venom2410/enzyme-knn

Enzyme classification: alignment + k-mer + ESM fusion

Three evidence sources are combined per query:

1. DIAMOND alignment search against the unique training sequences (hits vote for their ECs, weight exp((bits - best_bits)/T)).
2. Hashed k-mer TF-IDF nearest neighbours.
3. A fine-tuned ESM-2 8M classifier, run on CPU only for weakly aligned queries until a time limit.

Fusion weights are tuned on a hold-out of training sequences. Every stage has a time limit, and missing evidence is skipped, so inference always finishes in time. DIAMOND (https://github.com/bbuchfink/diamond) is shipped as a static Linux binary in bin/.