Abstract 1016: Reduction of social disparities in clinical cancer feature prediction via somatic evolutionary embedding of mutational data

Whole exome and whole genome sequencing (WES/WGS) of tumors provide rich repositories of information towards understanding the etiology of various cancer and how best to prevent and intervene in its progression. Studies utilizing these technologies, however, notably skew towards the inclusion of patients from white and well-resourced backgrounds. Considerable and ongoing effort has gone into reducing the disparate representation of WES/WGS data from diverse and global populations in the collective digital tumor corpus; however, the rapid development and deployment of machine learning models for cancer has not paused in wait of data reflective of the populace. This dissonance has resulted in imbalanced and, in an absolute sense, small WES/WGS data being used by increasingly complex machine learning models. Targeted sequencing panels, such as those hosted by AACR Project GENIE, are still skewed in a relative sense, but contain many more samples from all populations due to their routine clinical deployment and comparatively inexpensive. However, panel sequencing data is fundamentally limited in the amount of information captured. It is incumbent, then, to maximize the information that can be derived from WES/WGS data as it applies to somatic mutations assayed in these abundant targeted sequencing panels. As cancer is ultimately a disease of somatic evolution, with even social and germline contributors to cancer often actualized via somatic mutation, impact scores—derived from observed evolutionary conservation over extended organismal timescales—are frequently used to contextualize somatic mutations for input into predictive machine learning models. We show that cancer effect size, a disease-specific, empirical measure of evolutionary benefit conferred to the tumor by a mutation, matches or outperforms impact measures to this end across a wide range of clinical features of cancer. Further, we apply a transfer learning approach to leverage the feature importance discovered by the cancer effect size annotated WES/WGS models to train models using targeted panel data. We find this embedding to not only improve the performance on predictive tasks overall, but also that it disproportionately improves performance for underrepresented groups. Lastly, we show that, while improvements to models and clever embedding of data undoubtedly can reduce disparities in model performance, the simple inclusion of samples from more diverse, representative individuals reduces that disparity more effectively. Citation Format: Nic Fisk, Jeffrey P. Townsend. Reduction of social disparities in clinical cancer feature prediction via somatic evolutionary embedding of mutational data [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 1016.