Abstract 7383: G2PT: Mechanistic genotype-phenotype translation using hierarchical transformers

Abstract

Abstract Motivation: Many machine learning models have been proposed that use genotypes to predict various phenotypes. Recently, these models have focused not only on an accurate prediction but mechanistic interpretation, in that they attempt to describe the hierarchy of biological systems underlying the predicted phenotype. Such models still face major challenges, however, including how to robustly quantify the importance of systems mediating phenotypic outcomes and how to represent the bidirectional flow of information among genes, systems, phenotypes, and cell environment. Results: To address these challenges we introduce G2PT, a general Genotype-to-Phenotype Transformer, which we apply here to the classic G2P problem of modeling tumor cell sensitivity to drugs. We show that the framework of hierarchical transformers can represent biological systems spanning a wide range of physical scales and provide a quantitative assessment of the interplay between biological components important for phenotype prediction. When used to analyze a high-throughput cell screen of 436 drugs exposed to 1,097 cell types, we show that the model achieves good performance in predicting cell sensitivity for 83% of drugs (⍴ > 0.5 in cross-validation). G2PT provides a robust and interpretable model with potential application to many challenges in genotype-phenotype translation. Citation Format: Ingoo Lee, Sungjoon Park, Hojung Nam, Trey Ideker. G2PT: Mechanistic genotype-phenotype translation using hierarchical transformers [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 7383.