YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

kcat / KM predictor β€” full-data version 5

python predict.py --input INPUT.parquet --output OUTPUT.csv (input columns id, smiles, sequence; output columns id, log10_kcat, log10_km)

How it works (all in model.py)

  1. Features: ESM2-35M mean embedding of the first 510 residues (ONNX, CPU) + log sequence length; Morgan fingerprint (1024 bits) + 9 RDKit descriptors of the substrate.
  2. Fallback model: one LightGBM per target and one small PyTorch network (features -> 256 -> 64 -> 2 outputs), averaged 50/50. Both are trained on the competition data plus public data (CatPred-DB, DLKcat at weight 0.5, EnzyExtractDB at weight 0.25), including public rows of the validation enzymes.
  3. Lookup table of published values (CatPred-DB, BRENDA 2026.1, DLKcat, EnzyExtractDB), matched on enzyme sequence + substrate InChIKey, or on the InChIKey skeleton (first 14 characters) when there is no exact match.
  4. Combining: KM uses the published value when found. For kcat, a small LightGBM combines the published values and the fallback prediction (this also handles SABIO-RK values stored per hour instead of per second). Pairs not in the table get the fallback prediction.

Compared with the full-data v4 model it has no large ensemble, no within-enzyme adjustment and no Swiss-Prot or Rhea features. Held-out folds: kcat 0.715 / 0.679, KM 0.828 / 0.803 (v4: 0.720 / 0.698, 0.836 / 0.811; without the network: 0.700 / 0.668, 0.817 / 0.802).

weights/: esm2_35m.onnx (ESM2-35M, Meta AI, MIT licence, ONNX export by Xenova), fallback_kcat.txt, fallback_km.txt, stacker_kcat.txt (LightGBM), net.pt (network weights and feature mean / std), lookup_table.parquet, config.json (training means).

Data sources: CatPred-DB, DLKcat, EnzyExtractDB, BRENDA 2026.1 (CC BY 4.0), SABIO-RK (via CatPred-DB), PubChem (substrate names), EnzymeMap (substrate names).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support