Abstract 3656: Using a hierarchical transformer for mechanistic response prediction to chemotherapy

Abstract How a cancer patient responds to chemotherapy drugs can be notoriously difficult to predict, as the mechanisms triggering these responses are not completely understood. The superiority of attention-based models in machine learning, such as Transformers, offers the potential to improve drug response prediction while gaining deeper mechanistic understanding. Here, we develop and evaluate a hierarchically structured transformer model for chemotherapy response. The transformer architecture leverages prior knowledge of biological structure and function, enabling it to learn how genetic alterations perturb the states of genes and higher order multigenic systems in governing a drug response. This model accurately predicts response to 12 chemotherapies with high attention to systems known to function in replication stress, such as cell cycle regulation and genome maintenance via p53, as well as unexpected systems related to the actin cytoskeleton and regulation of cellular structure. This work demonstrates how transformer-based modeling can improve chemotherapy response prediction and reveal deeper understanding of molecular systems that underlie therapeutic outcomes. Citation Format: Zach Wallace, Ingoo Lee, Sungjoon Park, Akshat Singhal, Xiaoyu Zhao, Trey Ideker. Using a hierarchical transformer for mechanistic response prediction to chemotherapy [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 3656.