Rapid and interpretable protein contact map prediction using a pattern-matching strategy

Abstract

Protein sequence determines structure, function, and dynamics, yet the gap between sequenced proteins and experimentally determined structures continues to widen. While machine learning approaches like AlphaFold2 have transformed structural biology, they require substantial computational resources. Coevolution-based methods such as mutual information (MI) and direct coupling analysis (DCA), such as GREMLIN, offer alternatives but depend on extensive multiple sequence alignments with thousands of homologs. Here, we present a template-based pattern-matching approach that predicts protein contact maps by identifying conserved structural motifs from homologous experimental structures. Our method encodes spatial arrangements of up to five residues within 8.0 Å distance as sequence patterns, then aligns these patterns to query sequences to predict residue–residue contacts. Critically, our approach requires only a modest number of structural templates (typically 50–500) and runs on standard hardware without graphics processing units or high-performance computing clusters, processing proteins in 12–16 min regardless of length. We validated our method on 25 well-characterized protein domains, achieving correlations of 0.735–0.942 with experimental contact maps. Comparative analysis against MI and GREMLIN demonstrated that our method achieved better contact coverage while maintaining comparable accuracy. To demonstrate broader applicability, we tested on 7599 poorly annotated sequences using high-confidence AlphaFold structures as reference, achieving mean F 1-score of 0.609 ± 0.095 and accuracy of 0.954 ± 0.036. Our pattern matching approach provides a computationally efficient, interpretable alternative to both deep learning and coevolution-based methods, particularly valuable for proteins with limited sequence homologs or when rapid predictions are needed.