Multi-vector retrieval enables residue-resolved prediction of protein partners by ColBERT-PPI
- 1Shanghai Jiao Tong University School of Medicine
Abstract
Protein-protein interactions (PPIs) are central to biological processes, making the identification of both interacting partners and their binding sites important for understanding molecular function and guiding therapeutic discovery. However, connecting large-scale partner prediction to residue-level interaction evidence remains challenging. Here we present ColBERT-PPI, a structure-aware multi-vector framework that links protein partner retrieval to interaction-site localisation. The model retains reusable residue-level representations and compares them in a partner-dependent manner, producing raw residue compatibility maps for localisation and calibrated smooth MaxSim scores for retrieval. On PINDER, ColBERT-PPI achieved an area under the precision-recall curve of 0.482, compared with 0.151 for the strongest evaluated external comparator. Both complete and sequence-only multi-vector models outperformed the single-vector control on an independent human yeast two-hybrid screen. Residue-level comparisons highlighted experimental interfaces, captured changes in local evidence across partners and achieved a mean contact AUROC of 0.806 versus 0.526 for the single-vector control. After task-specific fine-tuning, PPI-trained representations also improved protein-RNA interaction retrieval, indicating transfer beyond the molecular context in which they were learned. By connecting candidate partners with their supporting residue-level evidence, ColBERT-PPI provides a computational basis for investigating physiological and disease-associated interactions and prioritizing proteins and interaction regions for therapeutic exploration. Code is available at https://github.com/UR-Free/ColBERT-PPI.
Related
- A Residue-Resolved Mechanism for Trapped-Twist Spectral Tuning in Cyanobacteriochromes.
- Scalable Enzyme Reaction Retrieval Enables Metagenomic Biocatalyst Discovery
- PathwayFinder enables linked exploration of cell types and gene modules in spatially resolved transcriptomics data
- Haplotype-resolved genetic mapping using selfed progeny enables simultaneous genome assembly validation and trait dissection in a heterozygous plant
- EV-Tracer enables lineage-resolved detection and molecular profiling of extracellular vesicle-associated signals in cancer-fibroblast co-culture
BibTeX
@misc{yang2026multivector,
title = {{Multi-vector retrieval enables residue-resolved prediction of protein partners by ColBERT-PPI}},
author = {H. Yang and R. Lei and Y. Zhuang and J. Zhang},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.19.752878},
url = {https://www.biorxiv.org/content/10.64898/2026.09.19.752878v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Multi-vector retrieval enables residue-resolved prediction of protein partners by ColBERT-PPI
AU - H. Yang
AU - R. Lei
AU - Y. Zhuang
AU - J. Zhang
PY - 2026
DA - 2026/09/19
PB - bioRxiv
DO - 10.64898/2026.09.19.752878
UR - https://www.biorxiv.org/content/10.64898/2026.09.19.752878v1
AB - Protein-protein interactions (PPIs) are central to biological processes, making the identification of both interacting partners and their binding sites important for understanding molecular function and guiding therapeutic discovery. However, connecting large-scale partner prediction to residue-level interaction evidence remains challenging. Here we present ColBERT-PPI, a structure-aware multi-vector framework that links protein partner retrieval to interaction-site localisation. The model retains reusable residue-level representations and compares them in a partner-dependent manner, producing raw residue compatibility maps for localisation and calibrated smooth MaxSim scores for retrieval. On PINDER, ColBERT-PPI achieved an area under the precision-recall curve of 0.482, compared with 0.151 for the strongest evaluated external comparator. Both complete and sequence-only multi-vector models outperformed the single-vector control on an independent human yeast two-hybrid screen. Residue-level comparisons highlighted experimental interfaces, captured changes in local evidence across partners and achieved a mean contact AUROC of 0.806 versus 0.526 for the single-vector control. After task-specific fine-tuning, PPI-trained representations also improved protein-RNA interaction retrieval, indicating transfer beyond the molecular context in which they were learned. By connecting candidate partners with their supporting residue-level evidence, ColBERT-PPI provides a computational basis for investigating physiological and disease-associated interactions and prioritizing proteins and interaction regions for therapeutic exploration. Code is available at https://github.com/UR-Free/ColBERT-PPI.
ER -