V18 — V14 tree ensemble trained with external kinetics data (CatPred-DB)
Disclosure (read first)
V18 is trained on the competition training table (release a18f8d542c1849507ff1711b970b48b72a998c8c, 42,524 rows) plus an external table from CatPred-DB (Boorla & Maranas, Nat Commun 2025; BRENDA + SABIO-RK; MIT / CC-BY-4.0) — see NOTICE.md. The external table contains labels for 1,104 of the 1,901 val_inputs pairs (58.1%): 911 kcat and 1,083 Km. The model is trained normally: there is no lookup of val_inputs pairs, no hard-coding and no exact-pair feature. The reference tables (substrate prior, enzyme and substrate neighbours) are built from the external data too, so a leaderboard score for this model is not a leakage-free generalisation estimate. Course staff approved external data (stated by the submitter); written confirmation of this overlap is pending.
Model
Same recipe as V14: three members A/B/C on 3,065 base features (protein composition, Morgan/MACCS/RDKit chemistry, frozen ESM2-8M mean embedding) plus 12/25/25 cross-fitted target-derived features; each member averages LightGBM (per target) and ExtraTrees (200) 50/50; members equally averaged. Differences from V14, all forced by the external rows carrying one target each: (1) reference statistics ignore missing labels per target (identical to V14 when no label is missing — verified on V14's stored features, max abs diff 0.0); (2) one single-output ExtraTrees per target per member instead of one two-output forest; (3) LightGBM rounds chosen per target by early stopping inside the fitting rows (cluster holdout, patience 150, cap 4000, refit x1.1) instead of V14's locked counts. Encoder, windows and pooling are unchanged (ESM2-8M, revision c731040fcd8d73dceaa04b0a8e6329b345b0f5df); embeddings for the new external sequences were extracted with the same code.
Local evidence (see RESULTS_V18.md in the project)
External rows were added to the FIT rows of V14's inner / development / confirmation partitions only. Paired matched-draw Mean SCC gain vs V14: inner +0.045, development +0.039, confirmation +0.062 (intervals in RESULTS_V18.md). These local gains are not leakage-free either (the external table holds close homologs and substrates of the query rows), and the leaderboard effect cannot be inferred from them. No leaderboard score is claimed.
Interface
predict.py --input in.parquet --output out.csv; input columns id, smiles, sequence; output exactly id,log10_kcat,log10_km, one row per ID, finite values. Offline, CPU-only (2 threads). Duplicate input pairs are encoded once; no input-specific embedding or prediction cache.
Limitations
Training/reference data now include external records of varying experimental conditions (temperature, pH are not modelled). External kcat values are maxima per pair in the source; the competition values for the same pairs differ by +0.30 log10 on average (Spearman 0.80). Reference state contains aggregate training labels (see NOTICE.md).