V12 — exploratory; failed local acceptance gates
This is an explicitly authorized exploratory comparison, not a confirmed improvement. The original failed decision, predictions and configuration remain unchanged. No V12 leaderboard score exists or is promised.
Model
- kcat: exact verified V11 full-training ExtraTrees component and its original preprocessing. Submitted V11 revision:
ea3e222f5c7057208b97c4e45fb0d78153dfee3d; deployment-file hashes matched the historical submitted files. - Km: the existing fixed CatBoost candidate, refitted on all 42,524 public training rows for exactly 1,021 iterations. CPU, depth 6, learning rate 0.05, L2 leaf regularization 10, seed 42. Iteration count came from the original inner protein-cluster validation by RMSE with patience 100; it was not reselected.
- Each target uses 2,212 features: the unchanged 2,148-feature sequence/molecular block plus 64 character-trigram TF-IDF/SVD features. Morgan fingerprints: radius 2, 2,048 bits, chirality enabled. Descriptor imputation uses fitting-row finite medians from float64 RDKit values. TF-IDF: case preserving, sublinear TF, L2, min_df 2, maximum vocabulary 10,000; SVD: 64 components, 5 iterations, seed 42.
- Km imputation, vocabulary, IDF and SVD were fitted on full public training only. Kcat keeps its own byte-identical V11 preprocessing. The two preprocessing branches are packaged separately.
- No new model search, routing, blend, target transformation, calibration, clipping, substrate statistics, encoder or retrieval.
Official data release: a18f8d542c1849507ff1711b970b48b72a998c8c. Only its training labels are used. Public input file has 1,901 rows and is used solely for coverage/runtime/equivalence checks, never fitting. Original IDs, SMILES and sequences, including X/B/U, are preserved.
Original local evidence — failed gate retained
| Canonical compound group | Model | Rows | Km SCC | RMSE | MAE |
|---|---|---|---|---|---|
| overall | ExtraTrees | 7955 | 0.626943 | 0.966191 | 0.743884 |
| overall | CatBoost | 7955 | 0.599754 | 0.989761 | 0.772189 |
| seen_compound | ExtraTrees | 6213 | 0.639978 | 0.948931 | 0.731000 |
| seen_compound | CatBoost | 6213 | 0.606760 | 0.976129 | 0.760653 |
| unseen_compound | ExtraTrees | 1742 | 0.585813 | 1.025386 | 0.789834 |
| unseen_compound | CatBoost | 1742 | 0.579219 | 1.036921 | 0.813335 |
Overall Km SCC change: −0.027189, paired protein-cluster-bootstrap 95% interval [−0.047564, −0.008452]. Overall RMSE increased 2.44% and MAE 3.81%, exceeding the original 1% guards. Unseen-compound SCC dropped 0.006594, within the allowed 0.01, but the other gates failed. Full refitting and deployment preparation proceed only by explicit user choice despite this result.
No absolute local Mean SCC is invented: no valid matched kcat predictions existed for the new cluster holdout, and kcat was not retrained for that comparison. Implied Mean SCC change was −0.013595, conditional on an unchanged kcat prediction. Local results are not leaderboard results. The V11 public score gap has no verified causal explanation here; target-wise public breakdown and evaluator revision were unavailable.
MMseqs2 clustered proteins at 80% identity and 80% coverage of both sequences; entire connected components were kept together. No qualifying cross-partition match was detected in the audit, but approximate search does not prove absence of relatives. The split is not homology-free, and compound novelty is not scaffold separation. Historical local results and failed guards are preserved outside this deployment payload. A full-training refit creates no new holdout score.
Actual full-training fit
Km preprocessing: 28.752 seconds. Km CatBoost fit: 69.908 seconds. Serialized Km model: 1,204,624 bytes. Fit-process lifetime peak working set: 1,107,824,640 bytes. These are measured local resources, not accuracy estimates.
Offline inference
/usr/local/bin/python /app/predict.py --input INPUT --output OUTPUT
Input Parquet columns must be exactly id,smiles,sequence. Starter nonempty-string, ID and length rules apply, plus file-global duplicate-ID rejection. Repeated inputs may share computations but remain separate output rows. Output CSV is exactly id,log10_kcat,log10_km, one row per original ID in original order, finite values with absolute value at most 1,000,000. Predictions remain logged values; no exponentiation.
All assets load locally. No GPU or internet is required. Inference uses two workers. Required limits: Linux amd64, 2 CPUs, 16 GiB RAM, 600 seconds, repository below 4 GiB and CSV below 16 MiB. Exact staged-package test evidence and complete SHA-256 manifests are kept in the project readiness record outside this payload. Development files, datasets, cluster data, caches, predictions, logs, environments and unused models are excluded.
CatBoost and ExtraTrees are classical ML; this package does not establish academic deep-learning compliance. Proposed public MODEL repository: VedaVachan/kcat-km-v12; display name: ALL-HAIL-DL-12. No automatic upload, submission or nomination changes.