Clustering and variable selection evaluation of 13 unsupervised methods for multi-omics data integration.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• High-dimensional omics data generation is advancing rapidly.
• Integrating diverse omics datasets presents statistical challenges like limited sample sizes and data heterogeneity.
• Existing tools for multi-omics integration include canonical correlation analysis and matrix factorization.

Purpose of the Study:

• To compare the performance of 13 unsupervised methods for multi-omics data integration.
• To evaluate methods' ability to identify subgroups and key variables driving clustering.
• To assess method robustness across simulated and real-world heterogeneous datasets.

Main Methods:

• Comparative analysis of 13 unsupervised multi-omics integration methods.
• Evaluation using eight simulation benchmarks for clustering and variable selection.
• Application of selected methods to three real-world heterogeneous multi-omics studies.

Main Results:

• SNF, MoCluster, CIMLR, LRAcluster, ConsensusClustering, and intNMF showed strong clustering performance.
• MoCluster demonstrated superior performance in variable selection.
• Method performance varied with dataset heterogeneity, impacting MCIA, intNMF, and iClusterPlus.
• MOFA did not yield results on the simulation benchmarks.

Conclusions:

• MoCluster is identified as the most effective method for analyzing heterogeneous multi-omics data.
• The choice of integration method should consider dataset characteristics, particularly heterogeneity.
• An R package, CrIMMix, is available for reproducing the study's findings.