songlab-cal/gpn

What it solves

GPN (Genomic Pretrained Network) provides a suite of genomic language models designed to predict the effects of genome-wide variants. It addresses the challenge of understanding how changes in DNA sequences across different species and evolutionary scales provide functional constraints on genetic mutations.

How it works

The project implements several model families based on different genomic data requirements:

• GPN: A base model that works with unaligned genomes.
• GPN-MSA: A model that utilizes multispecies alignment (MSA) for training and inference.
• PhyloGPN: A phylogenetic approach that uses alignment during training but does not require it for inference or fine-tuning.
• GPN-Star: The latest iteration which requires aligned genomes for both training and inference to predict functional constraints across evolutionary timescales.

Users can interact with the models via a Python API (integrating with Hugging Face Transformers) or a command-line interface for tasks like variant effect prediction (VEP), extracting logits, and generating embeddings.

Who it’s for

Researchers in genomics, computational biology, and bioinformatics who need to predict the effects of genetic variants and analyze genomic sequences using pretrained language models.

Highlights

• Multiple Model Architectures: Offers a variety of models tailored to different data availability (aligned vs. unaligned genomes).
• Transformers Integration: GPN-Star can be loaded using standard AutoModelForMaskedLM classes from the Transformers library.
• Comprehensive Tooling: Includes a CLI for multi-GPU inference and training recipes for custom data.
• Broad Evolutionary Scale: Capable of predicting functional constraints across various evolutionary timescales.