Fitting Gaussian mixture models on incomplete data.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Bioinformatics research often integrates diverse datasets, leading to incomplete data with missing values.
• Existing Gaussian Mixture Models (GMMs) struggle with missing data, often requiring restrictive assumptions or leading to biased results via complete case analysis or imputation.

Purpose of the Study:

• To introduce missingness-aware Gaussian mixture models (MGMM), an R package designed to fit GMMs robustly in the presence of missing data.
• To provide a statistically sound and user-friendly tool for clustering and density estimation with incomplete datasets.

Main Methods:

• Development of the MGMM R package, which accommodates missing data without imposing restrictions on the covariance matrix.
• Evaluation using three case studies involving real and simulated 'omics data.

Main Results:

• MGMM demonstrated superior performance in recovering true cluster assignments compared to existing GMM implementations and standard GMMs with imputation.
• The package accurately assesses cluster assignment uncertainty, even when data distributions deviate from a true GMM.

Conclusions:

• MGMM significantly improves cluster assignment recovery across various datasets and missingness rates compared to state-of-the-art methods.
• MGMM offers a powerful, accessible, and statistically valid solution for bioinformatics analyses involving missing data.