational pathology approachidentifies interpretable spatial biomarkers that predict ICI response and outcomes in advanced lung cancer. The spatial biomarker could help to select patients with high tumor PD-L1 expression who are most suited for ICI monotherapy. Further validation of these findings is warranted.