PhyloGPN

PhyloGPN learns position-specific parameters of the Felsenstein F81 substitution model from a phylogenetic alignment. Inference needs only DNA sequence, which makes the published model useful for transfer learning and zero-shot variant scoring away from a reference genome. The maintained package supports inference, not training or fine-tuning.

Quick start

from gpn import register_auto_classes
from transformers import AutoModel

register_auto_classes("phylo")

model = AutoModel.from_pretrained("songlab/PhyloGPN")

The PhyloGPN tutorial demonstrates rate parameters, nucleotide probabilities, and a zero-shot substitution score. The published model is songlab/PhyloGPN, described in A Phylogenetic Approach to Genomic Language Modeling.