Legend7001/ec-evidence-v6

Enzyme classification: MMseqs2 homology transfer

DA3410 project submission for the Leak CURBER enzyme classification task.

Each unique query sequence is searched against the training sequences with MMseqs2 (sensitivity 5.7, run on 2 threads, as separate prefilter and align steps so nothing is executed from /tmp). The ECs of the hits become candidates. A gradient-boosted classifier scores each candidate from homology evidence only (bitscores relative to the best hit, identity, coverage, rank, agreement among hits, length ratio, competition with sibling ECs). It uses no EC-identity priors such as EC frequency. It was trained on out-of-fold searches of every training sequence against the other four of five 40%-identity cluster folds. The best candidate is returned as the single prediction, since only the first EC is scored.

• predict.py: offline CPU inference (MMseqs2 search, reranking, time guard, vocabulary filtering).
• ecfeat.py: candidate features. weights/rerank_hgb.joblib + rerank_features.json: the reranker. weights/rerank_lr.json: an older logistic reranker, used only as a fallback.
• bin/: static MMseqs2 binaries (AVX2 and SSE4.1 builds; the right one is chosen from /proc/cpuinfo).
• db/: MMseqs2 database of the training sequences (names t<row>). The k-mer index is built at run time.
• weights/train_ecs.json: ECs of each training sequence, by row. weights/baseline.json: fallback EC.

MMseqs2: Steinegger and Soding, Nature Biotechnology 2017 (GPLv3). EC vocabulary: ExPASy ENZYME, release 02-Sep-2026, SIB, CC BY 4.0.