Identification of interactions using model-based multifactor dimensionality reduction.

Damian Gola,Inke R König

doi:10.1186/s12919-016-0019-8

Abstract

BackgroundCommon complex traits may involve multiple genetic and environmental factors and their interactions. Many methods have been proposed to identify these interaction effects, among them several machine learning and data mining methods. These are attractive for identifying interactions because they do not rely on specific genetic model assumptions. To handle the computational burden arising from an exhaustive search, including all possible combinations of factors, filter methods try to select promising factors in advance.MethodsModel-based multifactor dimensionality reduction (MB-MDR), a semiparametric machine learning method allowing adjustment for confounding variables and lower level effects, is applied to Genetic Analysis Workshop 19 (GAW19) data to identify interaction effects on different traits. Several filtering methods based on the nearest neighbor algorithm are assessed in terms of compatibility with MB-MDR.ResultsSingle nucleotide polymorphism (SNP) rs859400 shows a significant interaction effect (corrected p value <0.05) with age on systolic blood pressure (SBP). We identified 23 SNP–SNP interaction effects on hypertension status (HS), 42 interaction effects on SBP, and 26 interaction effects on diastolic blood pressure (DBP). Several of these SNPs are in strong linkage disequilibrium (LD). Three of the interaction effects on HS are identified in filtered subsets.ConclusionsThe considered filtering methods seem not to be appropriate to use with MB-MDR. LD pruning is further quality control to be incorporated, which can reduce the combinatorial burden by removing redundant SNPs.

Highlights

Common complex traits may involve multiple genetic and environmental factors and their interactions
We identified one Single nucleotide polymorphism (SNP) with a significant interaction effect with age on systolic blood pressure (SBP)
As an example for the internal representation of factor combinations in Model-based multifactor dimensionality reduction (MDR) (MB-MDR) we refer to Fig. 1, which shows the interaction rs859400 × age on SBP

Summary

Introduction

Common complex traits may involve multiple genetic and environmental factors and their interactions. Many methods have been proposed to identify these interaction effects, among them several machine learning and data mining methods. To handle the computational burden arising from an exhaustive search, including all possible combinations of factors, filter methods try to select promising factors in advance. Common complex traits, such as blood pressure, are at least partly based on complex molecular mechanisms likely to involve multiple genetic factors and their interactions. Two main classes of methods can be identified: parametric methods, like regression-based approaches, and nonparametric methods, which covers most machine learning and data-mining approaches. Model-based MDR (MB-MDR) [2] is one alteration that bridges the aforementioned classes by combining advantages of nonparametric approaches, that is, no distributional assumptions are imposed, with those of parametric approaches, that is, adjustment for confounding variables and lower level main effects

Methods

Results

Discussion

Conclusion