Machine and deep learning to predict viral fusion peptides.

Area of Science:

• Virology
• Bioinformatics
• Computational Biology

Background:

• Enveloped viruses like SARS-CoV-2 utilize surface fusion proteins for host cell entry.
• Fusion peptides (FPs) within these proteins are essential for viral fusion and represent therapeutic targets.
• Experimental FP identification is laborious and costly, necessitating computational prediction tools.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting fusion peptide locations in viral fusion proteins.
• To explore various ML approaches, sequence representations, and feature combinations for accurate FP identification.
• To identify novel putative fusion peptides, particularly for viruses with limited experimental data.

Main Methods:

• Employed token classification and sliding window techniques with machine and deep learning models.
• Evaluated diverse protein sequence representations: one-hot encoding, physicochemical features, NLP embeddings, and transformers.
• Tested over 50 combinations of ML models and sequence features.

Main Results:

• Achieved promising results using ML, especially with transformer-based models for amino acid token classification.
• Transformer models demonstrated high efficacy in predicting fusion peptide locations.
• Successfully predicted hypothetical fusion peptides for SARS-CoV-2 and analyzed existing annotations.

Conclusions:

• Developed effective ML models for predicting fusion peptide locations in viral fusion proteins.
• Transformer-based approaches show significant potential for identifying FPs, even with limited experimental data.
• This computational strategy can accelerate the discovery of new fusion peptides for therapeutic development.