deepHPI

deepHPI predicts host-pathogen protein-protein interactions using convolutional neural networks and protein sequence-derived features to support analysis of infectious disease mechanisms.

Key Features:

• Deep Learning-Based Prediction: Employs convolutional neural network (CNN) models selected through comprehensive evaluations of protein features and neural network architectures.
• Multiple Host-Pathogen Models: Implements four host-pathogen model types: plant-pathogen, human-bacteria, human-virus, and animal-pathogen.
• High Prediction Accuracy: Reports Matthews correlation coefficient (MCC) values on independent validation datasets of 0.87 for animal-pathogen interactions (using combined PAAC and CT features), 0.75 for human-bacteria (with additional normalized Moreau-Broto features), 0.96 for human-virus, and 0.94 for plant-pathogen interactions.
• Comprehensive Feature Utilization: Utilizes protein feature representations including pseudo-amino acid composition (PAAC), conjoint triad (CT), normalized Moreau-Broto (NMBroto), and composition and transition features (CTDC_CTDT).
• High-Performance Computing Deployment: Deployed on a high-performance computing cluster to support computationally intensive model training and inference.
• Network Visualization: Augments predicted interactions with host-pathogen network visualizations and links to protein annotation resources for interpretation.

Scientific Applications:

• Infectious disease mechanism analysis: Facilitates investigation of molecular mechanisms underlying host-pathogen interactions across multiple organismal systems.
• Discovery of novel and known interactions: Enables prediction-driven identification and exploration of both novel and previously reported host-pathogen protein-protein interactions.
• Support for therapeutic research: Provides interaction predictions and network context to aid the development and prioritization of potential therapeutic strategies.

Methodology:

Convolutional neural networks trained and selected via evaluations of protein features and network architectures using feature sets (PAAC, CT, NMBroto, CTDC_CTDT), with performance assessed on independent validation datasets using Matthews correlation coefficient (MCC).