High dimensional model representation of log likelihood ratio: binary classification with SNP data.

Area of Science:

• Bioinformatics
• Genetics
• Computational Biology

Background:

• Developing binary classification rules from SNP data is challenging for bioinformatics, especially for complex diseases like cancer.
• High-dimensional SNP data, weak individual SNP effects, and non-linear interactions complicate analysis.
• SNPs are categorical/ordinal but often treated as continuous variables by algorithms.

Purpose of the Study:

• To develop methods for building low-dimensional, interpretable models from high-dimensional SNP data.
• To account for single SNP effects and pairwise SNP interactions in classification tasks.
• To propose a regression-based approach (LABS-HDMR-CO) and a statistical test (FPT) for SNP interaction analysis.

Main Methods:

• Utilized High Dimensional Model Representation (HDMR) theory to model SNP effects and interactions.
• Computed second-order HDMR expansion of the log-likelihood ratio for SNP analysis.
• Developed Linear Approximation for Block Second Order HDMR Expansion of Categorical Observations (LABS-HDMR-CO) and Fixed Pattern Test (FPT).

Main Results:

• LABS-HDMR-CO demonstrated superior accuracy in SNP classification compared to other algorithms on synthetic and GWAS cancer data.
• FPT identified few significant pairwise interactions in small GWAS datasets due to the large number of tests.
• LABS-HDMR-CO leveraged numerous SNP pairs to enhance prediction accuracy, even if not all were statistically significant.

Conclusions:

• LABS-HDMR-CO and FPT offer valuable tools for creating prediction rules and detecting pairwise SNP interactions.
• Detecting interactions and using them for prediction accuracy are distinct but complementary objectives.
• Even with low detection power for interactions, utilizing potential interacting SNP pairs can improve predictive performance.