gpn-star

is a Data Source.

GPN-Star is a phylogeny-aware genomic language model from the Song lab at UC Berkeley that predicts functional constraint and variant effects across the genome from multispecies whole-genome alignments, using a phylogeny-informed attention design. It supersedes GPN-MSA. Three human (hg38) models were trained on vertebrate (100-way), mammalian (447-way) and primate (243-way) alignments, and further models cover mouse, chicken, Drosophila melanogaster, Caenorhabditis elegans and Arabidopsis thaliana. The authors distribute the trained models on Hugging Face (MIT license), genome-wide mutation-rate-calibrated constraint (entropy) and variant (log-likelihood ratio) scores as Parquet and BigWig files on Hugging Face (Apache 2.0), a multi-assembly UCSC Genome Browser track hub, and the gpn Python package for inference and training.

Product Summary

Products

From this Resource
ID Name URL Category Format Description
gpn-star.scores GPN-Star Genome-wide Scores gpn-star-scores Product parquet GPN-Star genome-wide scores on Huggin...
gpn-star.ucsc-hub GPN-Star UCSC Track Hub hgTracks?hubUrl=https%3A%2F%2Fhuggingface.co%2Fdatasets%2Fsonglab%2Fgpn-star-scores%2Fresolve%2Fmain%2Fucsc%2Fhub.txt Graphical Interface http Multi-assembly UCSC Genome Browser tr...
gpn-star.models GPN-Star Models gpn-star-68c0c055acc2ee51d5c4f129 Product mixed Hugging Face collection of trained GP...
gpn-star.software GPN Software gpn Process python The gpn Python package and repository...
gpn-star.docs GPN-Star Documentation gpn-star Documentation http GPN documentation pages for GPN-Star,...
From other Resources
ID Name URL Category Format Relation Description
biobtree.api BioBTree REST API api Programming Interface http had primary source REST API for searching identifiers an...

Details

GPN-Star

GPN-Star is a genomic language model developed in Yun Song’s lab at the University of California, Berkeley. It learns from whole-genome alignments of many species and uses the species phylogeny in its attention design, so it can weigh evidence from closely and distantly related species. It predicts how constrained each position in a genome is and how damaging a nucleotide change is likely to be. GPN-Star supersedes the earlier GPN-MSA model.

Human models were trained on alignments of 100 vertebrates, 447 mammals and 243 primates. Additional models cover mouse, chicken, Drosophila melanogaster, Caenorhabditis elegans and Arabidopsis thaliana.

Access

  • Precomputed scores: genome-wide, mutation-rate-calibrated constraint and variant scores for eight score sets, as Parquet and BigWig files in the songlab/gpn-star-scores Hugging Face dataset (Apache 2.0).
  • UCSC track hub: GPN-Star tracks for all supported assemblies in the UCSC Genome Browser.
  • Models: trained checkpoints in the songlab Hugging Face organization (MIT).
  • Software: the gpn Python package on GitHub and PyPI (MIT).

Is this information incorrect or incomplete? Request an update.

Created: November 12, 2026 | Last modified: November 12, 2026