V11: new-release ExtraTrees baseline
Trained from scratch using only the official new release starkAhmed43/leak-curber-kcat-km at full immutable revision a18f8d542c1849507ff1711b970b48b72a998c8c: 42,524 training rows. No labels or old-only records from the previous36,970-row release are added. No old weights, imputer, vocabulary, IDF or SVD is used. Public validation contains1,901 input-only rows and is used only for diagnostics/interface tests, never fitting or selection.
Selected model: ExtraTrees. Ridge is the unchanged56-feature starter with alpha10, training-fitted population mean/std and two direct logged outputs. The ExtraTrees comparison uses2,212 features: the verified2,148 starter/composition/simple-sequence/RDKit16-descriptor/missing-flag/Morgan2048 block plus64 character3-merTF-IDF/SVD values. Morgan radius2, chirality enabled. Original X/B/U are preserved. Descriptor finite medians, vocabulary/IDF and SVD are fitted on corresponding fitting rows only. TF-IDF uses sorted unique fitting sequences, character trigrams, lowercase=False, sublinearTF, L2, min_df2, max_features10000; SVD64 randomized,n_iter5,seed42. Forests use256trees,max_depth24,min_samples_split8,min_samples_leaf4,max_leaf_nodes4096,bootstrapFalse,seed42; max_features0.6 kcat and0.3 Km. Full target-specific parameters and schema are packaged when ExtraTrees is selected. No feature scaling for trees, target transformation, clipping, blends, substrate target statistics, retrieval or pretrained encoder.
Measured new-release local evaluation
| Partition | Model | kcat SCC | Km SCC | Mean SCC | Mean RMSE |
|---|---|---|---|---|---|
| Development | ridge | 0.343830 | 0.540274 | 0.442052 | 1.272389 |
| Development | ExtraTrees | 0.637912 | 0.793402 | 0.715657 | 0.996823 |
| Compound-disjoint diagnostic | ridge | 0.328689 | 0.536975 | 0.432832 | 1.243224 |
| Compound-disjoint diagnostic | ExtraTrees | 0.480582 | 0.667276 | 0.573929 | 1.137322 |
| Outer once-only | ridge | 0.320371 | 0.527853 | 0.424112 | 1.249337 |
| Outer once-only | ExtraTrees | 0.623915 | 0.778401 | 0.701158 | 0.974641 |
Mean SCC is the arithmetic mean of target Spearman correlations; Mean RMSE is the arithmetic mean of the two target RMSEs, not pooled RMSE. Complete MAE/R2, novelty/similarity subgroups, paired sequence-group bootstrap and runtime/resource evidence live in project development reports, excluded from deployment. Selection was frozen on grouped development results and the separately defined sequence-and-canonical-compound-disjoint diagnostic before one outer evaluation. Outer seed1101, development1102, diagnostic1103/1104. Exact-sequence grouping is not homology separation; triplet cosine is not aligned identity. Shared compounds alone do not invalidate the main sequence-generalization split. Small subgroup estimates are uncertain. No old/new-release score is presented as a matched comparison. Full-data fitting is not another holdout result; no V11 leaderboard score or rank exists.
Offline prediction
python /app/predict.py --input INPUT --output OUTPUT
Input Parquet columns exactly id,smiles,sequence; original strings and IDs are retained. Follow starter nonempty string/length/ID rules, with added file-global duplicate-ID rejection. Deduplicate computations only, never output rows; every required ID exactly once in original order. Output exactly id,log10_kcat,log10_km, direct logged predictions, finite abs<=1,000,000. All inference assets are local. No network/GPU needed. Two inference workers; Linuxamd64,16GiB,600seconds,CSV<=16MiB,repository<=4GiB. Actual exact-staged Docker evidence and complete deployment hashes are maintained outside this payload; never infer success from this README alone.
This is a classical machine-learning baseline. The official brief lists deep-learning methods; a classical leaderboard model alone is not asserted to satisfy all academic expectations. No report/video requirements are invented. Proposed future repository VedaVachan/kcat-km-v11, display ALL-HAIL-DL-11. No automatic upload, submission or nomination.