arnavthorat/submission11

Multimodal kcat–Km regression

Predicts log10_kcat and log10_km from protein sequence and substrate SMILES.

• Features: frozen ESM-2 35M (first 510 residues, mean-pooled), frozen MoLFormer, composition features and Morgan fingerprints with RDKit descriptors.
• Heads: TabM, two small fusion networks and LightGBM, trained on the official training set. The kcat heads also use upstream kcat training rows whose enzymes are in, or similar to, official training enzymes.
• Output: per target, an equal average of the heads' percentile ranks, moved part of the way toward the average for rows that share the same protein sequence, then mapped onto the training label distribution. recipe.json lists the heads.

Run python /app/predict.py --input INPUT --output OUTPUT. Input columns: id, sequence, smiles. Output columns: id, log10_kcat, log10_km. All weights are local; inference runs on CPU without network access. Temporary Hugging Face module files are written under /tmp.

Pretrained sources: Meta ESM-2 (MIT) and IBM MoLFormer (Apache-2.0); their original model cards are included with their weights. This repository contains model parameters and inference code, not raw training tables or prediction files.