Agentic and generative AI for in silico modeling of human pancreatic tissue in T1D

The National Institute of Diabetes and Digestive and Kidney Diseases awarded President and Fellows of Harvard College $698,321 on September 1, 2026, to develop agentic and generative artificial intelligence systems for in silico modeling of human pancreatic tissue in Type 1 Diabetes under the Diabetes, Digestive, and Kidney Diseases Extramural Research program (CFDA 93.847).

Harvard will develop two integrated computational platforms: FUSEMAP, extending a deep learning framework to integrate single-cell genomics, spatial transcriptomics, protein profiling, and functional imaging data from healthy and diseased pancreatic tissue to enable transcriptome-scale gene imputation, cross-sample alignment, and segmentation of anatomically and pathologically relevant domains; and TSAGENT, a multimodal generative AI system capable of autonomous data analysis, simulation of disease-associated perturbations, and mechanistic hypothesis generation that incorporates patient metadata to model inter-individual variability. Together, FUSEMAP and TSAGENT will form the computational foundation of a dynamic virtual pancreas tissue model to be validated for mechanistic dissection of spatial heterogeneity in beta-cell loss, immune-islet interactions, and tissue responses to therapeutic interventions.

Work is performed in Boston, Massachusetts with a period of performance from September 1, 2026, through May 31, 2031. The assistance type is a Project Grant.

Abstract

Type 1 diabetes (T1D) is characterized by the progressive autoimmune destruction of pancreatic β cells, accompanied by substantial spatial and temporal heterogeneity in tissue remodeling. Existing models fail to capture the multicellular complexity and spatial organization of disease progression, limiting our ability to understand pathogenesis or predict therapeutic responses. To address this critical GAP, we propose to develop the first unified, agentic, and generative virtual tissue platform for T1D, integrating recent advances in spatial transcriptomics, deep learning-based multimodal data harmonization, and large language model (LLM)-based AI agent for flexible reasoning, contextual learning, and zero- and few-shot generalization. The proposed research will extend our previously developed deep learning framework, fusemap, to integrate single-cell genomics, spatial transcriptomics, protein profiling, and functional imaging data from both healthy and diseased pancreatic tissue. This integration will enable transcriptome-scale gene imputation, robust cross- sample and multimodal alignment, and segmentation of anatomically and pathologically relevant domains. We will also advance our spatial transcriptomics AI agent into a tissue spatial AI agent (tsagent)-a multimodal, generative AI system capable of autonomous data analysis, simulation of disease-associated perturbations, and mechanistic hypothesis generation. Tsagent will incorporate patient metadata (e.g., age, SEX, BMI, HBA1C) to model inter-individual variability and iteratively refine its inferences based on new data. Together, fusemap and tsagent will form the computational foundation of a dynamic virtual pancreas tissue model. This platform will be validated for mechanistic dissection of spatial heterogeneity in β-cell loss, immune- islet interactions, and tissue responses to therapeutic interventions. Importantly, all models, software tools, and datasets will be disseminated through an open-access web portal, with extensive community outreach to promote adoption. By enabling continuous refinement through user contributions and supporting rigorous, interpretable in silico experiments, this work will establish a novel scientific infrastructure for understanding and treating T1D at the tissue level.