Sequence alone can now find distant protein relatives: S-PLM-TM leads below 30% identity and searches 430,140 targets in 8.6 seconds
Synopsis
Remote homologs often share less than 30% sequence identity, where conventional sequence alignment can no longer separate true homologs from unrelated proteins, while structure comparison requires a structure for every entry. S-PLM-TM writes structural information into a sequence-only protein language model representation during training, so inference encodes sequence alone and a lightweight calibration head converts embedding similarity into a structural-similarity score. On SCOPe family and superfamily retrieval it reaches the highest mean AUC-1FP in all three bins below 30% identity, and it searches 957 queries against 430,140 Swiss-Prot targets in 8.6 seconds with a mean average precision of 0.638, the highest among sequence-based methods.
Interpretation
Remote homology search can now take amino-acid sequences alone and return candidates ranked by structural similarity, without predicting structures for the query or the database entries. Earlier sequence-search tools lose sensitivity once pairwise identity falls below 25–30%, and structure-comparison tools, though more sensitive, require a reliable structure for the query and for every database entry. In SCOPe family-level retrieval, S-PLM-TM reaches mean AUC-1FP of 0.482, 0.788 and 0.942 in the [0,10), [10,20) and [20,30)% identity bins, against 0.419, 0.717 and 0.886 for PLMSearch and 0.252, 0.698 and 0.909 for TM-Vec.
Structural knowledge enters the sequence representation during pretraining, and inference keeps only the frozen sequence encoder plus a calibration head. Earlier representation-based methods attached structure to protein pairs only as a scalar label, so the encoder itself never observed a three-dimensional structure. Training maps residue-level sequence representations onto structure representations with cross-modal flow matching and organizes protein representations with structure-softened contrastive negatives; the calibration head is trained on TM-align TM-scores with the encoder frozen and converts cosine similarity into a structural-similarity score.
The same sequence representation also supports enzyme-function retrieval and residue-level alignment, extending what a retrieved hit can be used for. Global structural similarity alone cannot separate enzymes that share a fold but differ in catalytic function, because catalysis is often set by a few active-site residues. On the DEER microbial–host isozyme task, S-PLM-TM without EC supervision reaches mean recall 0.266±0.040, level with the strongest structural aligner Dali at 0.263±0.040; with the EC head it reaches 0.382±0.042 against 0.393±0.043 for the structure-input DEER model, a difference that is not significant (adjusted P=0.73).
Perspective
The demonstrated scope covers SCOPe, CATH-S20, GO functional consistency, DEER isozymes and a 430,140-target SCOPe-to-Swiss-Prot search; the large-scale query and target sets were rebuilt by the authors from data released with PLMSearch, and the ground truth is an all-against-all TM-align matrix computed by the authors. The Pfam-clan prefilter raises retrieval on a domain database but lowers mean average precision on the full-length protein library, so the large-scale results are reported without it.
The authors note that the best use of the Pfam-clan prefilter depends on the target database, with different optimal settings for a domain database and a whole-proteome database, so users must decide whether to apply it or run each query both ways and merge the lists. They also call for evaluation across a wider range of taxa and database compositions, especially proteins from underrepresented organisms and metagenomic samples, to determine how broadly the method generalizes beyond the current benchmarks.
