What question did this study set out to answer?

This research aims to define how regulatory variation and cell-state transitions affect the progression of type 1 diabetes (T1D) using AI models.

June 7, 2026

3002-LB: Multimodal AI System to Infer the Progression of Type 1 Diabetes

Key Points

This research aims to define how regulatory variation and cell-state transitions affect the progression of type 1 diabetes (T1D) using AI models.
Integrated genotype, chromatin accessibility, transcriptomic, proteomic, and spatial profiling.
Utilized three AI foundation models on datasets from the Human Pancreas Analysis Program (HPAP).
Validated findings with longitudinal samples from the TEDDY cohort.
Developed a genomic model linking risk variants to β-cell gene expression and regulatory networks.
Identified β-cell subpopulations showing stress activation and reduced maturity through single-cell analysis.
Characterized disruption of islet organization with immune cells near stressed β-cells, indicating inflammatory signaling.

Abstract

Introduction and Objective: Type 1 diabetes (T1D) is characterized by immune-mediated destruction of pancreatic β cells, progressing from autoantibody (AAB) positivity to insulin deficiency and islet failure. We integrated genotype, chromatin accessibility, transcriptomic, proteomic, and spatial profiling of human pancreatic tissue using three AI foundation models to define how regulatory variation, cell-state transitions, and microenvironmental context change across progression. Methods: Our models were initially trained on large-scale publicly available datasets, then applied to pancreatic datasets from the Human Pancreas Analysis Program (HPAP) and validated using longitudinal samples from the TEDDY cohort, enabling cross-cohort evaluation of disease-associated regulatory, cellular, and spatial changes. Results: First, we developed a genomic foundation model integrating whole-genome sequencing and ATAC-seq to predict gene expression by linking risk variants to cell type-specific regulatory elements and target genes. Many variants localized to β-cell-active enhancers and were connected to genes governing β-cell identity and stress responses, suggesting inherited variation reshapes regulatory networks in these cells. Second, a single-cell foundation model integrating scATAC-seq, scRNA-seq, and CITE-seq resolved endocrine cell states across T1D progression. Beyond canonical α and β cells, we identified β-cell subpopulations marked by stress activation, reduced maturity signatures, and metabolic reprogramming. Third, a spatial foundation model characterized islet architecture and cellular neighborhoods. We observed progressive disruption of islet organization, with immune cells enriched near stress-associated β-cell states and local inflammatory signaling, alongside diminished α-β spatial coupling, as discovered via the Pancreatlas Platform. Conclusion: Together, these findings integrate genetic risk, regulatory remodeling, cellular plasticity, and spatial reorganization to provide a unified view of T1D progression. Disclosure X. Luo: None. K. Liu: None. X.B. Bao: None. H. Zeng: None. Y. Tao: None. H.T. Vu: None. D. Leng: None. A. Mohammed: None. M. Fayyaz: None. F. Feng: None. Y. Wang: None. S. Lee: None. J. Vandana: None. A.K. Taylor: None. Z. Zhang: None. S. Johnson: None. R. Mao: None. D.C. Saunders: None. J. Cartailler: None. S. Parker: Research Support; Current; Pfizer Inc. Consultant; Ended; Novo Nordisk. Y. Huang: None. K.G. Young: None. D. Tewey: None. W. Wang: None. M. Brissova: None. S. Chen: Stock/Shareholder; Current; iOrganBio Inc. Stock/Shareholder; Ended; Oncobeat. J. Liu: None. Funding NIH and NIDDK

Mark Helpful

Bookmark

Relay