DeepBGC (BGC-Bench mirror)

This repository redistributes the DeepBGC weights released by the original authors, so that BGC-Bench can download them at a fixed tag. It is not an official release of the original authors, and the weights are not retrained. Licence: MIT, as upstream. Please cite the original paper (below). Intended use: frozen feature extraction in BGC-Bench; see the original paper for the model's training data and limitations.

Upstream

• Paper: Hannigan et al. (2019). A deep learning genome-mining strategy for biosynthetic gene cluster prediction. Nucleic Acids Research 47(18): e110. https://doi.org/10.1093/nar/gkz654
• Code: https://github.com/Merck/deepbgc
• Weights: GitHub release v0.1.0 (MIT), https://github.com/Merck/deepbgc/releases/download/v0.1.0/deepbgc.pkl

Format

The Keras detector was loaded with DeepBGC 0.1.12 / TensorFlow 2.15.1 and its arrays written to model.safetensors (float32): the forward and backward LSTM kernels, recurrent kernels and biases (Keras gate order i, f, c, o), the output head, and the Pfam2vec table (rows in the order of pfam2vec_vocab.json). A PyTorch re-implementation (hard sigmoid recurrent activation) reproduces the Keras hidden states to max |diff| 4.5e-7.

Model

Bidirectional LSTM (2 x 128 units) over Pfam2vec domain vectors plus two gene-border flags (102-d input). BGC-Bench uses the mean of the concatenated hidden states over the BGC's Pfam accessions (256-d), approximating the gene-border flags with the first and last token of each cluster.

Files

Use

git clone https://github.com/Nugkta/BGC-Bench && cd BGC-Bench
uv sync
uv run bgcbench run --encoder deepbgc

Citation

@article{deepbgc,
  title   = {A Deep Learning Genome-Mining Strategy for Biosynthetic Gene Cluster Prediction},
  author  = {Hannigan, Geoffrey D. and Prihoda, David and Palicka, Andrej and others},
  journal = {Nucleic Acids Research}, volume = {47}, number = {18}, pages = {e110}, year = {2019},
  doi     = {10.1093/nar/gkz654}
}