Genentech/gReLU

What it solves

gReLU tackles the difficulty of applying deep learning to genomics by offering a unified framework for building, training, interpreting, and using neural networks on DNA sequences. It simplifies tasks such as predicting gene regulatory elements, designing sequences, and understanding genomic features, reducing the need for custom code.

How it works

The library provides APIs for data preprocessing, model training, and model interpretation. It supports a range of architectures, including transformer models with flash attention, and includes a model zoo of pretrained models that can be loaded from Hugging Face. It also includes utilities for working with genome annotations (via UCSC tools) and generating bigWig coverage files for ATAC-seq or similar data.

Who it's for

This is aimed at bioinformaticians and researchers who use deep learning on DNA sequences, particularly in regulatory genomics, sequence-to-function prediction, and sequence design. It suits those who want a ready-to-use framework with pretrained models and standard preprocessing pipelines rather than starting from scratch.

Highlights

• End-to-end framework covering training, interpretation, and application of DNA sequence models.
• Model zoo with pretrained models hosted on Hugging Face, replacing the previous Weights & Biases integration.
• Support for transformer models and flash attention for efficient training.
• Integrates with standard genomics tools (UCSC utilities, bedtools) for annotations and coverage tracks.
• Published in Nature Methods and includes a DOI, indicating peer-reviewed quality.
• Active documentation and a contribution guide for community involvement.