Abstract 1068P
Background
Over the past 5 years, immune checkpoint inhibitors (ICI) have become the standard treatment for metastatic non-small cell lung cancer (mNSCLC). However, only a small proportion of patients (pts) derive durable clinical benefit (DCB). The predictive value of PD-L1 score is limited and better predictive biomarkers are needed. Spatial arrangement of immune cells in the tumor microenvironment (TME) is a potential biomarker for ICI efficacy. In this work, we utilized deep-learning (DL) models to extract TME features from digitized H&E whole slide images (WSI) and evaluated their role in predicting outcomes in mNSCLC pts treated with pembrolizumab.
Methods
A total of 76 mNSCLC pts treated with 1st line single-agent pembrolizumab in four medical centers in Israel and the US were identified. Pts were randomly assigned into training (n=57) and test (n=19) sets. Pre-treatment H&E WSI were analyzed using a DL model to identify and classify tumor, immune and fibroblast cells as well as tumor, necrotic and stromal areas; 72 spatial features were calculated per pt. We used 1-year progression-free survival (PFS) to determine DCB and correlated it with the spatial features to train a binary classifier to identify pts with DCB. The resulting classifier was then applied to the test set, and differences in PFS between pts with positive and negative scores were assessed.
Results
The classifier included three different features related to the arrangement of immune cells, tumor cells and necrotic area; the median value was used to determine patients in the test set with either a positive or a negative score. Baseline pt characteristics and PD-L1 scores were similar between the positive and negative groups. In a Kaplan-Meier (KM) analysis, PFS was significantly higher in pts with a positive score compared to pts with a negative score (HR=0.29, 95% CI 0.087-1; p<0.05). Positive pts had a significantly higher median PFS (NR vs. 7.33; p<0.05) and 1-year PFS (66% vs. 22%; p<0.05) than negative pts.
Conclusions
DL models that analyze the TME from H&E WSI can a priori identify mNSCLC pts with DCB on pembrolizumab. Identifying NSCLC pts likely to be exceptionally sensitive to ICI as monotherapy may improve clinical decision making.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Nucleai.
Disclosure
J. Bar: Financial Interests, Personal, Advisory Role: AstraZeneca, MSD, BMS, Roche, Takeda, AbbVie, Bayer, Novartis. A. Zer: Financial Interests, Personal, Invited Speaker: Roche, BMS, MSD, Takeda, Pfizer, Novartis; Financial Interests, Personal, Advisory Board: AstraZeneca, Steba, Oncohost; Financial Interests, Personal, Stocks/Shares: Nixio; Financial Interests, Institutional, Research Grant: BMS. C.J. Langer: Financial Interests, Personal, Advisory Role: Novocure, AstraZeneca, Takeda, Genentech, Merck, Gilead, GSK, Pfizer, OncLive, RTP, Regeneron, Sanofi, Boehringer Ingelheim; Financial Interests, Personal, Research Grant: Eli Lilly, Trizell, Merck, Takeda, Inovio, AstraZeneca, Oncocyte, Janssen, Amgen, Novartis. J.M. Johnson: Financial Interests, Personal, Other, Consulting for Molecular Tumor Boards: Foundation Medicine. Z. Rachmiel, A. Laniado, M. Matalon, E. Markovits: Financial Interests, Personal, Full or part-time Employment: Nucleai. O. Zelichov: Financial Interests, Personal, Leadership Role: Nucleai. All other authors have declared no conflicts of interest.