Multiple kernel learning for integrative consensus clustering of omic datasets.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Mining

Background:

• Integrative clustering combines multiple data sources for enhanced analysis, particularly in tumor subtyping.
• Cluster Of Clusters Analysis (COCA) is a widely used technique, but its properties and robustness to noisy data are not well understood.

Purpose of the Study:

• To rigorously benchmark Cluster Of Clusters Analysis (COCA).
• To introduce Kernel Learning Integrative Clustering (KLIC) as a novel, robust alternative for integrative clustering.
• To evaluate the performance of KLIC and COCA in various scenarios, including real-world applications.

Main Methods:

• Framing integrative clustering as a multiple kernel learning problem.
• Developing KLIC to allow weighted contributions from different datasets, down-weighting noisy data.
• Conducting simulation studies and applying methods to cancer subtyping and transcriptional module discovery.

Main Results:

• KLIC demonstrates improved performance by adaptively down-weighting contributions from potentially noisy datasets.
• Benchmarking reveals the limitations of COCA's robustness when faced with data heterogeneity.
• Both KLIC and COCA were applied to real-world cancer subtyping and transcriptional module discovery datasets.

Conclusions:

• KLIC provides a more robust and flexible approach to integrative clustering compared to COCA.
• The developed KLIC method enhances the reliability of analyses that combine information from multiple biological datasets.
• R packages for both KLIC and COCA are publicly available, facilitating their use in research.