GPN-Star AI model decodes genome using evolutionary data

Researchers at UC Berkeley have developed GPN-Star, a new genomic language model designed to decode the human genome by learning from evolutionary history. Unlike standard models adapted from natural language processing, GPN-Star utilizes a phylogeny-aware architecture that incorporates species trees and whole-genome alignments to model evolutionary relationships explicitly.

The model demonstrates state-of-the-art performance in predicting the effects of genetic variants in both coding and non-coding regions. It is notably more computationally efficient than larger existing models, requiring significantly fewer resources for training. GPN-Star has shown particular strength in prioritizing pathogenic variants and identifying functional elements within the genome that contribute to complex traits and diseases.

Beyond human applications, the framework has been successfully trained on five other model organisms: Mus musculus, Gallus gallus, Drosophila melanogaster, Caenorhabditis elegans, and Arabidopsis thaliana, demonstrating its robustness and generalizability for broader biological discovery.