Simple Feed Forward Neural Network to Predict Artificial Disulfide Bonds.

Given a PDB or CIF structure, predict which 2 residues could be mutated to create an artificial disulfide bond.

You can install and predict directly (skipping build/train) by supplying the included ss_model.pt, which was trained on the same mmcif files that are used as part of the AF3 database (but includes newer files up to 5/24/25).

1 | What the script does

build

Parse every .mmCIF in a directory, pull out all annotated disulfide bonds (inter- & intra-chain) and generate matching negative examples. Uses gemmi for blazing-fast CIF access and multiprocessing to scale across cores.

train

Learn to score cysteine-pair compatibility. A lightweight PyTorch feed-forward network (3 numeric features → 16→8→1). Default training loop + early model checkpointing.

predict

Scan a new structure, enumerate residue pairs close enough to fuse, and rank them by probability they could form a disulfide once both are cysteines. CA-distance cutoff defaults to 8 Å; adjust with --cutoff. Results land in predictions.csv (chain, residue numbers, prob).

All three stages are wrapped as CLI sub-commands, so one file drives the whole pipeline.

2 | Installation:

git clone https://github.com/linuxfold/NNdisulfide
cd NNdisulfide

conda create -n ssbond python=3.11

conda activate ssbond

pip install gemmi torch pandas numpy tqdm scikit-learn

3 | Build & Extract Data

  python NNdisulfide.py build \
  --data_dir /data/pdb-mmCIF \
  --out_csv disulfides.csv \
  --nproc 32        # adapt to your CPU budget

4 | Train

  python NNdisulfide.py train \
  --dataset disulfides.csv \
  --model ss_model.pt \
  --epochs 1000

5 | Predict

  python NNdisulfide.py predict \
  --model ss_model.pt \
  --structure my_enzyme.cif \
  --top_k 25 \
  --out my_enzyme_ss_predictions.csv