iarmourgarb/ArCHer

ArCHer is a method for remote homology search and retrieval based on sequence alone. It works in two phases:

1. Shortlist (ArCHer_shortlist): each sequence is embedded as one mean-pooled vector. Each query keeps the database sequences most similar to it by cosine similarity (default: 4096).
2. Rerank (ArCHer_reranker): each shortlisted sequence is scored by a learned alignment of per-residue embeddings. The learned Needleman-Wunsch alignment is scored just over the query sequence, with learned affine gap costs. Homolog hits are ranked by this alignment score.