NNdisulfide

Simple feed-forward neural network to predict candidate artificial disulfide bonds.

Given a PDB or mmCIF structure, NNdisulfide scans nearby residue pairs and ranks pairs that may be suitable for mutation to cysteine to create an engineered disulfide bond.

You can use the included ss_model.pt directly for prediction. It was trained on the same mmCIF files used as part of the AF3 database, including newer files up to 5/24/25.

Quick start

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

Run prediction with the included model:

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

For PDB input:

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

The output CSV contains:

Column	Description
chain1	Chain ID for residue 1
res1	Residue number for residue 1
chain2	Chain ID for residue 2
res2	Residue number for residue 2
prob	Model score for the candidate disulfide pair

Commands

NNdisulfide has three CLI subcommands:

Command	What it does
predict	Scan a PDB/mmCIF file and rank candidate artificial disulfide bonds
build	Parse a directory of structures and build a training CSV
train	Train a new PyTorch model from the CSV

Predict

python NNdisulfide.py predict \
  --model ss_model.pt \
  --structure input.cif \
  --cutoff 8.0 \
  --top_k 25 \
  --out predictions.csv

Arguments:

Argument	Description
--model	Path to a trained .pt model file
--structure	Input .pdb, .cif, or .mmcif structure
--cutoff	CB–CB distance cutoff for candidate pairs, default 8.0 Å
--top_k	Number of top predictions to write
--out	Output CSV file

Build training data

python NNdisulfide.py build \
  --data_dir /data/pdb-mmCIF \
  --out_csv disulfides.csv \
  --nproc 32

The build command searches recursively for:

• .cif
• .mmcif
• .cif.gz
• .mmcif.gz
• .pdb

It extracts positive cysteine-pair examples from annotated disulfide bonds and close SG–SG geometry, then generates nearby negative examples.

Train

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

Training uses PyTorch with Adam and binary cross-entropy with logits. The best validation checkpoint is saved to --model.

Model details

NNdisulfide does not feed the whole protein into the neural network. Instead, it converts each structure into residue-pair candidates. Each candidate pair is represented by 16 numeric features, and the model scores each pair independently.

Current network architecture:

16 input features → 32 hidden units → 16 hidden units → 1 output logit

The 16 features are:

ca_dist, cb_dist, seq_sep, same_chain,
phi1, psi1, chi1_1, ang1, asa1, b1,
phi2, psi2, chi1_2, ang2, asa2, b2

Distances and sequence separation are transformed with log1p; backbone and side-chain angles are scaled by /180.

Notes

• Predictions are ranked candidates for further inspection, modeling, or experimental validation.
• The model is a pairwise geometric classifier, not a full protein-design model.
• Only the first model in a multi-model structure is used.
• Missing atoms or unusual residue numbering may affect results.
• Glycine uses CA as a fallback for CB.

License

This project is licensed under the MIT License.

Permission is granted to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of this software, subject to preservation of this notice and the standard MIT License warranty disclaimer.