Representation-aware calibration enables statistically reliable protein structure search

Struct-Caliper improves protein structure search significance estimates by using representation-aware query randomization to produce reliable SG p-values, boosting homolog recovery by 18.2% in ASTRAL40 searches.

• Why it matters: Native E-values from tools like Foldseek and Reseek can be unreliable, risking false discoveries or missed remote homologs when prioritizing hits from large predicted-structure databases.
• What they did: Struct-Caliper generates background queries via representation-aware query randomization and computes query-specific p-values by comparing observed alignment scores to the best randomized-search scores from the same database.
• The result: In simulations, SG p-values kept false positives at or below the chosen cutoff; in 15,177 ASTRAL40 structures, SG p-values increased Foldseek recovery by 18.2% at cutoff 0.01, and for Reseek false discoveries decreased by 99.5%.

AI summary of the abstract · may contain errors · not medical advice · read the paper.