ProDive: detecting cross-family local conservation

Abstract

Short conserved sequence-structure elements can recur across protein families even when whole-domain homology is weak or absent, but systematic detection of such short local signals remains difficult. Here we present ProDive, a profile hidden Markov model (pHMM)-based framework for detecting cross-family local conservation. ProDive combines a closed-form approximation to the divergence between equal-length pHMM windows, a background-suppression stage that identifies and removes low-information similarities, diagonal-path extraction in window-pair similarity landscapes, and filtering and rescoring of retained paths according to the fraction of seed-alignment sequences supporting the matched positions. GPU benchmarking showed numerical agreement with CPU calculations and predictable near-linear scaling with the number of window-pair matrix elements. When applied to all 25,545 Pfam families in release 38.0, ProDive yielded 318,289 high confidence pHMM segment correspondences after filtering. For fragment pairs with suitable representative structures, length-stratified controls showed that ProDive matches in the 8-13 aligned-residue regime have significantly lower C root-mean-square deviation (RMSD) than same-length random fragment pairs. Comparison with HHsearch showed that ProDive occupies a distinct fragment-scale detection niche: ProDive paths were largely outside HHsearch-reported family-pair space, and in shared family pairs ProDive boundaries isolated shorter compact cores. Together, these results establish ProDive as a practical method for local conservation identification and a resource for downstream studies of protein structure, evolution, and design.