GEPBind

GEPBind is a hybrid antibody-antigen affinity predictor for (\Delta G) regression that combines:

• sequence representations from protein language models (winner uses ESM2), and
• graph-based structural encoding (winner uses GINE + Performer).

Included Contents

• Core training/evaluation code:
  • train_hybrid.py
  • esm2_embedder.py
  • src/
  • graphgps/
  • mamba/
• Reproducibility data assets (75% clustered split):
  • datasets/pairs_sabdab_clean_clustered75_noprune.csv
  • datasets/seq_natural.fasta
  • datasets/ABAG-DG/sabdab_clean/processed/data.pt
• Winner checkpoint and metadata:
  • checkpoints/win75_hybrid_s6/checkpoint_best.pt
  • checkpoints/win75_hybrid_s6/test_metrics.json
  • checkpoints/win75_hybrid_s6/hparams.json
• Curated results:
  • results/winner/
  • results/ablation_core/
  • results/ablation_graph_encoder/

Environment Setup

pip install -r requirements.txt
pip install -e mamba

Optional (only for antibody-specific PLM ablations):

pip install -r requirements-optional.txt

Quick Start

1) Evaluate shipped winner checkpoint

bash scripts/eval_winning_checkpoint.sh

2) Re-train winner configuration

bash scripts/run_winning_holdout_train.sh

3) Run seq-only and graph-only holdout ablations

bash scripts/run_seq_graph_ablation_holdout.sh

4) Run graph encoder ablation subset

bash scripts/run_graph_encoder_ablation_holdout.sh

Winner Configuration (Reference)

See configs/winning_holdout_s6.txt for the exact setting used in the reported winner.

Reported Metrics

• Winner holdout: RMSE 1.5693, Pearson 0.5752
• Seq-only holdout: RMSE 1.7586, Pearson 0.4423
• Graph-only holdout: RMSE 1.7445, Pearson 0.3707

Grouped 10-fold CV:

• Hybrid: RMSE 1.7259 +/- 0.3176, Pearson 0.3387 +/- 0.2116
• Seq-only: RMSE 1.7475 +/- 0.3224, Pearson 0.3255 +/- 0.2660
• Graph-only: RMSE 1.8008 +/- 0.2932, Pearson 0.2943 +/- 0.1275

Full summary: results/RESULTS_SUMMARY.md