Artificial neural networks trained to detect viral and phage structural proteins

Area of Science:

• Virology
• Bioinformatics
• Computational Biology

Background:

• Bacteriophages are crucial in microbial ecology and host pathogenicity.
• Identifying viral genes, especially structural proteins, is challenging due to high diversity and low sequence conservation.
• Accurate viral identification aids disease diagnosis, vaccine development, and environmental potential assessment.

Purpose of the Study:

• To develop a novel computational method for predicting phage structural protein sequences.
• To enhance the identification of unknown viral proteins for better understanding of viral functions and impact.

Main Methods:

• Designed and trained Artificial Neural Networks (ANNs) using amino acid frequency to classify viral structural proteins.
• Incorporated protein isoelectric points as features for specialized ANNs targeting major capsid and tail proteins.
• Experimentally validated ANN predictions using transmission electron microscopy on predicted structural proteins.

Main Results:

• ANNs trained on amino acid frequency achieved high specificity and sensitivity in classifying structural proteins.
• Specialized ANNs incorporating isoelectric points demonstrated improved accuracy for capsid and tail proteins.
• Experimental validation confirmed that ANN-predicted proteins self-assembled into structures resembling virions.

Conclusions:

• ANNs provide a powerful new tool for identifying phage and prophage structural proteins, particularly those with no known homologs.
• This method overcomes limitations of traditional bioinformatic analyses for functionally uncharacterized viral proteins.
• The developed ANNs are valuable for viral research when in vitro phage propagation is impractical.