The T1GRS algorithm maps complex interactions among 199 genetic variants to identify vulnerable populations across diverse demographics. The model enables earlier diagnosis and targeted preventative therapies.
Researchers at the University of California San Diego have engineered T1GRS, an advanced machine learning model that leverages non-linear genomic interactions to predict Type 1 diabetes risk with unprecedented accuracy across diverse global populations.
Detailed in a comprehensive study published on April 30, 2026, in the journal Nature Genetics, the novel architecture moves beyond traditional linear scoring methodologies by meticulously mapping complex epigenetic and genomic relationships among 199 distinct genetic variants.
To construct the foundational machine learning training dataset, the research team processed high-resolution genomic information from more than 20,000 individuals of European ancestry formally diagnosed with the autoimmune disorder, alongside a massive control group of nearly 800,000 subjects without the condition. This extensive data parsing confirmed 79 previously established risk loci—specific physical locations of genes on chromosomes—while successfully isolating 13 novel loci directly associated with complex gene regulation, blood sugar control, and overarching adaptive immune function.
The investigators placed particular algorithmic emphasis on the major histocompatibility complex, a highly specific region located on chromosome 6 that contains the strongest known genetic associations with the disease and dictates critical immune responses. By analyzing targeted genomic data from over 29,000 individuals, the researchers uncovered several novel variants within this complex that strictly govern immune function and gene activation, providing a clearer picture of the disease’s underlying biological mechanisms.
By capturing the non-linear interplay between these 199 genome-wide and major histocompatibility complex variants, the T1GRS algorithm calculates a highly individualized genetic risk score with elevated precision compared to legacy diagnostic systems that previously struggled with complex genetic profiles. This advanced mathematical modeling ensures that the predictive tool maintains high accuracy across a significantly larger and more diverse set of individuals, rather than being limited solely to those with well-known high-risk variants.
“We were able to identify individuals who get diabetes but don’t have known high‐risk genetic regions at a much higher rate than the previous diagnostic,” said TJ Sears, a postdoctoral fellow in the lab of Hannah Carter, associate professor of medicine at the UC San Diego School of Medicine.
To rigorously validate the model’s generalizability and prevent computational overfitting, the research team tested the architecture against independent datasets sourced from the National Institutes of Health All of Us Research Program and the National Pancreatic Organ Donor biobank. While the overall predictive accuracy of the machine learning model was marginally reduced when applied to these smaller sample sizes, the system still predicted disease risk with an impressive 87 percent accuracy rate and successfully classified individuals from non-European populations into the correct genetic sub-types.
The model’s unique capacity to evaluate non-linear feature interactions allowed the team to stratify the patient population into four distinct clinical sub-types based on the specific genetic features most strongly influencing each individual’s generated risk score. The first cluster, identified as the major histocompatibility complex-driven group, is characterized predominantly by well-known high-risk genetic variants and typically correlates with early childhood disease onset, while a second major histocompatibility complex-enriched group exhibits slightly later onset driven by a complex mixture of variants located both inside and outside the chromosome 6 region.
The algorithm also isolated a T-cell-enriched group driven largely by non-major histocompatibility complex variants affecting the adaptive immune response, alongside a pancreas-enriched group primarily influenced by gene variants that directly impact insulin-producing beta cells. Despite experiencing a significantly later age of onset, individuals classified within this final pancreatic group face the highest rates of severe clinical complications, including irreversible nerve damage, chronic heart problems, and debilitating kidney disease.
Moving forward, the deployment of T1GRS as a widespread clinical screening tool could fundamentally alter the diagnostic timeline for autoimmune conditions by identifying vulnerable populations long before clinical symptoms manifest, thereby setting the stage for targeted preventative interventions. This expanded detection capability allows clinicians to closely monitor patients to reduce the risk of acute complications like diabetic ketoacidosis at diagnosis, while simultaneously identifying candidates eligible for preventative therapies such as teplizumab before the immune system systematically shuts down the body’s ability to manufacture insulin.
“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,” stated Carolyn McGrail, a former graduate student in the lab of Kyle J. Gaulton and current senior associate consultant at L.E.K. Consulting.
