V9 exploratory: frozen ESM2 8M features plus ExtraTrees
Kcat is the unchanged verified LOCAL V6 component with its exact preprocessing. Km uses selected candidateB: embedding-augmented ExtraTrees in original logged-target units. No V9 leaderboard score or rank exists; no specific score is promised. Historical remote V3/V4 source identity remains unresolved in older records; actual local V6 asset hashes/output equivalence are checked separately.
The unchanged2212 V6 features comprise the2148length/composition/simple-sequence/RDKit16descriptor/missingflag/Morgan2048block plus64 character3merTF-IDF/SVD features. Frozen ESM2 adds320final-layer residue-pooled values, total2532. All descriptor imputation,vocabulary/IDF andSVD are fitted only on the corresponding training subset. Final fits use36970publictraining rows only. No target statistics,retrieval,scaling,calibration,clipping,exponentiation or extra blend weights. ExtraTrees settings match V6 Km:256trees,max_depth24,min_samples_split8,min_samples_leaf4,max_features.3,max_leaf_nodes4096,bootstrapFalse,seed42,other verified defaults. Inference uses2threads/workers,noGPU/network.
Encoder: official facebook/esm2_t6_8M_UR50D, full revisionc731040fcd8d73dceaa04b0a8e6329b345b0f5df. Entire sequences<=512residues are encoded whole. Longer ones use first256 and last256 as TWO SEPARATE windows; mean final-layer residue states in each, excluding padding/special tokens, then mean of the two vectors. Do not concatenate termini into an artificial sequence. Middle residues are omitted and biologically useful domains/active-site context may be lost. X/B/U have real tokenizer IDs and are preserved. Original strings/IDs remain unchanged; exact sequences/SMILES are deduplicated only for computation then row alignment restored. Unknown characters map individually to unk/special and are excluded from pooling, zero vector if none. FP32,CPUbatch4,2threads; model frozen eval with no kinetic fine-tuning; last_hidden_state,notMLM logits.
Local evidence
| Stage | V6MeanSCC | SelectedMeanSCC | Pairedgroup95%gainCI |
|---|---|---|---|
| Repeatedly observed primaryseed43 | 0.608434 | 0.619658 | [+0.007170,+0.015768] |
| Prospective once-only confirmationseed53 | 0.633564 | 0.637919 | [+0.000422,+0.008391] |
Only B and its fixed50/50V6 blend were compared; choice frozen before confirmation. The original primary gates passed, but confirmation FAILED the unseen-compound MeanSCC and KmSCC guards (KmSCC0.480874 to0.450885). Full refit and packaging proceed by explicit user choice as an EXPLORATORY submission. This is NOT a confirmed improvement; the historical failed-gate decisions are unchanged. Exact-sequence grouping is not homology separation. Primary/outer evaluations were observed historically; outer was not reevaluated. Multiple historical experiments/primaryselection and prior leaderboard-informed V6 component choice limit independence. Local metrics are not leaderboard metrics and do not guarantee improvement there. Detailed SCC/RMSE/MAE/R2,seen/unseencompounds,triplet-similarity subgroups,bootstrap/resource measurements are preserved outside the payload.
Prospective runtime amendment
Historical conservative projection502.021s failed the original480s gate and remains recorded. The user separately authorized one representative1502DISTINCT training-protein end-to-end measurement with unchanged8Mpolicy:229.576s,includingimports,encoder,actualV6features/forests/CSV. It passed the prospective measured480s preference. This did not change model-selection gates or imply accuracy improvement. Final exact-staged augmented-model Docker results and package hashes are recorded in the project readiness report; the proxy benchmark alone is not final deployment verification.
Actual full-training refit
Feature generation: 45.214s. Km forest fit: 766.838s. Km model bytes: 55,611,876. Fit-process lifetime peak RAM: 1227005952bytes. No new holdout metrics are produced by full refit. Final cold augmented-model timing/coverage/memory and native/Linux agreement evidence is stored outside this payload in the dated readiness record.
Offline inference
python /app/predict.py --input INPUT --output OUTPUT
InputParquetcolumns exactly id,smiles,sequence. Starter nonemptystring/ID/length rules apply, with globalduplicate-ID rejection. No output grouping/deduplication:everyinputID exactlyonce,originalorder. CSVexactlyid,log10_kcat,log10_km; finite abs<=1e6,CSV<16MiB. All assets local, arbitrary unseen sequences supported by the same boundedwindowpolicy. RequiredLinuxamd64,2CPUs,16GiB,600s,/usr/local/bin/python,repo<4GiB. UPLOAD_THIS contains only required inference code,dependencies,weights,preprocessors,license/notice/config. No datasets,trainingembeddings,predictioncaches,credentials or development logs. No automatic upload/submission/nomination.