Genentech/borzoi-model

Borzoi Model (Linder et al. 2025)

Model Description

This repository contains the weights for the Borzoi model, a model designed to predict functional genomic tracks from genomic DNA sequences.

• Architecture: Convolutions followed by Transformer and U-net layers.
• Input: 524,288 bp of genomic DNA sequence.
• Output Resolution: 32 bp bins.
• Source: Linder, J. et al. Nature Genetics (2025)

Repository Content

The repository includes both full PyTorch Lightning checkpoints and raw state dictionaries for the human and mouse versions of the model. Each model has 4 independent replicates. Note that the weights are derived from the publication but the model has been converted into the PyTorch Lightning format used by gReLU (https://github.com/Genentech/gReLU).

File	Type	Description
human_rep0.ckpt	PyTorch Lightning	Full checkpoint including base model and human head, replicate 0.
human_rep1.ckpt	PyTorch Lightning	Full checkpoint including base model and human head, replicate 1.
human_rep2.ckpt	PyTorch Lightning	Full checkpoint including base model and human head, replicate 2.
human_rep3.ckpt	PyTorch Lightning	Full checkpoint including base model and human head, replicate 3.
mouse_rep0.ckpt	PyTorch Lightning	Full checkpoint including base model and mouse head, replicate 0.
mouse_rep1.ckpt	PyTorch Lightning	Full checkpoint including base model and mouse head, replicate 1.
mouse_rep2.ckpt	PyTorch Lightning	Full checkpoint including base model and mouse head, replicate 2.
mouse_rep3.ckpt	PyTorch Lightning	Full checkpoint including base model and mouse head, replicate 3.
human_state_dict_rep0.h5	HDF5	Weights-only state dictionary for the human model, replicate 0.
human_state_dict_rep1.h5	HDF5	Weights-only state dictionary for the human model, replicate 1.
human_state_dict_rep2.h5	HDF5	Weights-only state dictionary for the human model, replicate 2.
human_state_dict_rep3.h5	HDF5	Weights-only state dictionary for the human model, replicate 3.
mouse_state_dict_rep0.h5	HDF5	Weights-only state dictionary for the mouse model, replicate 0.
mouse_state_dict_rep1.h5	HDF5	Weights-only state dictionary for the mouse model, replicate 1.
mouse_state_dict_rep2.h5	HDF5	Weights-only state dictionary for the mouse model, replicate 2.
mouse_state_dict_rep3.h5	HDF5	Weights-only state dictionary for the mouse model, replicate 3.
save_wandb_borzoi_human.ipynb	Jupyter Notebook	Code used to create human checkpoints
save_wandb_borzoi_mouse.ipynb	Jupyter Notebook	Code used to create mouse checkpoints

Model Heads & Output Tracks

The uman and mouse .ckpt files for the same replicate utilize the same core transformer trunk but differ in their species-specific output heads.

Outputs

Human Head: 7,611 total tracks Mouse Head: 2,608 total tracks

Usage

The models are intended for use with the grelu library.

from grelu.lightning import LightningModel
from huggingface_hub import hf_hub_download

# Download the desired checkpoint
ckpt_path = hf_hub_download(
    repo_id="Genentech/borzoi-model", 
    filename="human_rep0.ckpt"
)

# Load the model
model = LightningModel.load_from_checkpoint(ckpt_path)
model.eval()