EUVIMEDEUVIMED entdecken
Crossref

AttnSeq-PPI: Enhancing protein-protein interaction network prediction using transfer learning-driven hybrid attention

Dipayan Sarkar, Chiranjib Sarkar

Biochimica et Biophysica Acta (BBA) - Proteins and Proteomics · 2026

DOI: 10.1016/j.bbapap.2025.141102 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Biochimica et Biophysica Acta (BBA) - Proteins and Proteomics, 2026. Verfasst von Dipayan Sarkar und 1 weiteren Autor:innen. Beteiligte Einrichtung: Computational System Biology Laboratory, Department of Bioinformatics, University of North Bengal, India. Hauptthemen laut MEDLINE: Protein Interaction Maps, Deep Learning, Protein Interaction Mapping, Computational Biology. Die Arbeit nennt 43 Literaturangaben. Gefördert von: University Grants Commission.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Study of protein-protein interaction (PPI) network is fundamental to all cellular processes in living organisms. PPI study based on experimental techniques such as high-throughput assays, mass spectrometry is time-consuming and expensive. Computational techniques like molecular docking are restricted to availability of 3D protein structures and not suitable for large numbers of proteins. On the contrary Sequence-based PPI study is advantageous over the limitations of above techniques. Sequence-based PPI study has gained a broader scope with the help of deep learning techniques and large language model (LLM). In this study, we proposed AttnSeq-PPI, a deep learning framework based on two channels of hybrid attention mechanism. The protein sequences are embedded in high dimensional space using ProtT5 language model. In our model hybrid attention mechanism is designed by combining self-attention and cross-attention which enable the model to extract features from each protein with respect to the contextual features of both the proteins. The hybrid attention mechanism effectively captures long-range dependencies within protein sequences as well as interacting features of both proteins. The model was trained and evaluated based on 5-fold cross validation on intra-species and multi-species datasets. Additionally, four datasets of independent species and true PPI network datasets were used for validation. AttnSeq-PPI demonstrates superior generalization and outperforms existing models, achieving 99 % accuracy for both human and multi-species datasets. It can provide prediction of novel PPIs with fewer false negatives and higher precision. Additionally, we developed a web-based tool on AttnSeq-PPI accessible at https://compbiosysnbu.in/attnseqppi/ to provide PPI prediction based on protein sequences.

Schlagwörter

MeSH-Schlagwörter (MEDLINE)

Protein Interaction MapsDeep LearningProtein Interaction MappingComputational BiologyProteinsHumansDatabases, ProteinSoftware

Stichwörter der Autor:innen

Cross-attentionHybrid attention mechanismProtT5Protein-protein interaction networkSelf-attention

Substanzen

Proteins

Autor:innen und Institutionen

  • Computational System Biology Laboratory, Department of Bioinformatics, University of North Bengal, India.
  • Computational System Biology Laboratory, Department of Bioinformatics, University of North Bengal, India. Electronic address: cschiranjib9@gmail.com.

Förderung

  • University Grants Commission

Zitieren

APA

Sarkar, D., & Sarkar, C. (2026). AttnSeq-PPI: Enhancing protein-protein interaction network prediction using transfer learning-driven hybrid attention. Biochimica et Biophysica Acta (BBA) - Proteins and Proteomics. https://doi.org/10.1016/j.bbapap.2025.141102

Vancouver

Sarkar D, Sarkar C. AttnSeq-PPI: Enhancing protein-protein interaction network prediction using transfer learning-driven hybrid attention. Biochimica et Biophysica Acta (BBA) - Proteins and Proteomics. 2026. doi:10.1016/j.bbapap.2025.141102

BibTeX
@article{Sarkar2026,
  author = {Sarkar, Dipayan and Sarkar, Chiranjib},
  title = {AttnSeq-PPI: Enhancing protein-protein interaction network prediction using transfer learning-driven hybrid attention},
  journal = {Biochimica et Biophysica Acta (BBA) - Proteins and Proteomics},
  year = {2026},
  doi = {10.1016/j.bbapap.2025.141102}
}
Kostenlos starten →