NLM DIR Seminar Schedule

Abstract:

Drug response prediction is a central challenge in computational biology, with the ultimate goal of enabling clinically actionable predictions that generalize from cell lines to patients. Despite significant progress, existing approaches face key limitations: (1) limited generalization to unseen drugs and cell types, and (2) lack of biological interpretability and evidence grounding.
In this talk, I present research that addresses these challenges from complementary perspectives. First, I introduce a graph neural network-based model (drGT) that enhances generalization to unseen drugs and cell lines by modeling interactions between drugs, genes, and cells with attention-based interpretability. Second, I present DrugAgent, a large language model (LLM)-based multi-agent framework for evidence-based reasoning that integrates knowledge graphs, machine learning predictions, and PubMed-derived literature via retrieval-augmented generation (RAG) to produce biologically grounded explanations. Together, these approaches move beyond prediction toward models that are generalizable, interpretable, and grounded in biological processes, providing a foundation for clinically meaningful drug response prediction. Lastly, I will discuss future plans for this research area related to "world models" (i.e., models with internal representations of environments that better fit physical constraints) in how such machine learning models may improve the accuracy and interpretability of drug response predictions.