Predicting response to pembrolizumab in non-small cell lung cancer, by analyzing the spatial arrangement of tumor infiltrating lymphocytes using deep learning.

Abstract

9045 Background: Immune checkpoint inhibitors (ICI) have become the standard treatment for metastatic NSCLC, although only a small proportion of patients derive durable benefit. PDL1 expression is the only approved biomarker to select NSCLC patients for treatment with single-agent pembrolizumab, however its predictive value is limited and better predictive biomarkers are needed. The spatial arrangement of immune cells in the tumor microenvironment (TME), namely tumor infiltrating lymphocytes (TILs), emerges as a potential biomarker for ICI efficacy. In this work, we utilized deep-learning (DL) models to extract TME features from digitized H&E slides and evaluated their predictive and prognostic role in patients with mNSCLC treated with Pembrolizumab. Methods: NSCLC patients (n=90) treated with single-agent 1st line pembrolizumab in two centers were identified. 47 patients from one center were used to train the model, and 43 patients from another center were used for validating the model. Pre-treatment H&E whole slide images (WSI) were analyzed using a deep-learning model to identify and classify tumor cells, TILs, tumor and stromal areas, and spatial features were calculated. Spatial features were correlated with clinical outcome data to train a binary classifier that identifies patients with a favorable clinical outcome. The resulting classifier combined three spatial features and three clinical features. The classifier was then applied to the validation set and differences in duration of treatment (DOT), and overall survival (OS) between patients with positive and negative scores were assessed. Results: The classifier identified patients in the validation set to have either positive (n=18) or negative (n=25) scores. Baseline patient characteristics and PDL1 score were similar between the positive and negative groups. In a Kaplan-Meier (KM) analysis, OS was significantly higher in patients with a positive score compared to patients with a negative score (HR=0.35, 95% CI 0.13-0.98; p<0.05). Positive patients had a significantly higher median OS (NR vs.17.8m, p<0.05) and 2-year OS (70.8% vs. 33%, p=0.02) than negative patients. Median DOT was also higher in positive patients compared to negative patients (10.1m vs. 6.5m). Conclusions: Deep-learning models that analyze the TME from H&E whole-slide images can identify NSCLC patients with durable benefit on Pembrolizumab. Identifying NSCLC patients who are exceptionally sensitive to anti-PD-1 therapy as monotherapy may improve clinical decision making and spare patients the unnecessary adverse effects associated with the addition chemotherapy or another IO agent.