P-PPI: accurate prediction of peroxisomal protein-protein interactions (P-PPI) using deep learning-based protein sequence embeddings

Protein-protein interactions (PPIs) are crucial for various biological processes, and their prediction is typically accomplished through experimental methods, which can be time-consuming and costly. Computational methods provide a faster and more cost-effective approach, leveraging protein sequences and other data sources to infer PPIs. Deep learning (DL) approaches have shown promising results in various protein-related tasks, including PPI prediction. However, DL-based embeddings are often not thoroughly compared or evaluated against state-of-the-art tools. Additionally, existing PPI predictors incorporate different types of information beyond protein sequence representation, making it important to assess the effectiveness of DL-based embeddings solely relying on protein sequences. In this work, we benchmark and compare commonly used DL-based embeddings for PPI prediction based solely on protein sequence information. We utilize high-quality training data, including experimentally validated negative interactions from the Negatome database. The best model, obtained through double cross-validation and hyperparameter optimization, is selected and evaluated to predict peroxisomal PPIs. The resulting tool, P-PPI, is further enhanced by combining AlphaFold2-Multimer predictions with the P-PPI model, leveraging DL-based embeddings and protein structure predictions for a comprehensive analysis of peroxisomal PPIs. This integrated approach holds significant potential to advance our understanding of complex protein networks and their functions.