BELMM: Bayesian model selection and random walk smoothing in time-series clustering

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• Advances in omics technologies generate large time-course datasets.
• Clustering analysis is crucial for uncovering structures in these complex datasets.

Purpose of the Study:

• Introduce Bayesian Estimation of Latent Mixture Models (BELMM) for time-series data analysis.
• Provide a flexible Bayesian framework for clustering and modeling time-series data.

Main Methods:

• Utilize mixture modeling with random walk smoothing priors for mean curves.
• Employ Reversible-jump Markov chain Monte Carlo for model selection and determining the number of mixture components.
• Assign time series to clusters based on similarity to latent random walk-derived trends.

Main Results:

• Demonstrate BELMM's application on simulated and real-world omics time-series data.
• Showcase both fast and slow implementations of the BELMM framework.
• Successfully clustered French mortality and Drosophila melanogaster gene expression data.

Conclusions:

• BELMM offers a robust Bayesian approach for time-series clustering.
• The framework is implemented in R and Stan, with code available for reproducibility.
• BELMM facilitates the analysis of complex biological time-course data.