VirusHound-I

VirusHound-I predicts viral proteins that mediate evasion of the host adaptive immune response.

Key Features:

• Machine Learning Foundation: Uses a multilayer perceptron algorithm to model relationships between protein features and immune-evasion function.
• Molecular Descriptor Utilization: Represents proteins using dipeptide composition descriptors.
• Data Augmentation Strategy: Augments the positive dataset using generative adversarial networks (GANs) to improve model training.
• High Predictive Performance: Validated with training accuracy 0.947 (precision 0.994, F1 0.943, specificity 0.995, sensitivity 0.896, kappa 0.894, Matthew's correlation coefficient 0.898, AUC 0.989) and testing accuracy 0.964 (precision 1.0, F1 0.967, specificity 1.0, sensitivity 0.936, kappa 0.929, Matthew's correlation coefficient 0.931, AUC 1.0).

Scientific Applications:

• Immune-evasion mechanism analysis: Facilitates investigation of molecular mechanisms by which pathogenic viruses evade the host adaptive immune response.
• Virus-host interaction studies: Identifies viral proteins involved in adaptive immune evasion to inform studies of virus-host interactions.
• Antiviral target discovery: Supports identification of candidate viral proteins that could serve as therapeutic targets for antiviral drug development.

Methodology:

Multilayer perceptron models were trained and validated on protein inputs encoded by dipeptide composition descriptors with positive-set augmentation performed using generative adversarial networks (GANs).