MoSBi: Automated signature mining for molecular stratification and subtyping

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Increasing biomedical data necessitates advanced computational tools for patient stratification and disease subtyping.
• Unsupervised machine learning, particularly biclustering, can identify patterns in heterogeneous molecular data.
• Existing biclustering algorithms require specific parameterization and show variable performance.

Purpose of the Study:

• To develop an automated, robust biclustering approach for analyzing heterogeneous omics data.
• To integrate multiple biclustering algorithms to overcome individual limitations.
• To provide a tool for improved molecular sample stratification and biological discovery.

Main Methods:

• Developed MoSBi (molecular signature identification using biclustering), an ensemble approach integrating multiple biclustering algorithms.
• Utilized an error model-supported similarity network for result integration.
• Systematically evaluated 11 biclustering algorithms and MoSBi using transcriptomics, proteomics, metabolomics, and synthetic datasets.

Main Results:

• MoSBi achieved robust identification of group and disease-specific signatures across diverse datasets.
• The ensemble approach overcame the specificities and parameter-dependency of single algorithms.
• A scalable network visualization tool was developed to aid biological hypothesis generation.

Conclusions:

• MoSBi provides a robust and accessible automated biclustering solution for molecular sample stratification.
• The ensemble methodology enhances the reliability and generalizability of biclustering results.
• MoSBi facilitates biological discovery through integrated analysis and visualization of omics data.