Knowledge graph-aided Bayesian active learning for top-K genetic interaction discovery.

Area of Science:

• Computational Biology
• Genomics
• Infectious Disease Modeling

Background:

• In silico prediction of multi-gene perturbations is crucial for functional genomics, drug discovery, and disease modeling.
• Developing predictive algorithms for mammalian systems is challenging due to limited data and high experimental costs.

Purpose of the Study:

• To develop a Bayesian active learning framework for discovering pairwise host gene knockdowns that inhibit HIV-1 proliferation.
• To leverage biological knowledge graphs and batch diversification for efficient identification of gene interactions.

Main Methods:

• Implemented a Bayesian active learning framework incorporating a biological knowledge graph.
• Employed a computationally efficient batch diversification approach.
• Evaluated the framework on a dataset of viral load measurements from dual-gene depletion experiments across over 350 host genes.

Main Results:

• The framework rapidly identified effective gene knockdown pairs for reducing HIV-1 viral load.
• Incorporating side information (knowledge graph) improved early-stage active learning performance.
• Batch diversification significantly enhanced performance in later stages (high data regime).

Conclusions:

• The developed framework efficiently identifies gene pairs to inhibit viral proliferation in an HIV-1 model.
• The approach is generalizable for exploring gene interactions in other biological contexts, such as synthetic lethality and epistasis.
• This method offers a cost-effective and rapid approach to functional genomics and disease modeling.