Abstract 910: Clinical inference and biological dissection of tumor ploidy and heterogeneity in cutaneous melanoma for immunotherapy response using deep learning

Abstract

Abstract Background: Tumor ploidy and heterogeneity demonstrated to be pivotal in predicting immunotherapy response in cutaneous melanoma in several independent cohorts (Liu NatMed2019, Tarantino BioRXiv2022). Their estimation can guide more personalized and rational utilization of these immunotherapies. However, 1) the biology underpinning ploidy and heterogeneity is unknown; and 2) these findings were derived in patients using retrospective research tumor-normal paired whole exome sequencing which is not performed for clinical management. Methods: This study addresses this gap by employing deep learning models to predict these crucial markers using routinely available clinical assays, including H&E images. Attention-based computer vision models enable the identification of key morphological features in H&E slides. Segmentation of tumor tissue to perform automated masking, enhances ploidy inference.Moreover, biologically informed neural networks (P-Net, Elmarakeby Nature2021) uncover transcriptional and genomic features linked to ploidy and tumor heterogeneity. Our models are trained on publicly available data (e.g., Liu et al Nature Medicine 2019; TCGA SKCM) from melanoma cohorts and further validated in independent cohorts to ensure robustness. Results: We developed and validated automated tumor tissue masking, enabling the prediction of Whole Genome Doubling (WGD) from H&E Slides with an AUC > 0.75. Attention-based models identify distinct Tumor Microenvironment (TME) structures predictive of high tumor heterogeneity. P-Net application revealed the Calmodulin pathway, previously associated with regulating proliferation in cancer and targeted with chemotherapy, as intricately linked to high tumor heterogeneity, providing valuable insights into underlying mechanisms. Conclusion: In conclusion, our study strategically harnesses and integrates existing datasets to rigorously test, refine, and validate hypotheses concerning the biological and therapeutic implications of genomic heterogeneity and ploidy. Notably, our predictive models with automated tumor masking demonstrate a remarkable AUC >0.75 for biomarkers traditionally challenging to derive from clinical assays. This breakthrough opens avenues for novel therapeutic strategies targeting genomic heterogeneity and ploidy, providing a transformative potential to enhance patient care and outcomes. Citation Format: Marc Glettig, Giuseppe Tarantino, Tyler Aprati, Haitham Elmarakeby, David Liu. Clinical inference and biological dissection of tumor ploidy and heterogeneity in cutaneous melanoma for immunotherapy response using deep learning [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 910.