| --- |
| license: mit |
| tags: |
| - enzyme-classification |
| - ec-number |
| --- |
| |
| # Leak CURBER enzyme classification — homology kNN |
|
|
| Predicts one EC number per protein sequence by transferring labels from the most similar training enzymes. |
|
|
| ## Method |
| 1. **Prefilter**: TF-IDF-weighted 5-mer presence vectors; cosine similarity against all 186k unique training sequences, top 1,000 candidates per query (sparse matrix product). |
| 2. **Re-rank**: Smith-Waterman local alignment (BLOSUM62, gap open 11, extend 1) of the query against each candidate with SIMD Opal (`pyopal`). |
| 3. **Vote**: ECs of all hits scoring within 10% of the best hit, weighted by alignment score (multi-EC hits split their weight). |
|
|
| Falls back to the k-mer nearest neighbour if the 480 s time budget is reached, and to the most frequent EC if a query has no candidates. |
|
|
| ## Local validation |
| Validation sequences have no exact matches in train and very low k-mer overlap (median top-1 cosine 0.05), so a random split is misleading (median 0.65). The local hold-out is 4,915 training sequences drawn from the low-similarity pool, scored against the remaining 181k and importance-weighted to match the validation similarity distribution (effective sample size ~2,300). |
|
|
| | Metric | Weighted hold-out | |
| |---|---| |
| | L1 | 0.943 | |
| | L2 | 0.912 | |
| | L3 | 0.899 | |
| | L4 | 0.746 | |
|
|
| ## Runtime |
| 11,367 validation inputs: 233 s, 4.0 GB peak RAM on 2 vCPU. |
|
|
| ## Files |
| - `predict.py` — inference entry point |
| - `kmer.py` — sparse k-mer index |
| - `train.py` — builds `weights/db.parquet` from `train.parquet` |
| - `inputs.py`, `task.json`, `ec_vocabulary.json` — from the starter |
| - `weights/db.parquet` — unique training sequences with their ECs |
|
|
