DeepVISP

DeepVISP predicts oncogenic virus integration sites (VISs) in the human genome using a convolutional neural network with an attention architecture to identify sequence features and cis-regulatory factors associated with viral integration and tumorigenesis.

Key Features:

• Model architecture: Convolutional neural network (CNN) with an attention architecture that automatically learns informative sequence features and essential genomic positions from DNA sequences.
• Target viruses and data: Trained on curated benchmark integration data from hepatitis B virus (HBV), human papillomavirus (HPV), and Epstein-Barr virus (EBV).
• Prediction output: Predicts oncogenic virus integration sites (VISs) in the human genome.
• Regulatory decoding: Decodes potential cis-regulatory factors involved in virus integration and tumorigenesis, including HOXB7, IKZF1, and LHX6.
• Motif discovery: Performs clustering analysis of informative motifs to identify representative k-mers that guide virus recognition of host genes.
• Performance: Demonstrates area under the curve (AUC) improvements of 8.43% to 34.33% compared with conventional machine learning methods.

Scientific Applications:

• Identification of VISs: Locating putative oncogenic virus integration sites in the human genome for HBV, HPV, and EBV.
• Mechanistic studies of viral oncogenesis: Investigating cis-regulatory factors such as HOXB7, IKZF1, and LHX6 associated with integration and tumorigenesis.
• Motif-guided host–virus interaction analysis: Revealing representative k-mers to study how viruses recognize and target host genes.
• Method benchmarking: Comparing deep-learning predictive performance against conventional machine learning using AUC metrics.

Methodology:

DeepVISP employs a convolutional neural network with an attention architecture trained on curated HBV, HPV, and EBV integration datasets to automatically learn informative sequence features and genomic positions from DNA sequences, followed by clustering analysis of informative motifs to identify representative k-mers.