Abstract 1055: Inclusion of large genomics databases in disease modeling improves outcomes across ancestral populations

Abstract Rapid advancements in genomic sequencing technologies have revolutionized the study and treatment of human diseases. Even with the wealth of genomic information available, successful application of developed treatments is often limited to the human populations represented in the original study, suggesting that these studies are overfitting to their data and thus limiting scientific understanding of complex traits and diseases. Human populations have developed population-specific variants in response to evolutionary changes as a result of human migrations and other stressors which can impact disease occurrence and outcome. Lack of sufficient sampling for equal representation of this variation in disease studies has led to disparate outcomes and treatments. While obtaining a representative study population for a disease is often unrealistic, equitable machine learning approaches can bridge the gap. Here, we demonstrate an equitable machine learning approach, PhyloFrame, showing that the inclusion of publicly available genomic resources in disease modeling methods can force machine learning and artificial intelligence models to be more robust to data distribution shifts caused by unequal sampling of human populations, thereby increasing model generalizability. We demonstrate PhyloFrame’s utility across several diseases at high risk of ancestral bias, which we identified via an analysis of heritable disease signatures. We first perform statistical enrichment of ancestry-specific alleles within existing disease signatures for Mendelian disorders by calculating variant frequency enrichment for each of the populations in gnomADv4.1 and the 1000 Genomes Project. After identifying the diseases at highest risk for ancestral bias, we apply our PhyloFrame method to RNA-seq datasets from those diseases and contrast PhyloFrame’s performance in each ancestry group to a comparable benchmark model that does not incorporate equitable machine learning techniques. Our results demonstrate that incorporating publicly available genomic data into disease modeling methods improves outcomes for all. Growing resources for publicly available large-scale genomic databases provide unprecedented opportunities to understand disease. Resulting analysis of variation across human populations is a powerful approach to supplement disease models. Here we demonstrate how to incorporate this information into disease modeling, and show it is a unique and feasible approach to advancing equity in precision medicine. Citation Format: Leslie A. Smith, Franco Krepel, James A. Cahill, Ji-Hyun Lee, Kiley Graim. Inclusion of large genomics databases in disease modeling improves outcomes across ancestral populations [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 1055.