kcat / Km predictor β variant v4 (public-value lookup layer incl. BRENDA 2026.1)
CPU-only model for python /app/predict.py --input X.parquet --output Y.csv (columns id, log10_kcat, log10_km).
kcat: equal committee in the style of a reference design shared in class: LightGBM (L2 and L1), CatBoost, two fusion networks and a regional-attention network over ESM2-35M protein embeddings (first 510 residues, mean + 32 regions) and MoLFormer-XL / Morgan / RDKit substrate features. Members are averaged in label space.
Extra members (v12): a LightGBM classifier over 10 order-of-magnitude bins (expected value; both targets) and a substrate-only LightGBM (Km only).
Km: the same committee's Km output. One ESM2 pass (first 510 residues) and one MoLFormer pass serve both targets, so inference takes about 4 minutes on 2 CPUs. A small LightGBM model on hand-built features is kept as a fallback.
Training data: the competition training split, plus public rows from CatPred-DB (Boorla & Maranas 2025, CC BY 4.0) and DLKcat (Li et al. 2022) at half weight, and wild-type rows from EnzyExtractDB (Wei et al. 2025, literature-mined) at quarter weight. Every public row whose enzyme is a validation enzyme, or shares a 5-mer cluster with one (cosine > 0.08), was removed before training, so no validation answers were used.
Neural networks run in NumPy at inference; ESM2 and MoLFormer run through ONNX Runtime.
Leak guard (v13): the public kinetics databases may contain pairs from the hidden test split, which cannot be removed in advance because the test inputs are not public. At prediction time every input enzyme is compared with all public-data enzymes used in training (5-mer cosine > 0.08 = related). Rows of related enzymes are predicted by a committee trained on the competition training data only (weights/refc_safe), so no public label of that enzyme or a close relative reaches the output. On the validation inputs no enzyme is related (they were removed from the public data), so validation predictions are identical to v12.
Big protein LM members (v14): two extra LightGBM members whose protein block is not ESM2-35M but a larger protein language model, without running that model: per-protein embeddings are looked up by exact sequence in public Swiss-Prot embedding tables (ProtT5-XL-U50 from UniProt, CC BY 4.0; ESM-C 600M from Protein Dimension DB 2.0, pitagoras-alves on Hugging Face, MIT), reduced to 256 PCA dimensions. Enzymes not in Swiss-Prot get the embedding predicted from their ESM2-35M mean by a small network trained on ~385k Swiss-Prot proteins (no kinetics labels involved). Each of these members counts twice in the committee average.
Within-enzyme adjustment (v14): most kcat variance between unseen enzymes is at the enzyme level, and the committee orders the substrates of one enzyme weakly. For each enzyme in the input (its rows), predictions are pulled toward the enzyme's mean (kcat x0.5, Km x0.7) and a small within-enzyme LightGBM (trained on competition data only, label-free features of each substrate relative to the enzyme's other substrates) adds 0.2 x its centred deviation. Enzyme-level means are unchanged. Applied after the leak guard, to all rows.
All v14 choices were made on two held-out cluster folds of the competition training data.
Leak-guard committee upgrade (v15): the competition-only committee (weights/refc_safe) gains the same extra members as the main committee, all trained on competition data only: the bins classifier, the substrate-only member (Km) and the two big protein LM lookup members (weight 2). On two held-out folds this raised the competition-only path by +0.017 / +0.007 (kcat) and +0.015 / +0.004 (Km). Validation predictions are identical to v14 (no validation enzyme is related to public data); only hidden-test rows routed by the leak guard change.
Public-value lookup layer (v2): inputs whose exact (enzyme sequence, substrate InChIKey) pair has a published value in CatPred-DB (BRENDA/SABIO-RK), BRENDA 2026.1 (CC BY 4.0, wild-type entries with UniProt accessions), DLKcat or EnzyExtractDB get that value: Km directly (mean of raw values, as in the competition labels); kcat through a small LightGBM that combines the public value(s) with the v15 committee prediction (it learns, among other things, the x3600 unit slip in some SABIO-RK kcat entries). All other inputs keep the full-data committee. Data used: competition training split + CatPred-DB + BRENDA 2026.1 + DLKcat + EnzyExtractDB, with no leak filtering.
v4 changes: (1) Lookup matches also accept the same substrate up to stereochemistry and protonation (InChIKey skeleton block). These are used only when no exact (sequence, InChIKey) entry exists. On training rows these matches have Spearman 0.80 kcat / 0.93 Km to the labels (median error 0), and validation coverage rises by about 5%. (2) BRENDA substrate names missing from EnzymeMap's name table are mapped through PubChem PUG REST (name -> SMILES), adding 3,804 BRENDA entries. On two held-out folds (1) + (2) gave kcat +0.013 / +0.018 and Km +0.005 / +0.007. (3) The full-data committee gains the four annotation / reaction members of the v16 honest model, trained on all public data, and the within-enzyme kcat model with reaction-match columns.
Enzyme annotation and reaction members (from the v16 honest model): four extra LightGBM members, each with a label-free block that describes what the enzyme is and which reaction it runs:
- Annotation block (289 columns): the input enzyme's 10 nearest Swiss-Prot proteins in ESM2-35M space (public per-protein embeddings from Protein Dimension DB 2.0, MIT; PCA-128, int8). Similarity-weighted shares of their kingdom, top phyla, EC classes (levels 1-3) and Pfam families come from Swiss-Prot annotations (UniProt, CC BY 4.0).
- Reaction block (1033 columns): among the Rhea reactions (CC BY 4.0) of the 5 nearest Swiss-Prot proteins, the one whose molecules best match the input substrate (Morgan Tanimoto). It is oriented with the substrate on the left and encoded as a DRFP reaction fingerprint (1024 bits; DRFP by Probst, Schwaller & Reymond 2022, MIT, vendored in drfp_fp.py), plus the match quality, neighbour similarity, candidate count, EC of the reaction and molecule counts.
- Members and committee weights: reaction block replacing the ESM columns (x3), reaction block appended (x2), annotation block appended (x1), annotation block replacing the ESM columns (x1). Each exists in both the main and the leak-guard committee; the leak-guard versions are trained on competition data only.
- No kinetic values are involved in either block. On two held-out folds (committee level): kcat +0.001 / +0.007, Km +0.007 / +0.004. A production-pipeline check on fold 0: kcat +0.003, Km +0.006.
- The within-enzyme kcat model also gets the 9 reaction-match columns and is weighted 0.3: kcat +0.0020 / +0.0021.
PubChem (NCBI) name-to-structure service was used to map BRENDA substrate names (public domain data).