NLM DIR Seminar Schedule

Graph-Aware Interpretable Drug Response Prediction and LLM-Driven Multi-Agent Drug-Target Interaction Prediction

Machine learning (ML) holds promise for accelerating drug discovery, a lengthy and expensive process. However, the black-box nature of deep learning (DL) models hinders their clinical applicability. Here, we propose two methods: 1) drGT, graph-aware interpretable methods for drug response prediction (DRP), and 2) DrugAgent, a large language model (LLM)-driven multi-agent tool for drug-target interaction (DTI) prediction.

drGT is a graph neural network (GNN)-based approach utilizing a heterogeneous network (e.g., a graph with nodes representing genes, drugs, and cell lines). drGT was evaluated for DRP under randomly masked 5-fold cross-validation and for unseen drugs and cell lines. For prediction, drGT achieved an AUROC of up to 94.5% under random splitting, 84.4% for unseen drugs, and 70.6% for unseen cell lines, comparable to existing benchmark methods, while providing interpretability. Crucially, 63.67% of the drug-gene associations identified by drGT are independently supported by PubMed literature or an established DTI prediction model, validating its interpretability.

Regarding DrugAgent, our multi-agent LLM system for DTI prediction combines multiple specialized perspectives with transparent reasoning. We adapt and extend existing multi-agent frameworks by (1) applying a coordinator-based architecture to the DTI domain, (2) integrating domain-specific data sources (including ML predictions, knowledge graphs, and literature evidence), and (3) incorporating Chain-of-Thought (CoT) and ReAct (Reason+Act) frameworks for transparent DTI reasoning. In comprehensive experiments using a kinase inhibitor dataset, our multi-agent LLM method significantly outperformed a non-reasoning GPT-4o mini baseline, achieving a 45% higher F1 score (0.514 vs 0.355).

Using Ontologies to Make Knowledge Computable

Ontologies have played a critical role in the last 25 years in knowledge representation, data indexing and retrieval, and data integration. Projects like the Gene Ontology and the Human Phenotype Ontology, coordinated through the Open Biomedical Ontology (OBO) Foundry, demonstrate the ability to encode knowledge in a computable format and link it to external sources of knowledge from other domains at scale. As the volume of data used in biomedical experiments grows, efforts to improve data stewardship, such as the FAIR principles, have identified ontologies as a key tool for making data and metadata machine actionable. In parallel with this are efforts to produce computable knowledge, defined as knowledge that is explicitly represented such that it can be parsed and reasoned upon using computational methods to derive new knowledge, and represent it in resource like knowledge graphs. In this presentation, I provide background on what ontologies are and how they are used, and introduce the Cell Knowledge Network—a DIR-funded knowledge network designed to connect data and knowledge about cell phenotypes to knowledge about anatomy, drug targets, human phenotypes and disease, and other domains.

TBD