RohanTheZ/leak-curber-enzyme-v2

Leak CURBER enzyme classification: DIAMOND homology + learned reranker

1. Each query is aligned against all unique training sequences with DIAMOND (--very-sensitive, top 50 hits, E-value < 1e-3), in chunks so that no file exceeds 16 MiB.
2. Every EC carried by a hit becomes a candidate. rerank.py describes each candidate with alignment features (best bitscore and identity for that EC, vote shares, rank of first supporting hit, coverage, agreement at the first three EC digits, EC frequency in training).
3. A gradient-boosted classifier (weights/reranker.pkl, scikit-learn) scores each candidate and the highest-scoring EC is predicted. Queries with no hit get the most frequent training EC.

The reranker was trained on five held-out folds of whole 40%-identity clusters of the training set, so its training queries have no close relatives in the reference database, as in the hidden split.

• train.py --train train.parquet builds the DIAMOND database and label files.
• predict.py --input INPUT --output OUTPUT runs inference on CPU.
• bin/diamond is DIAMOND v2.1.10 (linux x86-64 static binary, https://github.com/bbuchfink/diamond).

EC vocabulary: ExPASy ENZYME release 02-Sep-2026, SIB, CC BY 4.0.