Partitioned learning of deep Boltzmann machines for SNP data

Area of Science:

• Genetics and Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Deep learning excels at learning joint distributions and identifying low-dimensional manifolds in data.
• Traditional deep learning struggles with single nucleotide polymorphism (SNP) data due to high feature dimensionality compared to sample size.
• Novel methods are needed to apply deep learning effectively to complex genetic datasets like SNPs.

Purpose of the Study:

• To adapt deep Boltzmann machines (DBMs) for analyzing single nucleotide polymorphism (SNP) data.
• To overcome the high dimensionality challenge in SNP analysis using a partitioned learning approach.
• To identify complex SNP patterns and their influence on biological outcomes.

Main Methods:

• A sparse regression approach is used for coarse screening of SNP joint distributions.
• Partitioned learning involves training multiple DBMs on identified SNP subsets.
• Aggregate features and SNP patterns are extracted using statistical tests and sparse regression.

Main Results:

• The proposed method effectively handles high-dimensional SNP data by partitioning.
• Complex SNP patterns were uncovered in simulated case-control data, augmenting univariate analyses.
• In acute myeloid leukemia patients, three SNPs jointly influencing survival were identified using gene expression pre-screening.

Conclusions:

• Partitioned learning of DBMs is a viable approach for SNP data analysis.
• Jointly analyzing SNPs offers added value beyond standard univariate methods.
• This approach provides a complementary tool for complex genetic data analysis.