Genetic Association and Machine Learning Improve the Prediction of Type 1 Diabetes Risk-Dr. Carolyn McGrail, PhD & Dr. Alexandra Ghaben, MD PhD
On Demand video from the May 21 2026 talk
🎬 ON DEMAND VIDEO
Genetic Association and Machine Learning Improve the Prediction of Type 1 Diabetes Risk Dr. Carolyn McGrail, PhD & Dr. Alexandra Ghaben, MD PhD | May 2026
📺 Watch the full talk here → 📄 Read the paper → Nature Genetics, 58:1062–1072 (2026)
💬 Key Quote
“Genetic risk scoring allows us to capture a broader pool of both children and adults who are at high risk for T1D but who might otherwise be missed. This supports close monitoring to reduce the risk of complications such as diabetic ketoacidosis at diagnosis and helps identify individuals eligible for preventative therapies like teplizumab.” — Dr. Carolyn McGrail, PhD
🔬 Foundational Insights as They Apply to T1D
Every tool in the current T1D risk prediction toolkit carries a blind spot. Autoantibody screening — the workhorse of programs like TrialNet’s Pathway to Prevention — is powerful but requires repeated testing over time, is most accessible to individuals with an affected family member, and captures risk only after the immune system has already mobilized. HLA typing identifies the classical DR3/DR4 high-risk haplotypes that confer 10–50-fold elevated risk, but roughly half of T1D cases arise in individuals who lack these canonical haplotypes entirely — a population that current genetic screening consistently misses. And existing genetic risk scores (GRS), including the current gold standard GRS2, are built on additive statistical models: they sum the effects of individual risk variants as if each acts independently, treating the genome as a list of parts rather than a network of interacting systems. That assumption is false, and the consequence is a measurable gap between the risk these scores capture and the full genetic architecture of the disease.
The deeper problem this exposes is not merely technical. T1D is not one disease. It presents across a spectrum from early childhood onset with powerful HLA-driven autoimmunity, to adult-onset forms with attenuated immune signatures and greater beta cell vulnerability. It affects individuals of all ancestries, in family contexts and without, with widely varying rates of progression, complication burden, and response to immunotherapy. A risk prediction framework that treats all T1D as genetically equivalent — driven by the same HLA haplotypes, modeled by the same additive formula — will inevitably be most useful for the patients most studied and least useful for everyone else.
T1GRS, developed by Dr. Carolyn McGrail, Dr. Alexandra Ghaben, Dr. Kyle Gaulton and collaborators at UC San Diego and the Broad Institute, is a direct response to these compounding limitations. It is built on the most expansive T1D genome-wide association study ever conducted — over 20,000 T1D cases and nearly 800,000 controls imputed across 60 million variants — and trained using gradient boosting, a machine learning method that inherently captures the nonlinear interactions between features that additive scores cannot model. The result is not merely a better risk score. It is a framework that reveals, for the first time, 154 significant nonlinear interactions between genetic variants, identifies four distinct genetic subtypes of T1D with meaningfully different clinical trajectories, and is openly available via GitHub for any research group or clinical program to apply to existing genotype data. T1GRS is a tool, a biological discovery engine, and a first step toward precision medicine in T1D — all at once.
🎯 Core Premise
T1GRS is a gradient boosting machine learning model trained on 199 genetic variants — 27 MHC signals from stepwise conditional fine-mapping, 70 previously characterized HLA alleles, and 133 non-MHC signals from SuSiE fine-mapping across the genome — that captures nonlinear interactions between genetic features that additive models miss. In head-to-head comparison with GRS2 across every tested population, T1GRS outperforms: AUC 0.937 vs. 0.916 in the discovery cohort, with the most pronounced gains in the NIH All of Us population-based validation cohort and specifically in individuals without high-risk HLA-DR3/DR4 haplotypes — the population most poorly served by current tools. SHAP feature importance analysis of T1GRS scores additionally enables unsupervised clustering that reveals four genetically distinct T1D subtypes with different ages of onset and dramatically different complication profiles, opening a path to blood-based genetic endotyping of T1D that does not require pancreatic tissue.
🌟 Why This Talk Matters to T1D Scientists and Clinicians
For scientists: The 154 nonlinear interactions T1GRS reveals represent a genuinely new layer of T1D genetic architecture. The most striking is an interaction between the INS insulin gene locus and the HLA-DQB1 position 57 amino acid polymorphism — the strongest known T1D risk signal in the MHC — suggesting that the same insulin gene variant influencing central thymic tolerance interacts with the specific HLA molecule presenting the antigen, compounding risk in ways that additive models simply cannot detect. The gradient boosting framework and SHAP analysis that reveal this interaction also produce individual-level feature importance “fingerprints” that can be clustered to distinguish patient subtypes. The model is fully containerized, uses imputed genotype data compatible with any standard array, leverages the TOPMed and Michigan HLA imputation reference panels, and is deposited on GitHub — meaning the genetic architecture discoveries embedded in T1GRS are immediately accessible for any group to validate, extend, or apply in their own cohort. The finding that the pancreas-enriched genetic cluster has a higher burden of nephropathy, neuropathy, and cardiovascular complications independent of HbA1c is a concrete scientific hypothesis replication-ready in any cohort with genetic and clinical outcome data.
For clinicians: Three implications are immediately actionable. First, T1GRS meaningfully improves identification of at-risk individuals who would be missed by current tools — specifically those without DR3/DR4, who represent a large fraction of adult-onset T1D cases and are the hardest to catch before clinical onset. With teplizumab FDA-approved and able to delay clinical T1D onset by a median of two years, finding these individuals earlier is not theoretical — it is a treatment opportunity. Second, T1GRS performs comparably to ancestry-specific scores in African American populations despite being trained on European ancestry data, and the non-MHC model alone (AUC 0.803 vs. GRS2’s 0.692) demonstrates the untapped signal in genome-wide variants relevant across ancestries. Third, the four-cluster genetic subtyping framework offers what the T1D field has long needed but lacked: a clinically scalable endotyping tool. Prior endotype work required pancreatic donor tissue from nPOD — invaluable mechanistically but inaccessible in living patients. T1GRS-based cluster assignment requires only a genotype array and can be performed at birth, diagnosis, or any point in adult life.
3️⃣ Big Takeaways
1. T1GRS outperforms GRS2 across every tested population — and the gains are largest precisely where the clinical need is greatest: individuals without classical HLA risk who current tools most often miss. The gradient boosting framework at the heart of T1GRS is not simply a bigger additive model. It learns which combinations of genetic features, taken together, best discriminate T1D from non-T1D — and it is allowed to weight those combinations nonlinearly. In the discovery cohort (trained data), T1GRS achieved AUC 0.937 versus GRS2’s 0.916. In the All of Us population-based validation cohort — a real-world, diverse dataset not used for training — T1GRS showed consistent superiority. Most importantly, the improvement was sharpest in individuals with complex non-DR3/DR4 genetic risk profiles: exactly the individuals who remain undetected by HLA-centric screening. The non-MHC model alone — trained only on genome-wide variants outside the HLA region — achieved AUC 0.803 versus GRS2’s 0.692, a substantial gap that underscores how much predictive signal has been left on the table by HLA-dominant approaches. T1GRS also performed comparably to an African American-specific GRS in that population, a preliminary signal of cross-ancestry portability that the field will need to formally validate but that is immediately encouraging.
2. 154 nonlinear genetic interactions — including a striking INS × HLA-DQB1 interaction — reveal that T1D risk is not an additive sum of independent variants but an interacting network, with implications for mechanism, therapy, and pharmacogenetics. Additive genetic models assume that the effect of carrying a risk allele at the INS locus is the same regardless of what allele you carry at HLA-DQB1. T1GRS demonstrated this assumption is wrong: 154 statistically significant nonlinear interactions exist between MHC and non-MHC loci, and the strongest involves exactly that INS × HLA-DQB1 pair. The biological interpretation is compelling: the insulin gene variant affects how much proinsulin the beta cell produces and presents in the thymus, influencing central tolerance. The HLA-DQB1 polymorphism at position 57 determines what peptides the MHC class II molecule binds and presents to T cells. Together, a risk allele at INS interacting with a high-risk DQB1 configuration may create a compounded vulnerability — more antigen, better presented, to a system primed for autoimmunity — that neither variant produces in isolation. These interactions also explain why T1GRS particularly outperforms additive models in individuals with “higher complexity” genetic profiles: the more interactions at play in a given individual’s genome, the more GRS2 underestimates their true risk, and the more T1GRS corrects for it.
3. Four genetic subtypes of T1D — MHC-driven, MHC-enriched, T cell-enriched, and pancreas-enriched — differ meaningfully in age of onset and complication burden, and can be assigned from genotype data alone, without pancreatic tissue. When SHAP feature importance values were calculated for every individual in the training cohort and subjected to dimensionality reduction and clustering, four distinct genetic “fingerprint” patterns emerged — and they mapped onto biologically meaningful cell types. The MHC-driven and MHC-enriched clusters weight classical HLA DR3/DR4 risk most heavily and tend toward earlier disease onset. The T cell-enriched cluster is dominated by PTPN22 and maps predominantly to CD8+ T cell biology. Most intriguingly, the pancreas-enriched cluster — dominated by INS and beta cell loci, with enrichment for T2D-associated loci and a partial protective MHC signal — has later disease onset but a substantially higher burden of nephropathy (1.29–1.44-fold), neuropathy (1.35–1.44-fold), and cardiovascular disease (1.46–2.34-fold) relative to other clusters. This complication pattern replicated in the All of Us validation cohort, suggesting that a genetically distinct subset of T1D patients faces disproportionate long-term complications — not because of worse glycemic control, but because of their underlying biology. If this subtype can be reliably identified from genotype data at birth or diagnosis, it would immediately inform surveillance intensity, screening recommendations, and eventually subtype-specific therapeutic targeting.
❓ Key Questions from the Discussion
Can T1GRS predict the sequence of autoantibody appearance — IAA-first vs. GADA-first — as distinct phenotypes, and therefore identify who is moving toward stage one and stage three disease? Åke Lernmark raised this directly in the Q&A, pointing to the known phenomenon that IAA-first seroconversion is associated with younger onset and higher HLA risk, while GADA-first appearance tracks with older onset and different genetic architecture. Dr. McGrail acknowledged that T1GRS’s richer genetic signal — particularly its non-MHC component — should in theory track these differences better than HLA-only tools, but confirmed this question has not yet been formally answered. Accessing the longitudinal autoantibody onset data from cohorts like TEDDY, DAISY, and TrialNet with paired genetic data is the next essential step, and the team is in active conversations with collaborators toward this.
Can the genetic cluster assignment predict which patients will respond best to teplizumab, ATG, or other immune-modifying therapies — and is this the path to pharmacogenetics in T1D prevention? This is the question that visibly excited both speakers most. The T cell-enriched cluster — dominated by PTPN22 and CD8+ T cell biology — may respond differently to teplizumab (which works by targeting CD3 on T cells) than the pancreas-enriched cluster, which has more beta cell-intrinsic biology and less canonical autoimmune signal. Alex Ghaben confirmed this is an active line of inquiry: the team is seeking pharmacologic outcome data from teplizumab and ATG trials that can be linked to T1GRS cluster assignments. The framework is in place — T1GRS is portable, runs on any standard genotype array, and is available via GitHub. The bottleneck is access to paired genetic and treatment response data, which the T1D clinical trial community is in a position to provide.
Do the CLNK mast cell signal and the ZMIZ1 locus — two of the eight newly discovered risk loci — point to mast cells or exocrine pancreatic cells as underappreciated players in T1D pathogenesis? Monica raised the CLNK signal as a particular surprise — CLNK (also called MIST) is an adapter protein in the SLP-76 family that regulates FcεRI-mediated mast cell degranulation, and it has been largely absent from T1D biology discussions for decades. The question is whether the CLNK risk variant makes mast cells more trigger-happy (lower degranulation threshold), or whether it impairs a Treg-supporting function of mast cells, preventing them from dampening autoreactive responses. Dr. McGrail was appropriately cautious — the locus has a smaller effect size than HLA or INS — but noted that the four-subtype framework raises the possibility of a mast cell-enriched subpopulation of T1D patients in whom this biology is more prominent, potentially identifiable from genotype data and relevant to specific therapeutic strategies. ZMIZ1 — which has roles in both beta cell function and T2D — strengthens the case that the pancreas-enriched cluster has genuine beta cell-intrinsic vulnerability worth investigating in nPOD and HPAP tissue.
Do the genetic clusters map onto distinct tissue pathology in nPOD or HPAP donors — and could T1GRS feature weights serve as priors to guide where the MAI multimodal AI T1D consortium should focus single-cell and spatial transcriptomic attention in pancreatic tissue? Todd Brusko raised this in the chat, and both speakers endorsed the connection enthusiastically. HPAP (Human Pancreas Analysis Program) has pancreatic tissue from donors across T1D disease stages with genetic data; nPOD has a larger tissue collection with varying degrees of available genotyping. Mapping cluster assignment onto insulitis grade, beta cell mass, CD8+ T cell infiltration patterns, and acinar cell pathology in these collections would test whether the genetic fingerprints T1GRS identifies correspond to histologically distinguishable forms of the disease. The MAI consortium’s use of single-cell and spatial transcriptomic data provides exactly the resolution needed to ask whether CLNK expression in mast cells or ZMIZ1 expression in acinar cells follows the pattern the genetic clusters predict. T1GRS feature weights as priors for which cell types to focus on in each cluster is a particularly tractable computational proposal.
🔗 3 TSS Talks That Connect With This One
1. Panel: ATG — Clinical History, Mechanistic Insights, and Future Directions for the Treatment of T1D ▶️ Read the TSS pregame & resources → The ATG panel — featuring Todd Brusko, Mike Haller, and Laura Jacobsen — addresses the exact pharmacogenetics question raised in today’s Q&A. ATG, like teplizumab, works by depleting or modulating T cells, and the panel examines in detail what distinguishes responders from non-responders in these T cell-targeting prevention trials. The T cell-enriched T1GRS cluster — dominated by PTPN22 and CD8 biology — is precisely the patient subtype most likely to respond differently to T cell-targeting immunotherapy than the pancreas-enriched or MHC-driven clusters. Understanding the immunological mechanism ATG attempts to restore is essential context for designing the cluster-stratified pharmacogenetic studies that T1GRS now makes possible. Brusko also raised the nPOD/HPAP tissue question in today’s Q&A — his work on tissue immunopathology in T1D is the natural experimental partner for genetic cluster validation.
2. What Would the Earliest Detection of Type 1 Diabetes Look Like? — Drs. Ahrens, Triplett, Ludvigsson & Oresic ▶️ TSS Substack page → This multi-speaker panel frames the scientific and clinical landscape that T1GRS enters. Its central question — what would detecting T1D look like before autoantibodies, before any clinical sign, using biology rather than pathology as the signal — is precisely the question T1GRS addresses at the genetic level. T1GRS is static: it can be calculated at birth from a dried blood spot and never changes. Combined with the environmental, metabolomic, and microbiome signals the panelists discuss, genetic subtyping via T1GRS represents the earliest possible layer of a multi-modal early detection strategy. The panel’s discussion of what intervention would be warranted at Stage 0 — before any autoantibodies — maps directly onto the T1GRS finding that individuals with complex non-HLA risk profiles can be identified genetically before any immune signal appears.
3. Ask the Expert: Kyle Gaulton, PhD — UCSD ▶️ Listen on TheSugarScience Podcast → Dr. Kyle Gaulton is the senior author of today’s paper and Carolyn McGrail’s PhD mentor. His earlier TSS appearance covers the foundational T1D genetics and single-cell epigenomics program at UCSD — including the landmark Chiou et al. Nature 2021 paper linking T1D GWAS signals to cell-type-specific regulatory elements in islets and immune cells — that T1GRS builds directly upon. Understanding Gaulton’s prior work on non-coding variant interpretation and islet chromatin accessibility is essential context for appreciating how T1GRS connects genetic risk signals to the specific cell types and regulatory programs that drive T1D. The CLNK mast cell signal, the acinar cell enrichment in cluster 3, and the INS × HLA interaction all flow from the analytical infrastructure Gaulton’s lab has been building for a decade.
🔔 Subscribe to The Sugar Science newsletter to get the latest T1D research drops straight to your inbox.