GPN-Star maps genomic constraints with 200M parameters

Researchers published GPN-Star, a 200-million-parameter genomic language model trained for several days on 8 NVIDIA A100 GPUs, in Nature’s September 9 paper. It uses multispecies alignments and phylogenetic information to improve coding and non-coding variant-effect prediction, while the authors’ open repository publishes model and score resources that still require aligned genomes at inference time. Bioinformatics teams should treat it as a focused evaluation candidate rather than a clinical oracle, and compare its held-out results with AlphaGenome’s broader regulatory-genomics approach.