Learning genetic epistasis using Bayesian network scoring criteria.

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Gene-gene epistatic interactions are crucial for common diseases.
• Machine learning methods like Multifactor Dimensionality Reduction (MDR) and Bayesian Network (BN) models are used to detect epistasis.
• Previous methods like BNMBL showed success but may not generalize to unknown numbers of interacting SNPs.

Purpose of the Study:

• To evaluate various BN scoring criteria for identifying epistatic interactions.
• To determine the optimal scoring criterion for learning epistatic models without prior knowledge of the number of interacting SNPs.
• To compare the performance of different BN scoring criteria against MDR.

Main Methods:

• Evaluated 22 BN scoring criteria on 28,000 simulated datasets.
• Utilized a real-world Alzheimer's Genome-Wide Association Study (GWAS) dataset.
• Compared performance based on recall and detection of complex epistatic models.

Main Results:

• The Bayesian scoring criterion with high alpha values demonstrated superior performance.
• This criterion outperformed other BN scoring methods and MDR in recall on simulated data.
• It also excelled at detecting challenging epistatic models and validated findings in the Alzheimer's GWAS data.

Conclusions:

• BN models combined with appropriate scoring criteria are effective for identifying epistatic genetic variants.
• The Bayesian scoring criterion with large alpha values is a highly promising approach for epistasis detection.
• This method offers advantages over existing alternatives for uncovering complex genetic interactions in disease.