IEEE BIBM 2024

ProtLOCA

Structure-only local geometry alignment for protein homology detection — asking whether amino-acid type embeddings always help, and when geometry alone is the better cue.

Yang Tan, Lirong Zheng, Bozitao Zhong, Liang Hong, Bingxin Zhou

Paper arXiv Code

Key idea

Many structure models still inject amino-acid identities by default. ProtLOCA shows that, for structure alignment, sequence embeddings are not always helpful — dissimilar sequences can share folds, and similar sequences can diverge structurally, so type features can dilute the geometric signal.

The model therefore encodes structure-only local geometry with roto-equivariant graph networks over backbone neighborhoods, without relying on residue-type embeddings for matching.

What ProtLOCA evaluates

  • Global matching on CATH — binary domain-pair similarity labels from CATH; ProtLOCA matches structurally consistent domains more quickly and accurately than common sequence- and structure-based baselines on held-out CATH-aligns benchmarks.
  • Local common-structure pairing — residue-level point matching that highlights shared local folds across proteins with different overall structures but the same function (e.g. DNA-binding motifs), where global aligners such as TM-align can miss the local correspondence.

Together, the two settings argue for geometry-first representation when the inference goal is structural homology rather than sequence identity.

Citation

If you use ProtLOCA, please cite:

@inproceedings{tan2024protloca,
  title={Protein representation learning with sequence information embedding: Does it always lead to a better performance?},
  author={Tan, Yang and Zheng, Lirong and Zhong, Bozitao and Hong, Liang and Zhou, Bingxin},
  booktitle={2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)},
  pages={233--239},
  year={2024},
  organization={IEEE}
}