Bayesian Multi-View Clustering given complex inter-view structure.

Area of Science:

• Computational biology
• Data science
• Statistical modeling

Background:

• Multi-view datasets are common in biology, offering complementary information but posing analytical challenges.
• Existing clustering methods struggle with complex relationships, missing data, and varying sample sizes across views.
• Standard approaches often assume simple cross-view relationships and similar clustering structures.

Purpose of the Study:

• To develop a flexible Bayesian multi-view clustering (BMVC) approach for complex, heterogeneous datasets.
• To handle intricate many-to-many relationships between entities across different data modalities.
• To jointly infer view-specific clusterings while allowing mutual information flow and estimating inter-view dependence.

Main Methods:

• Proposed a probabilistic graphical model for Bayesian multi-view clustering (BMVC).
• Incorporated many-to-many relationships between entities across views.
• Developed methods to estimate the strength of relationships between views, moderating dependence constraints.

Main Results:

• BMVC accurately estimated inter-view dependence in simulated data with non-one-to-one relationships.
• Demonstrated interpretable inter-view structure in a public health survey dataset.
• Improved biological homogeneity of clusters in multi-omic breast cancer data, generating novel hypotheses.

Conclusions:

• BMVC effectively leverages complex inter-view structure for higher-quality clustering compared to standard methods.
• BMVC is a valuable tool for real-world data integration, discovery, and hypothesis generation.
• The approach enhances understanding of complex biological systems through integrated multi-view analysis.