New AI model for DNA learns from evolution to unlock secrets of the human genome
BiologyData ScienceMedicine
THE AI ANGLE
Predicting pathogenic and functional genetic variants across genomesResearchers at UC Berkeley developed GPN-Star, a genomic language model trained on whole-genome alignments that accurately identifies functional elements and predicts the pathogenicity of genetic variants. By leveraging evolutionary conservation across different timescales, the model outpaces competitors while drastically reducing training requirements from thousands of processors over months to a handful of processors over hours or days. This approach provides an accessible computational framework to prioritize experimentally testing noncoding and coding mutations linked to complex diseases like cancer and schizophrenia.
THE TEACHING ANGLE
Instructors can explore the pedagogical tension between brute-force scaling on raw, unaligned sequence data versus engineering models with biological priors, showing how evolutionary timescales dictate model efficacy for protein mutations versus complex noncoding traits.Read the original at phys.org Generate teaching or study materials
More in Biology
- AI’s best coding agent fails 60% of the time — and the data backs it upThe New Stack · September 15, 2026
- Report covers advances in surgical robots, deploying warehouse automation in the real worldThe Robot Report · September 15, 2026
- Observing Earth from orbit is unpredictable and messy: Could 'liquid' AI clear the view?Phys.org — Technology · September 15, 2026
- Can Culta take Japan’s premium fruit model global?AgFunderNews · September 15, 2026
- AI models enable cross-species mapping of cell biologyPhys.org — Biology · September 15, 2026