Combinatorial Mixtures of Multiparameter Distributions: An Application to Bivariate Data

Area of Science:

• Statistics
• Computational Biology
• Bioinformatics

Background:

• Mixture models are widely used for statistical inference.
• Existing models often struggle with multidimensional parameters and flexible component sharing.
• Analyzing complex biological data, such as cancer subtypes, requires sophisticated modeling approaches.

Purpose of the Study:

• To introduce combinatorial mixtures, a novel class of models for flexible inference on mixture distributions.
• To develop Bayesian inference and computational methods for these models.
• To apply the models to analyze molecular subtypes of lung and prostate cancers.

Main Methods:

• Introduced combinatorial mixtures allowing shared multidimensional parameters across components.
• Developed Bayesian inference and computational algorithms for parameter estimation and model selection.
• Applied the models to publicly available cancer genomics data.

Main Results:

• Demonstrated the flexibility and parsimony of combinatorial mixtures.
• Successfully applied the models to identify distinct molecular subtypes in lung and prostate cancer data.
• Showcased the unification of component-specific parameter inference and component number inference.

Conclusions:

• Combinatorial mixtures offer a powerful and flexible framework for mixture model inference.
• The developed methods are effective for analyzing complex biological data, including cancer subtypes.
• This approach provides a unified way to handle multidimensional parameters and model complexity.