Abstract 6179: Machine learning (ML)-spatial quantification of the tumor microenvironment (TME) identifies differences associated with response to bintrafusp alfa (BA) vs pembrolizumab (PEM) treatment in the Phase 3 INTR@PID Lung 037 study

Abstract

Abstract Background: ML digital pathology models can accurately quantify predictive biomarkers like PD-L1.1 We combined tissue conserving ML assessment of the TME in hematoxylin and eosin (H&E)-stained whole slide images (WSI) with RNAseq for high-dimensional TME quantification for biomarker discovery. The orthogonal methods were applied to a retrospective analysis of a Phase 3 trial directly comparing BA, a first-in-class bifunctional fusion protein composed of the extracellular domain of TGF-βRII (a TGF-β “trap”) fused to a human IgG1 monoclonal antibody blocking PD-L1, and the anti-PD-1 PEM. Methods: Non-small cell lung cancer ML models previously trained to exhaustively quantify cell types and tissue regions in H&E-stained tissue were refined and deployed on 273 WSI from the INTR@PID Lung 037 study/NCT03631706. Human interpretable features (HIFs) describing the complex spatial organization of the TME were extracted from model predictions. HIFs were selected for analysis using K-means consensus clustering and describe the frequency of lymphocytes in cancer epithelium (cTILs); macrophages, fibroblasts, and granulocytes in stroma; and the abundance of cancer cells relative to all other cells in tumor. Bulk RNAseq was performed for 186 samples yielding 177 WSI-RNAseq pairs. Spearman correlation between HIFs and gene expression or signature was computed. Top gene associations were analyzed via gene-set enrichment analysis. HIFs were used for immunophenotyping (immune excluded, inflamed, and desert) via prespecified cutoffs on tumor-infiltrating lymphocyte abundance, and for biomarker discovery via Cox modeling of PFS and OS. Results: Gene expression and signatures were associated with HIFs quantifying immune spatial abundance in TME (e.g., cTILs vs. immune checkpoint signature2). Improved PFS and OS were associated with an immune-inflamed phenotype for PEM (PFS, HR=0.36, p=0.007, 95% CI 0.17-0.75; OS, HR=0.37, p=0.065, 95% CI 0.13-1.06). In interaction with BA, stromal macrophages were associated with improved survival (PFS, HR=0.72, p=0.062, 95% CI 0.53-1.02; OS, HR=0.62, p=0.058, 95% CI 0.38-1.02) and an increased likelihood of treatment response (odds ratio=1.61, p=0.015). PFS and OS were nearly identical between arms in immune-excluded cases (PEM vs. BA PFS, HR=0.93, p=0.73, 95% CI 0.63-1.39; OS, HR=1.09, p=0.77, 95% CI 0.62-1.92. Conclusion: Together HIFs and immunophenotyping enabled response prediction to BA and PEM. PEM response was greater for immune-inflamed tumors, while BA response was associated with macrophage infiltration of stroma. ML-based spatial analysis of the TME shows promise for immunotherapy biomarker discovery and validation. 1V Baxi, et al. Modern Pathology. 2022; 35(1):23-32. 2S Mariathasan, et al. Nature. 2018; 554(7693):544-548. Citation Format: John Abel, Andreas Machl, Aslihan Gerhold-Ay, Limin Yu, Darpan Sanghavi, Ben Trotter, Neel Patel, Ylaine Gerardin, Ramprakash Srinivasan, Sergine Brutus, Thomas Mrowiec. Machine learning (ML)-spatial quantification of the tumor microenvironment (TME) identifies differences associated with response to bintrafusp alfa (BA) vs pembrolizumab (PEM) treatment in the Phase 3 INTR@PID Lung 037 study [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 6179.