SSAlign: Fast and Sensitive Protein Structure Search at Scale.

Abstract

MOTIVATION Predicted protein structure resources now contain hundreds of millions of models, turning structural comparison from a pairwise alignment task into a database-scale retrieval problem. Exhaustive structural aligners are computationally impractical at this scale, whereas fast structural-alphabet searches may discard remote or low-complexity homologs during prefiltering. Database-scale search therefore requires methods that combine rapid retrieval with sensitivity to global structural similarity. RESULTS We present SSAlign, a protein structure search method that integrates SaProt sequence-structure embeddings, entropy-reduced dense retrieval, an SS-score predictor and two-stage alignment refinement. In benchmarks on the Swiss-Prot, SCOPe40 and AFDB50 datasets, SSAlign showed competitive alignment quality and returned more TM-align-supported hits than Foldseek in recall-oriented evaluations. On AFDB50, 1,000 queries against 53.7 million structures completed in 3,150 s on CPUs and 2,290 s with GPU prefiltering, corresponding to 103-fold and 142-fold speedups over Foldseek, respectively. On SCOPe40, SSAlign increased the area under the curve (AUC) by 20.2% at family level and 33.3% at superfamily level. In simple-fold antimicrobial peptide case studies, SSAlign retrieved candidates that were not returned by Foldseek under the tested settings, indicating its utility as a fast candidate-generation tool for proteome-scale structural search. AVAILABILITY AND IMPLEMENTATION The source code for SSAlign is publicly available at https://github.com/ISYSLAB-HUST/SSAlign. The software version and test data used in this study are archived on Zenodo at https://doi.org/10.5281/zenodo.23432919. Datasets and prebuilt indices can be downloaded from http://bioinfo.isyslab.info/ssalign/download/section/ssalign/. SUPPLEMENTARY INFORMATION Supplementary data are available online.