Data-driven discovery of drug response mechanisms through biologically informed deep learning and large language models

About This Grant

Summary/Abstract Developing more effective therapies requires a comprehensive understanding of the molecular mechanisms that govern drug responses. In pursuit of this goal, large-scale pharmacogenomic resources have been generated, capturing baseline multi-omics, systematic viability screens of drug and gene perturbations, and transcriptomic signatures of treatment-induced changes. Although each dataset reflects an important aspect of treatment response, these resources remain siloed across modalities and underutilized due to the lack of integrative computational frameworks. Building on our prior work in deep learning and bioinformatics tool development, we propose to address two critical gaps: i) the need for a unified framework to systematically integrate multi-modal pharmacogenomic data for modeling the drug–gene–pathway–response axis; and ii) the need to make these tools and data resources more accessible to biomedical researchers without programming expertise. Our central hypothesis is that biology-guided, multi-modal integration using deep learning will enable accurate prediction and interpretation of drug effects, from molecular perturbations to phenotypic outcomes. We will develop a novel deep learning architecture that uses transfer learning to combine knowledge from transcriptomics, drug features, drug and CRISPR screens, and perturbation signatures. The model will bridge drug-induced molecular changes and phenotypic viability effects through pathway-level representations, enabling mechanistic insight and generalizable prediction of drug responses across a wide range of biological contexts (Aim 1). To complement this systems-level model and enhance its real-world applicability, we will develop a scalable computational framework for inferring and evaluating drug mechanisms directly from transcriptomic perturbation signatures, leveraging embeddings derived from large language models. This approach enables gene set-free discovery of both known and novel drug mechanisms, particularly in under-annotated or complex settings (Aim 2). All models and findings will be rigorously validated using independent datasets. To maximize impact and accessibility, we will develop a user-friendly web platform that integrates these tools and data resources, allowing users to submit their own data, explore predictive outputs, and visualize pathway- and mechanism-level interpretations, without requiring programming expertise (Aim 3). Proposed in response to PAR-25-238, this study aligns closely with the National Library of Medicine’s mission to advance data-driven discovery in biomedical science. The project is supported by a multidisciplinary team with expertise in bioinformatics, pharmacogenomics, artificial intelligence, and software development. Successful completion will yield: i) the first deep learning framework to comprehensively model the drug–gene–pathway–response axis through systematic integration of multi-modal pharmacogenomic data; ii) deeper biological knowledge of how drugs drive outcomes; iii) a novel methodology for high-resolution, gene set-free drug mechanism discovery; and iv) an open-access platform that democratizes advanced pharmacogenomic modeling across a broad range of diseases.