VNVSH-DA24B034/aura

kcat–Km predictor (VNVSH-DA24B034)

Predicts log10_kcat and log10_km for an enzyme sequence + substrate SMILES.

Model

One LightGBM regressor per target on:

• Substrate (RDKit): 2048 Morgan count bits (radius 2), 167 MACCS keys, ~210 RDKit descriptors.
• Enzyme: length, hydropathy, amino-acid groups, amino-acid and dipeptide composition, composition of each third of the sequence. v2: plus mean-pooled ESM-2 embeddings (weights/esm/).
• Substrate prior: shrunk mean, count, std, max, min and median of each substrate's training labels (training rows use out-of-fold values, so no row sees its own label).

Validation during development: 5-fold CV grouped by enzyme clusters (3-mer cosine similarity > 0.3), which mimics the unseen-enzyme split of the leaderboard.

Files

• predict.py — grader entry point (python /app/predict.py --input X.parquet --output Y.csv)
• model.py, features.py — inference code; train.py — retrains weights/ from train.parquet
• weights/ — LightGBM models, substrate prior table, meta.json, optional esm/
• Dockerfile (v1, no ESM) and Dockerfile.esm (v2)