Bridging Graphs and Foundation Models for Biomedicine and Biology
Dr. Ritambhara Singh
February 18th, 2026
Bridging Graphs and Foundation Models for Biomedicine and Biology
Abstract:
Recent advances in foundation models, ranging from large language models (LLMs) to single-cell representation learners, have enabled new ways to analyze biomedical and biological data. Many systems in these domains are inherently relational, with structure encoded in graphs such as drug-protein or cell-cell interactions. However, bridging these structured representations with foundation models requires computationally intensive methods or retraining, reducing accessibility.
In this talk, I will present two computationally inexpensive methods developed in our lab for integrating graph-based representations with foundation models. (1) K-paths extracts structured and biologically meaningful paths from dense biomedical knowledge graphs. Using this information improves the generalization and accuracy of LLMs and enables explainable inference for drug-based prediction tasks. (2) DRIFT leverages diffusion on spatial graphs derived from single-cell spatial transcriptomics (ST) data. Integrating this spatial information into pre-trained single-cell foundation models improves performance for ST tasks such as annotation, alignment, and clustering, outperforming specialized state-of-the-art methods and computationally expensive ST foundation models.
Together, these approaches point toward unifying paradigms for graph-aware foundation models, where structured knowledge and AI models work together to better capture the complexity of biomedical and biological datasets.
Dr. Ritambhara Singh
Associate Professor of Computer Science and Data Science, Brown University
ritambhara@brown.edu | Personal page
Host: Evanjelin Mahmoodi
