Loading Curie Briefs...
Getting the latest healthcare news for you
Getting the latest healthcare news for you
UC Berkeley researchers have developed GPN-Star, a genomic AI model that outperforms competitors at identifying disease-linked genetic variants — and does it in a fraction of the time. Trained on whole-genome alignments across multiple species, the model can predict which genetic variants are most likely to influence inherited traits like cancer, heart disease, and schizophrenia. Its efficiency also makes it accessible enough for research teams worldwide to build upon.
Scientists at UC Berkeley have unveiled GPN-Star, a new AI-powered genomic language model that predicts the pathogenicity of genetic variants — essentially, which DNA changes are most likely to cause disease. Published in Nature, the model was trained on whole-genome alignments (WGAs) that compare DNA sequences across hundreds of species, helping it zero in on the parts of the genome that matter most for human health.
What sets GPN-Star apart isn't just accuracy — it's efficiency. While massive models like Evo 2 required 2,000 powerful processors and months of training, GPN-Star can be trained in days or even hours using just a handful of processors. The team trained it across multiple evolutionary timescales (primates, mammals, vertebrates), and found that each timescale excelled at predicting different types of variants — from rare protein-coding mutations to complex trait risks like schizophrenia.
Key Takeaways:
Why it matters: Most of the human genome remains poorly understood, yet non-coding regions may hold critical clues to diseases like cancer and autism. GPN-Star gives researchers a powerful, accessible tool to prioritize which genetic variants to study — potentially accelerating breakthroughs in precision medicine.
