Graph Neural Networks for Molecular Networks and Drug Discovery
This presentation by Stanford PhD student Huang Kexin explores the challenges and innovations of applying graph machine learning to molecular and biomedical networks, introducing specialized GNN architectures, actionable hypothesis generation, domain‑scientist interfaces, few‑shot learning, and the Therapeutics Data Commons for accelerating drug discovery.
Speaker Huang Kexin, a first‑year Stanford computer‑science PhD student, introduces his research on applying machine learning, specifically graph neural networks (GNN), to biomedical problems, focusing on molecular networks.
Biological systems are modeled as complex multi‑scale graphs linking proteins, RNAs, drugs, and diseases; predicting interactions within these graphs can accelerate drug development.
01 – Special considerations for GNNs on molecular networks: Traditional GNNs rely on homophily, which does not hold for molecular graphs. The proposed SkipGNN rewires edges to capture “skip similarity,” improving performance over DTI, DDI, PPI, and GDI baselines.
02 – Generating actionable hypotheses: Beyond prediction, the work aims to produce interpretable hypotheses for biologists, exemplified by drug‑drug interaction (DDI) analysis. Using graph methods, the team predicts that the combination of Paroxetine and Hydroflumethiazide may increase risks of orthostatic hypotension and aplastic anemia.
03 – Domain‑scientist perspective and XAI: Effective AI for biologists requires clear explanations. Three visualization methods are proposed: Neighbor Nodes (important genes per disease/drug), Subgraph (knowledge‑graph extraction), and Paths (tracing drug‑disease relationships).
04 – Few‑shot learning for biomedical graphs: Data scarcity and cross‑species translation are addressed with meta‑learning. The G‑Meta framework extracts subgraphs and leverages their similarity to solve three related problems involving varying labels and graph structures.
05 – Graph opportunities in therapeutic discovery: The presentation highlights diverse graph types (molecular, protein, cellular, knowledge graphs) and introduces the Therapeutics Data Commons (TDC), which offers 66 ML‑ready datasets, tools, and benchmarks to advance treatment‑discovery research.
The talk concludes with a call to explore these methods, download TDC data with minimal code, and engage with the DataFun community.
DataFunTalk
Dedicated to sharing and discussing big data and AI technology applications, aiming to empower a million data scientists. Regularly hosts live tech talks and curates articles on big data, recommendation/search algorithms, advertising algorithms, NLP, intelligent risk control, autonomous driving, and machine learning/deep learning.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.