Human protein-protein interaction prediction using twin-stream bidirectional long short-term memory model

The study of proteomics is fundamental in regulating cellular processes and understanding complex biological systems. By revealing changes in protein expression and interaction networks, it provides insights into cellular pathways involved in cancer, cardiovascular diseases, neurological disorders, and infectious diseases. Accurate identification of Protein-Protein Interaction (PPI) is essential for explaining the molecular mechanisms underlying human diseases and advancing drug discovery. However, experimental methods for PPI detection are often labor-intensive, time-consuming, and limited in scalability. To address these challenges, this research proposed a novel deep twin-stream learning method, TS-BiLSTM-PPI (Twin-Stream Bidirectional LSTM for PPI), with shared protein sequence embedding weights. For conducting research experiments, we have used the Pan’s PPI benchmark database, comprising 73,132 human protein sequences. We have also applied amino acid sequence-based data analysis to check protein distributions and isoelectric points. We have built traditional Machine Learning (ML) methods for fair comparisons with the proposed deep twin-stream architecture. The proposed TS-BiLSTM-PPI outperformed state-of-the-art traditional methods, achieving a correct PPI rate of 99.3% (AUC score). The ablation study and external PiNUI data validation results confirm the generalization of the TS-BiLSTM-PPI model for PPI prediction. The real-time PPI inference is performed using the proposed twin-stream model to assess interaction outputs with confidence scores based on protein-protein sequence pairs. This study provides an approach with practical PPI applicability for real-world proteomics scenarios.