MAI-T1D Resource Hub
Foundation models, data platforms, tools, and publications for multimodal AI research in Type 1 Diabetes — all in one place.
About MAI-T1D
MAI-T1D is a five-institution project building multimodal AI foundation models for type 1 diabetes, trained and benchmarked across genomics, single-cell, spatial, and immunological data from existing T1D resources to comprehensively characterize disease progression. We also study the ethical side of this work — participation leakage, content leakage, and safeguards such as split learning for private whole-genome sequencing data. See the publications below.
This hub hosts the consortium's public output: the models, the data platforms behind them, the governance tooling that tracks how training data evolves, and the publications. Everything is openly available or points to the access process for controlled data.
- Award
- NIH / ODSS · OT2OD038003
- Lead institution
- University of Michigan
- Partners
- Vanderbilt · Cornell · USF · UCLA
- Data sources
- HPAP · TEDDY · TrialNet · TRIGR · ImmPort
- Contact
- Kai LiuProject Managerkailiua [at] umich.edu
This work used resources developed by the MAI-T1D consortium, supported by the National Institutes of Health Office of Data Science Strategy through the Advancing Health Research through Multimodal AI program under award OT2OD038003.
Foundation Models
4 modelsGenomics Foundation Model
Predicts epigenomic profiles and chromatin organization from DNA sequences. Core genomic FM of the MAI-T1D foundation stack.
Single-cell Foundation Model
Foundation model for single-cell transcriptomics and epigenomics analysis. Supports cell-type annotation, gene program discovery, and perturbation prediction.
Spatial Transcriptomics Foundation Model
Foundation model for spatial transcriptomics. Models tissue architecture and cell–cell interactions in pancreatic islet microenvironments.
Spatial Omics Foundation Model — VOICE
Vision–omics foundation model that predicts single-cell gene expression directly from H&E morphology, trained contrastively on 23M cells paired with Xenium data. Combines direct regression with retrieval from reference cells, weighted per gene.
Knowledge & Data Platforms
9 resources
GLKB — Genomics Literature Knowledge Base
PubMed-integrated knowledge platform with ML-enhanced search, agentic Q&A in biomedicine, and AI-driven hypothesis generation.
AI ready Datasets - HPAP
Curated, AI-ready datasets derived from the Human Pancreas Analysis Program — preprocessed and standardized for direct use in MAI-T1D foundation model training and benchmarking.
EPCOT Training Data
Training and evaluation data behind the Genomics Foundation Model (EPCOT v2), which predicts cell-type- and context-specific regulatory signals from ATAC-seq data and DNA sequences.
Use Cases
Showcased applications and end-to-end workflows demonstrating how MAI-T1D models and data platforms can be applied to real-world T1D research questions.
Tools, Agents & Governance
2 resourcesMAI-T1D Foundation Stack
Central GitHub repository that integrates the genomics, single-cell, spatial transcriptomics, and spatial omics foundation models into one pipeline for T1D research.
Data Governance Knowledge Graph
PROV-inspired knowledge graph for automated governance of evolving multimodal AI training data across the MAI-T1D consortium. Interactive demo with provenance graph, impact analysis, and governance agent views.
Ethical AI & Data Privacy
3 topicsMembership Inference
Participation leakage — whether an individual's presence in the training data can be inferred from a trained model. We audit MAI-T1D models for this risk and quantify how much it grows as modalities are combined.
Reconstruction of Training Records
Content leakage — whether original training records can be reconstructed out of a trained model. We test how much of a donor's multimodal record a model memorizes, and what limits that memorization.
Split Learning
A safeguard rather than a risk — the model is partitioned across sites so private whole-genome sequencing data stays protected during training. Quantized-TinyLLaVA compresses the intermediate features exchanged between partitions, cutting communication ~87.5% at 2-bit.
Publications
Consortium work on type 1 diabetes data, biomedical models, and the platforms listed on this hub.
VOICE: A Vision-Omics Foundation Model Integrating Direct and Retrieval-Based Prediction of In-situ Single-Cell Gene Expression
Aligns cell-centered H&E morphology with single-cell expression embeddings by contrastive learning over 23M cells, then predicts expression through a direct-regression branch and a retrieval branch fused with per-gene weights — generalizing to held-out patients, slides, and partially overlapping Xenium panels.
CelLink: integrating single-cell multi-omics data with weak feature linkage and imbalanced cell populations
Optimal-transport integration method that aligns scRNA-seq with spatial proteomics even when features link weakly and cell populations are imbalanced, enabling cross-modal imputation for multimodal foundation models.
Developing a general AI model for integrating diverse genomic modalities and comprehensive genomic knowledge
Multi-task genomic model (EPCOT) that takes ATAC-seq and DNA sequence as input and predicts nascent RNA, high-resolution chromatin organization, and non-coding variant effects across human and mouse.
Enhancing Biomedical AI Foundations: Genomic Literature Knowledge Base Boosts LLMs' Mastery of Biomedical Literature
GLKB is a knowledge graph of 14.6M relationships over 3.2M entities drawn from 33M PubMed abstracts; the paired agent lifts eight state-of-the-art LLMs by up to 27.5% on PubMedQA.
Biomedical Question Answering via Multi-Level Summarization on a Local Knowledge Graph
Builds a local knowledge graph from retrieved documents using propositional claims, then layerwise-summarizes it to contextualize a small language model — matching or beating RAG baselines on biomedical QA.
AutoPCR: Automated Phenotype Concept Recognition by Prompting
Prompt-based phenotype concept recognition that generalizes to new ontologies and unseen text without ontology-specific training, with an optional self-supervised boost.
Quantized-TinyLLaVA: A New Multimodal Foundation Model Enables Efficient Split Learning
Quantizes the intermediate features exchanged between split-learning partitions into discrete codes, cutting communication overhead ~87.5% at 2-bit while matching the 16-bit model across five benchmarks.
Fact or Guesswork? Evaluating Large Language Models' Medical Knowledge with Structured One-Hop Judgments
Builds the UMLS-derived Medical Knowledge Judgment Dataset to isolate factual medical recall from multi-hop reasoning, exposing weak recall on rare conditions and poor calibration.
Machine-learning methods work by MAI-T1D investigators. These papers are not T1D-specific, but the techniques feed the models and evaluation pipelines above.
Preference Leakage: A Contamination Problem in LLM-as-a-judge
Identifies a contamination effect where an LLM judge systematically favors student models trained on synthetic data from a related generator, quietly inflating evaluation scores.
Chasing the Tail: Effective Rubric-based Reward Modeling for Large Language Model Post-Training
Traces reward over-optimization to misspecification at the high-reward tail, then uses off-policy examples in a rubric-based reward workflow to separate "excellent" from merely "great" and curb reward hacking.
Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoning
REFRAIN decides when to stop chain-of-thought reasoning, pairing a two-stage stop discriminator with a sliding-window bandit that adapts the threshold to problem difficulty — no training required.
InsideOut: Measuring and Mitigating Insider–Outsider Bias in Interview Script Generation
A 4,000-prompt benchmark across 10 cultures showing LLMs write as cultural insiders for the US but default to an outsider stance elsewhere, plus prompt- and agent-based mitigations.
BLUR: A Bi-Level Optimization Approach for LLM Unlearning
Recasts unlearning as bi-level optimization — lower level minimizes forget loss, upper level preserves utility — instead of the usual weighted sum, yielding better forget/retain trade-offs.
Open-Domain Safety Policy Construction
Deep Policy Research drafts a full content-moderation policy from seed domain information alone, iteratively proposing web queries, distilling sources into rules, and indexing them into a policy document.
Self-Routing RAG: Binding Selective Retrieval with Knowledge Verbalization
Treats the model's own parametric knowledge as a first-class retrieval source, jointly training source selection, verbalization, and generation — ~29% fewer retrievals at ~5.1% higher accuracy.
Literature Review
2 resourcesTEDDY Metabolomics
Literature review and data explorer for TEDDY metabolomics and lipidomics datasets — supporting multimodal T1D research and biomarker discovery.
TrialNet Literature Review
Literature review of the TrialNet data cohort — covering natural history studies, autoantibody screening, and intervention trials in T1D progression.
Wiki & Documentation
2 resourcesMAI-T1D Project Wiki
Internal knowledge base covering project protocols, data dictionaries, onboarding guides, and cross-institutional coordination documentation.
MAI-T1D GitHub Organization
Central code repository for all MAI-T1D foundation models, analysis pipelines, governance tools, and shared utilities.