GlycoAudit: structural joinability and annotation degree shape glycan–agent relation ranking
Minqi Chen, Minrui Chen, Guanjie Zou, Lunxu LiuAbstract
Motivation
Benchmark scores can be misleading when identifier resolution, structural availability, and incomplete interaction annotation determine which glycans can be evaluated. GlycoAudit audits this evidence-to-structure path before model comparison.
Results
A strict accession, branch, and sequence contract retained 44 of 173 resolved SugarBind glycans (25.4%) and excluded 129 (74.6%). The joined domain was substantially more densely annotated. Under leakage-controlled held-out-positive ranking, the Web3 Unique Representation of Carbohydrate Structures (WURCS) exceeded the evaluated global structural summaries. Yet a leave-one-agent-out annotation-degree control achieved higher mean reciprocal rank (0.491 versus 0.381) and higher known-positive average precision (APknown), 0.489 versus 0.373, whereas WURCS achieved higher Recall@5 (0.688 versus 0.456). These metric-dependent reversals show that structural joinability, annotation degree, candidate construction, and evaluation choice jointly shape apparent benchmark performance. The release evaluates positive-relation ranking, not confirmed non-binding classification or cross-source transfer.
Availability
GlycoAudit makes evidence, joining decisions, exclusions, candidates, folds, controls, and machine-readable results inspectable in the public GlycoAudit repository.