Yun Song, Professor of Electrical Engineering & Computer Sciences and Statistics, and Director of the Center for Computational Biology at UC Berkeley, led a team of researchers in developing GPN-Star, a new genomic language model that uses evolutionary patterns across species to identify genetic variants and functional elements in the human genome.
The researchers found that models trained across different evolutionary timescales were optimized for different types of genetic variants, with models trained on primate genomes performing particularly well at predicting variants associated with complex traits such as schizophrenia. The study was co-led by UC Berkeley graduate students Chengzhong Ye and Gonzalo Benegas, with additional contributions from Carlos Albors, Canal Li, Sebastian Prillo, Peter Fields, and Brian Clarke.