Fast approximate inference for variable selection in Dirichlet process mixtures, with an application to pan-cancer...

Area of Science:

• Computational statistics
• Bioinformatics
• Machine learning

Background:

• Dirichlet Process (DP) mixture models are widely used for model-based clustering, enabling inference of the number of clusters.
• The sequential updating and greedy search (SUGS) algorithm provides efficient approximate Bayesian inference for DP mixture models.
• Existing methods often lack integrated variable selection and may not fully leverage Bayesian model averaging (BMA).

Purpose of the Study:

• To extend the SUGS algorithm for DP mixture models to incorporate variable selection for clustering.
• To demonstrate the advantages of using Bayesian model averaging (BMA) over Bayesian model selection (BMS) in this context.
• To provide a computationally efficient and robust method for analyzing high-dimensional biological data.

Main Methods:

• Extension of the sequential updating and greedy search (SUGS) algorithm to include variable selection.
• Application of Bayesian model averaging (BMA) for improved inference in DP mixture models.
• Implementation in an open-source R package (sugsvarsel) with C++ acceleration and parallel processing.

Main Results:

• The proposed method demonstrates competitive performance against state-of-the-art approaches in simulation studies and cancer transcriptomics examples.
• Significant computational benefits are observed compared to traditional Markov chain Monte Carlo methods.
• Successful application to The Cancer Genome Atlas (TCGA) reverse-phase protein array (RPPA) data for pan-cancer proteomic characterization.

Conclusions:

• The enhanced SUGS algorithm with variable selection and BMA offers a powerful and efficient tool for model-based clustering.
• The method provides valuable insights into complex biological datasets, such as cancer proteomic profiles.
• The open-source R package sugsvarsel facilitates broader adoption and accelerates research in computational biology and statistics.