GPN#
GPN, also known as GPN-SS, is a single-sequence genomic language model trained on unaligned genomes. The maintained package supports both inference and training on prepared datasets.
Quick start#
from gpn import register_auto_classes
from transformers import AutoModelForMaskedLM
register_auto_classes("ss")
model = AutoModelForMaskedLM.from_pretrained("songlab/gpn-brassicales")
The GPN tutorial demonstrates tokenization,
embeddings, masked nucleotide probabilities, and plots. File-backed inference is
available through gpn ss {vep,logits,embedding}; see the
command-line guide. The canonical prepared-data
training recipe lives in
recipes/gpn_training.
Published assets#
Brassicales model#
The GPN collection accompanies the PNAS paper.
Brassicales pretraining dataset: balanced sequences from Arabidopsis thaliana and seven other Brassicales.
songlab/gpn-brassicales: the published masked language model.songlab/tokenizer-dna-mlm: the seven-token DNA tokenizer used by the published model.Processed Arabidopsis data: supporting resources used by the published study.
Animal promoter model and TraitGym#
The TraitGym collection accompanies the causal regulatory variant benchmark.
Animal promoter pretraining dataset: promoter sequences used for GPN pretraining.
songlab/gpn-animal-promoter: the published promoter model.songlab/TraitGym: the benchmark datasets and model predictions.Training checkpoints: intermediate checkpoints from the published run.
TraitGym leaderboard: an interactive comparison of submitted predictions.
Sorghum gene-expression fine-tune#
The sorghum model is a fine-tuned application of GPN, not a separate model
family. It supports inference through the same register_auto_classes("ss")
path. The sorghum collection
accompanies the Nature Biotechnology paper.
songlab/gxa-sorghum-v1: sorghum gene-expression data from Gene Expression Atlas.songlab/gpn-brassicales-gxa-sorghum-v1: a gene-expression model fine-tuned fromgpn-brassicales.