Explainable deep neural networks for novel viral genome prediction

Area of Science:

• Bioinformatics
• Computational Biology
• Virology

Background:

• Viral infections cause diverse human diseases, including cancer and COVID-19.
• Accurate viral genome prediction is vital for understanding and combating diseases like AIDS and Ebola.
• Current computational methods often lack explainability, despite high classification performance.

Purpose of the Study:

• To develop an interpretable deep learning model for accurate viral genome prediction.
• To automatically extract biologically relevant features from viral sequences.
• To enhance the understanding of viral genome characteristics and improve disease prediction.

Main Methods:

• Proposed Convolutional Neural Network (CNN) and CNN-Long Short-Term Memory (CNN-LSTM) based methods, named EdeepVPP and EdeepVPP-hybrid.
• Implemented an interpretable CNN model (EdeepVPP) to extract vital patterns through learned filters.
• Utilized 19 human metagenomic contig experiment datasets with 10-fold cross-validation and leave-one-experiment-out cross-validation.

Main Results:

• The EdeepVPP-hybrid predictor achieved superior performance, with a mean AUC-ROC of 0.992 and AUC-PR of 0.990.
• Demonstrated the ability of CNN filters to detect critical patterns associated with viral sequences.
• Showcased the model's interpretability by identifying important features through learned filters.

Conclusions:

• EdeepVPP provides an interpretable approach to viral genome prediction, outperforming existing methods.
• The model can serve as a recommendation system for analyzing unknown viral sequences.
• Identified key biological patterns crucial for accurate virus sequence prediction.