borzoi-pytorch

The Borzoi model from Calico, but ported to Pytorch! Original implementation and weights are here.
We show that the Pytorch version produces the same predictions as the original implementation for all tested notebook examples.

We include a weight conversion script that ports Borzoi weights from TF-keras to Pytorch.

Installation

1. Clone the repo and cd
2. pip install -e .

Todo

• Test the model on more sequences to ensure equivalence to the original implementation.
• Support reverse complement and shift augmentation.
• Make model configurable, currently, almost everything is hardcoded.
• ...

Misc.

Enabling tf32 or bf16 and/or compiling with Pytorch 2.0 leads to a speed up (compared to the plain PT version).

References

[1] Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation
Johannes Linder, Divyanshi Srivastava, Han Yuan, Vikram Agarwal, David R. Kelley
bioRxiv 2023.08.30.555582; doi: https://doi.org/10.1101/2023.08.30.555582
[2] enformer-pytorch github, Phil Wang lucidrains