DeepHINT: understanding HIV-1 integration via deep learning with attention.

Area of Science:

• Virology
• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• Human immunodeficiency virus type 1 (HIV-1) genome integration is crucial for viral latency and rebound.
• Predicting HIV integration sites is challenging due to complex genomic context dependencies.
• Understanding integration mechanisms is vital for developing therapeutic strategies.

Purpose of the Study:

• To develop an explainable deep learning framework for accurate HIV-1 integration site prediction.
• To elucidate the mechanistic basis of HIV-1 integration site selection.
• To provide an open-source tool for researchers studying HIV-1 integration.

Main Methods:

• Developed DeepHINT, an attention-based deep learning framework.
• Utilized primary DNA sequence and epigenetic information for model training.
• Validated predictions against high-density HIV integration site datasets.
• Analyzed model attention values for mechanistic insights.

Main Results:

• DeepHINT outperforms conventional methods in predicting HIV-1 integration sites.
• The framework accurately identifies genomic contexts influencing integration.
• Attention analysis revealed potential roles of DNA-binding proteins in integration site selection.
• Demonstrated the biological relevance of predictions through systematic analyses.

Conclusions:

• DeepHINT is an effective and explainable deep learning framework for HIV-1 integration studies.
• The tool facilitates prediction and mechanistic understanding of HIV-1 integration.
• DeepHINT offers valuable insights into viral latency and rebound mechanisms.