DeepHPV

DeepHPV predicts human papillomavirus (HPV) integration sites within the human genome using an attention-based deep learning model to learn genomic environmental features that influence integration preferences.

Key Features:

• Attention-Based Deep Learning Model: Employs an attention-based architecture to focus on sequence regions and genomic context associated with HPV integration.
• Training and Testing Datasets: Trained on 3,608 known HPV integration sites and validated against a testing dataset of 584 reviewed HPV integration sites.
• Baseline Performance Metrics: Achieved a baseline AUROC of 0.6336 and AUPR of 0.5670.
• Improved Performance with RepeatMasker and TCGA Pan Cancer Peaks: Incorporation of RepeatMasker and TCGA Pan Cancer peaks improved performance with AUROC values of 0.8464 (RepeatMasker) and 0.8501 (TCGA Pan Cancer) and AUPR values of 0.7985 (RepeatMasker) and 0.8106 (TCGA Pan Cancer).
• Independent Validation on VISDB: On VISDB, the model incorporating TCGA Pan Cancer peaks produced AUROC 0.7175 and AUPR 0.6284, while the RepeatMasker-enhanced model produced AUROC 0.6102 and AUPR 0.5577.
• Enrichment of Transcription Factor Binding Sites: The attention mechanism highlighted regions enriched with transcription factor binding sites such as BHLHA15, CHR, COUP-TFII, and others near intensive attention sites.

Scientific Applications:

• Genomic feature characterization: Characterizes genomic features and environmental context associated with HPV integration preferences.
• Cervical carcinogenesis research: Supports investigation of molecular mechanisms by which HPV integration contributes to cervical carcinogenesis.
• Biomarker and therapeutic target prioritization: Prioritizes candidate integration sites for research into early detection biomarkers and therapeutic targets in HPV-related cancers.

Methodology:

Uses an attention-based deep learning approach trained on 3,608 known HPV integration sites and validated on a testing set of 584 reviewed sites, incorporating RepeatMasker annotations and TCGA Pan Cancer peaks; models were independently evaluated on VISDB.