Scoring Configs
A ScoringConfig defines how raw prediction signals are combined into a single [0, 1] confidence score — without re-running the KNN pipeline.
Available signals
Embedding similarity1 − cosine_distance / 2 — always available.
Identity NWNeedleman-Wunsch global sequence identity [0, 1].
Identity SWSmith-Waterman local sequence identity [0, 1].
Evidence weightQuality of the reference annotation's GO evidence code.
Taxonomic proximity1 / (1 + taxonomic_distance)
linear: Σ(w_i · s_i) / Σ(w_i) over all active (w_i > 0 and signal available) signals.
evidence_weighted: Same as linear, then multiplied by the resolved evidence weight — down-ranks IEA even when other signals are strong.
Loading…