V10 kcat-Km: exploratory; failed local acceptance gates
This package is a classical machine-learning comparison. Its full-data refit and deployment were explicitly authorized after its frozen local accuracy gates failed. It is not a confirmed improvement. No V10 leaderboard score is claimed. Confirmation was not run.
| Km development metric | Matched control | V10 |
|---|---|---|
| SCC, 7,955 rows | 0.626943 | 0.623946 |
| RMSE | 0.966191 | 0.970863 |
| MAE | 0.743884 | 0.744660 |
| Unseen-compound SCC, 1,742 rows | 0.585813 | 0.573956 |
The paired protein-cluster-bootstrap 95% interval for overall Km SCC change was [-0.013577, +0.007409]. The unseen-compound SCC decrease was 0.011858, exceeding the allowed 0.01. These failed decisions remain unchanged. The development split had previously been observed; it is not independent confirmation. MMseqs grouping is not proof of homology separation. Local results are not leaderboard results.
Architecture and assets
Kcat uses the unchanged V11 ExtraTrees forest, imputer, TF-IDF vectorizer, SVD and schema, copied without retraining. Its 2,212 features comprise the original 2,148 engineered sequence/molecular features plus 64 sequence 3-mer TF-IDF/SVD coordinates.
Km uses that same feature construction, with preprocessing newly fitted on all 42,524 rows of official release a18f8d542c1849507ff1711b970b48b72a998c8c, plus 167 RDKit MACCS positions and 2,048 atom-pair presence bits, giving 4,427 columns. MACCS includes reserved zero bit 0. Atom pairs use distances 1..30, topological 2D distances, chirality, 2,048 bits and no count simulation. Original sequence characters including X/B/U and original SMILES/IDs are retained.
Km forest settings are unchanged from the tested candidate: 256 trees, maximum depth 24, maximum leaf nodes 4096, minimum samples per leaf 4, minimum samples to split 8, max_features 0.3, squared-error criterion, bootstrap false, seed 42. Fractional feature sampling gives 1,328 sampled columns per split versus 663 for the 2,212-feature control. Full resolved settings and ordered names are in configuration.json. Full refitting took 1632.078 seconds of measured forest fitting, excluding checkpoint serialization. Full-refit results are not new holdout scores.
Both logged targets are predicted directly, without unit conversion, exponentiation, clipping, calibration, blending or target statistics. Empty strings are rejected. Invalid molecular parsing produces the original failure flag and descriptor imputation plus zero added fingerprint blocks; missing descriptors use fitted medians.
Offline inference
python /app/predict.py --input INPUT --output OUTPUT
Input Parquet columns must be exactly id,smiles,sequence. Output CSV columns are exactly id,log10_kcat,log10_km, with one prediction for each input row and its original ID. There is no row grouping or deduplication of predictions; exact input-string caches only avoid repeated feature computations. Duplicate IDs are rejected globally. All models and preprocessing load locally from separate weights/kcat/ and weights/km/ folders. No internet or GPU is required. No public-input prediction lookup tables, datasets or development caches are included.
Dependencies are pinned for Python 3.11.9/Linux amd64. Use two inference CPUs, 16 GiB memory and a 600-second timeout. The deployment-only package must remain below 4 GiB and output below 16 MiB. Measured Docker evidence is retained outside this upload package in the project development reports.
ExtraTrees is classical ML. This package does not establish satisfaction of the course's deep-learning academic requirements. See NOTICE.md for attribution.