Coevolutionary signals in multiple sequence alignments improve virulence factor prediction with an MSA Transformer

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Virulence factors (VFs) are crucial for understanding bacterial pathogenesis and developing treatments for infectious diseases.
• Coevolutionary information in protein sequences offers insights into the functional and structural roles of VFs.
• Previous VF prediction methods have not fully utilized coevolutionary data.

Purpose of the Study:

• To develop a novel deep learning method for predicting virulence factors (VFs) by incorporating coevolutionary information.
• To enhance the accuracy and understanding of bacterial pathogenicity through advanced protein sequence analysis.

Main Methods:

• Developed MSA-VF Predictor (MVP), a deep learning model utilizing Multiple Sequence Alignment (MSA) and MSA Transformer.
• Extracted coevolutionary and homologous protein features from MSA data.
• Proposed MSA-composition to represent amino acid latent vectors for prediction.

Main Results:

• Achieved a prediction accuracy of 0.869 for virulence factors, surpassing current state-of-the-art models.
• Demonstrated the significant contribution of coevolutionary information to MVP's predictive performance.
• Identified crucial attention blocks in the MSA Transformer model for VF prediction.

Conclusions:

• MSA-VF Predictor (MVP) effectively integrates coevolutionary information for accurate VF prediction.
• The study highlights the importance of evolutionary interdependencies in understanding protein function and pathogenicity.
• MVP provides a powerful new tool for bacterial pathogenesis research and drug development.