maozee/model_kcat_lgbm

Leak CURBER kcat–Km: LightGBM on ESM-2 + RDKit features

Predicts log10_kcat and log10_km for enzyme–substrate pairs (id, smiles, sequence).

Model

One LightGBM regressor per target on:

• Substrate: RDKit Morgan count fingerprint (radius 2, 1024 bits), MACCS keys, and RDKit 2D descriptors.
• Enzyme: mean-pooled ESM-2 (esm2_t6_8M_UR50D) final-layer embedding (first 1022 residues), log length, amino-acid and dipeptide composition.

Trained only on the provided train.parquet. Model selection used a homology-aware split: training sequences were clustered at 3-mer cosine similarity > 0.3, and whole clusters were held out, which matches the low similarity between the public validation and training proteins.

Inference

docker build --platform linux/amd64 -t kcat-km-lgbm .
docker run --rm --platform linux/amd64 --network none --cpus 2 --memory 16g \
  -v "$PWD/../data:/input:ro" -v "$PWD/output:/output" \
  kcat-km-lgbm python /app/predict.py --input /input/val_inputs.parquet --output /output/predictions.csv

Runs on CPU only and offline. ESM-2 runs in plain PyTorch (featurize.py) from weights/esm2_t6_8M.pt, so no transformers install is needed.