News‑Medical: AI predicts type 1 risk

Type 1 diabetes risk prediction just got a lot more interesting — and a lot more useful. A UC San Diego-led team built a machine-learning genetic model called T1GRS that does a better job than older risk scores at spotting who is likely to develop the disease. The bigger twist is that it does not just spit out one number. It also seems to sort people into biologically distinct subgroups with different ages of onset and complication patterns. ### What did they actually build? They built a new genetic risk score for type 1 diabetes, but not the usual add-up-the-variants version. T1GRS uses machine learning to combine signals across the genome, including interactions between genes that older models tend to miss. In plain English, it is trying to capture the fact that risk is not just one risky variant plus another — sometimes the combination matters more than either piece alone. ### Why was the old approach falling short? Older type 1 diabetes scores worked best in people carrying the classic high-risk HLA variants. That is useful, but incomplete. Many people who develop the disease do not fit that clean textbook pattern, especially when risk is spread across a more complicated mix of genes. The new model especially improved prediction in people with fewer of those obvious high-risk haplotypes and messier genetic profiles. ### How big was the study? Pretty big. The team ran genome-wide association and fine-mapping analyses in 20,355 people with type 1 diabetes and 797,363 people without diabetes of European ancestry. They also analyzed the major histocompatibility complex — the immune-system-heavy region that dominates type 1 genetic risk — in another 10,107 cases and 19,639 controls. From that, they identified 160 risk signals. ### What is the machine-learning part doing? The model learned patterns that standard linear scores flatten out. The paper says it uncovered 154 nonlinear interactions between MHC and non-MHC loci. That sounds technical, but the basic idea is simple — some genes change the meaning of other genes. Think of it less like adding points on a checklist and more like reading chords instead of single notes. ### What about the subtypes? This is the part that could end up mattering most. The researchers used the features driving each person’s T1GRS score to define four genetic subclusters. Those clusters were not just statistical curiosities. They showed significant differences in age of onset and in diabetic complications. That suggests type 1 diabetes may be genetically more heterogeneous than the one-label diagnosis implies. ### Does this change patient care now? Not immediately. This is still a genetics-driven prediction tool, not a standalone clinical test that replaces antibodies, glucose monitoring, or symptoms. But the timing matters. Type 1 screening is becoming more valuable because teplizumab can delay progression in some at-risk people, so better early risk sorting has real clinical value now in a way it did not a few years ago. ### Why would eye clinics care? Because earlier diagnosis changes the whole downstream timeline. Diabetic eye disease is usually thought of as a later complication, but the path to retina clinics starts long before vision symptoms. If clinicians can identify high-risk people earlier, they may be able to tighten follow-up, personalize monitoring, and possibly reduce the damage that accumulates. That is still an inference, but it is a plausible one. ### So what is the real takeaway? This is not “AI can diagnose diabetes from nowhere.” It is narrower and better than that. A large genetics study found more type 1 risk signals, then used machine learning to turn those signals into a sharper risk tool and a possible map of disease subtypes. Basically, the field is moving from “who is at risk?” to “what kind of type 1 risk is this?” — and that is a much more actionable question.