KO-BCE / gKO-BCE — Conformational B-Cell Epitope Prediction

Released weights for the paper "gKO-BCE: A Novel Deep Learning Approach to Predict B-Cell Epitopes, Even Within Glycosylated Antigens."

Two checkpoints share one transformer architecture (EpitopeTransformer); they differ only in whether the glycosylation-proximity feature is used.

Files

Each model directory contains:

• best_model.pt — PyTorch state_dict for EpitopeTransformer
• normalizer_quantile.pkl — fitted QuantileTransformer (output_distribution='normal'), applied to all feature columns
• results.json — training config + model hyper-parameters
• info.json — n_num_features, n_classes

pca/ holds the fitted ESM-2 / ESM-1v PCA reducers (1280→256) needed to reproduce the input features.

Architecture (both models)

Tabular transformer: 3 layers, d_token 256, 32 heads, Tanh-GLU feed-forward, prenormalization, attenuated attention init (init_scale 0.01), token bias. Inputs per residue (1047 / 1049 dims): ESM-2 (PCA 256) + ESM-1v (PCA 256) + ESM-IF1 (512) + 20-d residue one-hot + B-factor + sequence length + RSA (+ glycosylation indicator + glycosylation-proximity distance for gKO-BCE).

Usage

Use the KO-BCE / gKO-BCE repository — the inference code downloads these weights automatically:

python inference/main.py --model gko-bce --input_pdb_dir path/to/pdbs

Citation

@article{odeyemi_gkobce,
  title  = {gKO-BCE: A Novel Deep Learning Approach to Predict B-Cell Epitopes, Even Within Glycosylated Antigens},
  author = {Odeyemi, Jethro and Kashyap, Monika and Wilson, Heather L. and Khatooni, Zahed},
  year   = {2026}
}

License: MIT.