GPN-MSA

GPN-MSA is a genomic language model trained on a multispecies whole-genome alignment. It is deprecated in favor of GPN-Star: the maintained package supports inference from the published model, but not GPN-MSA training.

Quick start

from gpn import register_auto_classes
from transformers import AutoModelForMaskedLM

register_auto_classes("msa")
model = AutoModelForMaskedLM.from_pretrained("songlab/gpn-msa-sapiens")

File-backed inference uses gpn msa {vep,logits,embedding} and a compatible local Zarr alignment. The checkpoint and store must agree on target assembly, species count, species order, and preprocessing. See the command-line guide for the input contract.

Published assets

The GPN-MSA collection accompanies the Nature Biotechnology paper.

Alignment and training inputs

• songlab/multiz100way: processed 100-way vertebrate alignment in ZIP-backed Zarr stores.

• songlab/multiz100way-pigz: compressed 100-way vertebrate alignment archive.

• lpigou/89.zarr: the alignment representation used by the original workflow.

• songlab/gpn-msa-sapiens-dataset: prepared human training regions. The asset remains public, but dataset construction and GPN-MSA training are not maintained.

Model and evaluations

• songlab/gpn-msa-sapiens: the published human GPN-MSA checkpoint.

• songlab/clinvar: missense variants with clinical pathogenic or benign labels.

• songlab/cosmic: somatic missense mutations in cancer.

• songlab/omim: regulatory variants implicated in Mendelian disorders.

• songlab/gnomad: genome-wide variants with allele-frequency information.

Genome-wide predictions

• songlab/gpn-msa-hg38-scores: hg38 genome-wide variant scores.

• songlab/gpn-msa-hg38-gene-essentiality-scores: gene-essentiality predictions.

Historical analysis

The GPN-MSA paper analysis is preserved in the historical archive.