Latent Forests to Model Genetical Data for the Purpose of Multilocus Genome-wide Association Studies. Which clustering should be chosen?
Résumé
The aim of genetic association studies, and in particular genome-wide association stu-
dies (GWASs), is to unravel the genetics of complex diseases. In this domain, machine learning
offers an attractive alternative to classical statistical approaches. The seminal works of Mourad
et al. (2011) have led to the proposal of a novel class of probabilistic graphical models, the forest of
latent trees (FLTM). The design of this model was motivated by the necessity to model genet-
ical data at the genome scale, prior to a multilocus GWAS. A multilocus GWAS fully exploits
information about the complex dependences existing within genetical data, to help detect the loci
associated with the studied pathology. The FLTM framework also allows data dimension reduc-
tion. The FLTM model is a hierarchical Bayesian network with latent variables. Central to the
FLTM construction is the recursive clustering of variables, in a bottom up subsuming process.
This article focuses on the analysis of the impact of the choice of the clustering method used in
the FLTM learning algorithm, in a GWAS context. We rely on a real GWAS data set describing
41400 variables for each of 3004 controls and 2005 cases affected by Crohn’s disease, and compare
the impact of three clustering methods. We compare statistics about data dimension reduction
as well as trends concerning the ability to split or group putative causal SNPs in agreement with
the underlying biological reality. To assess the risk of missing significant association results due
to subsumption, we also compare the clustering methods through the corresponding FLTM-based
GWASs. In the GWAS context and in this framework, the choice of the clustering method does
not influence the satisfying performance of the GWAS.