AI model eliminates ancestry bias in genetic testing, works without reference databases

A machine-learning system can now fill in missing genetic data from patients without needing large reference databases — and it works equally well across different ancestries, where older methods often fail. This matters because genetic tests are used to diagnose disease risk and guide medical treatment, but have historically been unreliable for non-European populations due to bias in the training data.

For decades, genetic imputation — filling in gaps in DNA test results — has depended on massive reference databases that are overwhelmingly composed of European ancestry samples. This creates a real medical problem: the same test becomes less accurate for patients of African, Asian, or other ancestry. GenoBERT solves this by learning patterns directly from a patient's own genetic data rather than comparing it to a reference panel, which means accuracy no longer degrades across ancestry groups. If this holds in clinical use, it removes a structural source of medical inequality that has been embedded in the tools themselves.