Dissecting trait heterogeneity: a comparison of three clustering methods applied to genotypic data

Area of Science:

• Genetics
• Computational Biology
• Statistical Genetics

Background:

• Trait heterogeneity complicates the genetic study of complex human diseases.
• Unsupervised computational methods can identify trait heterogeneity when detailed phenotypic data is unavailable.
• Locus heterogeneity and gene-gene interactions further complicate genetic model discovery.

Purpose of the Study:

• To compare the performance of three unsupervised clustering methods (Bayesian Classification, Hypergraph-Based Clustering, Fuzzy k-Modes Clustering) for detecting trait heterogeneity.
• To evaluate the ability of these methods to handle locus heterogeneity and gene-gene interactions.
• To assess the reliability of Bayesian Classification's internal clustering metrics using permutation testing.

Main Methods:

• Simulation of genetic datasets with varying complexity.
• Application of Bayesian Classification, Hypergraph-Based Clustering, and Fuzzy k-Modes Clustering to simulated data.
• Evaluation of clustering performance using metrics like cluster recovery.
• Permutation testing to assess the validity of Bayesian Classification's internal metrics.

Main Results:

• Bayesian Classification demonstrated superior performance overall, accurately recovering underlying trait heterogeneity in most simulated datasets.
• Fuzzy k-Modes Clustering showed better performance on the most complex genetic models.
• Bayesian Classification achieved excellent recovery for 75% of datasets under the simplest model and moderate recovery for larger sample sizes.
• Internal clustering metrics for Bayesian Classification showed well-controlled false positive rates (≤3%) and acceptably low false negative rates (18% at p=0.10).

Conclusions:

• Bayesian Classification is a promising unsupervised method for dissecting trait heterogeneity in genotypic data.
• The method's robust control over false positive and false negative rates enhances confidence in its findings.
• Further research is needed to optimize Bayesian Classification parameters for complex genetic models.