Power of data mining methods to detect genetic associations and interactions

Area of Science:

• Genetics
• Bioinformatics
• Statistical genomics

Background:

• Genetic association studies traditionally focus on individual marker effects.
• There is a critical need to model epistasis (gene-gene interactions) for a comprehensive understanding of genetic associations.
• Tree-based methods are powerful tools for modeling complex variable interactions.

Purpose of the Study:

• To systematically evaluate the power of three leading algorithms for discovering genetic associations in the presence of complex interactions.
• To compare the performance of Random Forests (RF), Monte Carlo Logic Regression (MCLR), and Multifactor Dimensionality Reduction (MDR).

Main Methods:

• Utilized algorithm-specific variable importance measures (VIMs) as statistics.
• Employed permutation-based resampling to generate null distributions and p-values.
• Assessed the power of RF, MCLR, and MDR through simulation studies and a real-world data analysis of non-Hodgkin lymphoma risk.

Main Results:

• Random Forests (RF) exhibited the highest statistical power across all simulation models.
• Monte Carlo Logic Regression (MCLR) showed comparable power to RF in some scenarios.
• Multifactor Dimensionality Reduction (MDR) consistently demonstrated the lowest power among the evaluated methods.

Conclusions:

• Variable importance measures from Random Forests (RF) are the most reliable for detecting genetic associations involving complex interactions.
• The performance of RF is influenced by tuning parameters, variable type (continuous vs. categorical), and the specific VIM employed.