S-PLM-TM enables fast and accurate sequence-based search for remote homologs
Protein structures and functions can remain conserved after extensive sequence divergence, but detecting these distant relationships across large databases remains challenging. Here we present S-PLM-TM, a sequence-only remote homology search tool based on a structure-aware protein language model. Cross-modal flow matching and contrastive learning incorporate structural information into sequence representations. A lightweight calibration head enables structural-similarity search with the frozen encoder, while a residue-level alignment head generates structural alignments directly from sequence. S-PLM-TM outperforms the tested sequence-based methods in remote homology detection, rivals structure-based methods in structural classification and retrieves the most functionally consistent neighbors. It also retrieves distantly related enzymes with shared catalytic functions at a level comparable to the best-performing structure-based methods. In large-scale database search, S-PLM-TM achieves the highest average precision among both sequence-based and alignment-free methods, with speed approaching the fastest sequence-alignment tools. Together, these results establish S-PLM-TM as a practical tool that combines sensitive and biologically informative remote homology search with the speed required for database-scale retrieval.
Authors
- Duolin Wang (ORCID: https://orcid.org/0000-0003-1649-460X)
- Qing Shao (ORCID: https://orcid.org/0000-0001-9433-1131)
- Shuai Zeng (ORCID: https://orcid.org/0000-0001-7632-427X)
- Yichuan Zhang
- Dong Xu
- Yuexu Jiang (ORCID: https://orcid.org/0000-0001-8330-1113)
Institutions
- University of Kentucky (US)
- University of South Florida (US)
Publication Details
- Journal
- bioRxiv (Cold Spring Harbor Laboratory)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.64898/2026.11.16.767046
- Primary Topic
- Protein Structure and Dynamics
- Type
- preprint
Funders
- National Institutes of Health