Alternative contingency table measures improve the power and detection of multifactor dimensionality reduction

Area of Science:

• Genetics
• Biostatistics
• Computational Biology

Background:

• Multifactor Dimensionality Reduction (MDR) is a non-parametric statistical method for detecting gene-gene interactions.
• MDR assigns multi-locus genotypes to high- or low-risk groups, using classification error to score model quality.
• This study aimed to enhance MDR's gene-gene interaction detection by exploring alternative model scoring measures.

Purpose of the Study:

• To compare the detection and power of MDR using various measures for two-way contingency table analysis.
• To identify superior fitness functions for MDR beyond traditional classification error.
• To improve the accuracy and reduce spurious variable inclusion in MDR models.

Main Methods:

• Simulated 40 genetic models with varying numbers of disease loci (2-5), allele frequencies (.2/.8 or .4/.6), and heritability (.05-.3).
• Evaluated 10 different measures for two-way contingency table analysis to score MDR model quality.
• Compared the performance of MDR using normalized mutual information (NMI) and likelihood ratio against classification error.

Main Results:

• Normalized Mutual Information (NMI) achieved 65.36% overall detection and 59.4% specific detection across models.
• Classification error yielded 62% overall detection and 52.2% specific detection.
• Likelihood ratio and NMI demonstrated superior performance in detecting gene-gene interactions.

Conclusions:

• Normalized Mutual Information (NMI) and likelihood ratio consistently improve MDR's detection and power in simulated data.
• These alternative measures effectively reduce the inclusion of spurious variables in multi-locus models.
• MDR's capability for detecting gene-gene interactions can be significantly enhanced by employing these alternative fitness functions.