CAREER: Leveraging Large Language Models and Neural Operator Learning for Spatiotemporal Biomedical Discovery

Imagine being able to read the body's most complex signals and patterns inside cells, tissues, organs, and the brain—just as easily as reading text in a book. That is the goal of this research, which merges advanced artificial intelligence methods with cutting-edge mathematics to transform biological and medical studies. By adapting large language models—the same type of technology behind today's most powerful artificial intelligence tools—to biomedical data, this project seeks to unlock critical insights into how cells function, how diseases progress, how the brain operates, and how best to treat a range of health conditions. The research will develop new artificial intelligence methods that track continuous changes in biological systems over time and space, build foundation models that can handle multiple types of data such as brain scans, heart tests, and clinical notes, and create specialized tools for gene expression data to better understand how cells and tissues interact. These approaches could lead to more accurate diagnoses, more efficient drug development, personalized treatments, and a deeper knowledge of both cellular and brain processes—bringing healthcare closer to being more precise, more effective, and more attuned to each individual's unique biology. This project develops a new class of computational frameworks that unify large language models with operator learning techniques to address key challenges in modeling spatiotemporal phenomena in biomedical research. Neural operator learning generalizes deep neural networks from functions to operators, enabling flexible modeling of high-dimensional and continuous dynamical systems governed by integral equations or partial differential equations. By combining these operator-based approaches with the ability of large language models to interpret and generate symbolic or tokenized representations, the proposed methods capture both local and non-local interactions, manage complex boundary conditions, and accommodate the wide variety of scales and data types inherent in biology and medicine. The research extends operator learning to handle continuous dynamical systems characterized by memory effects and long-range dependencies, develops strategies to unify varied biomedical modalities—such as magnetic resonance imaging, electrocardiogram, and clinical metadata—in a text-based format, and builds new frameworks that transform genomic data into context-rich representations, enabling discovery of patterns in single-cell transcriptomics and multicellular interactions. These advances promise to significantly improve the predictive power and mechanistic interpretability of machine learning models in biomedical contexts, supporting breakthroughs in personalized medicine, disease modeling, and drug discovery. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.