A comparison of internal validation techniques for multifactor dimensionality reduction.

Area of Science:

• Genetics
• Computational Biology
• Statistical Genetics

Background:

• Complex diseases arise from gene-environment interactions, challenging traditional statistical methods.
• Multifactor Dimensionality Reduction (MDR) is a key data-mining technique for detecting epistasis.
• Internal validation is crucial in MDR to prevent overfitting and false positives.

Purpose of the Study:

• To evaluate a three-way split (3WS) with pruning as an alternative to cross-validation for MDR internal validation.
• To assess the computational efficiency and performance of 3WS compared to cross-validation.
• To compare the power of detecting true disease loci using different validation strategies.

Main Methods:

• MDR with 3WS and a post-hoc pruning procedure was developed as an alternative to cross-validation.
• The study compared 5-fold and 10-fold cross-validation with 3WS for various genetic models.
• A real-world HIV immunogenetics dataset was analyzed to demonstrate the methods.

Main Results:

• MDR with 3WS was approximately five times faster than 5-fold cross-validation.
• Before pruning, 5-fold cross-validation showed higher power for detecting exact disease loci.
• After pruning, 3WS demonstrated equivalent power to cross-validation in detecting true disease loci, including false positives.
• Both methods identified the same two-locus model in the HIV dataset.

Conclusions:

• The performance of 3WS with pruning is equivalent to cross-validation for MDR.
• 3WS offers a computationally efficient approach for screening epistatic effects in large genetic studies.
• The choice of pruning procedure involves a trade-off between identifying all genetic effects and minimizing false positives.