An integrative multiomics random forest framework for robust biomarker discovery.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• High-throughput technologies generate vast multi-omics data (genomics, transcriptomics, epigenomics, proteomics).
• Integrating these omics layers is crucial for identifying complex molecular interactions missed by single-layer analyses.
• Unsupervised methods are needed to discover biomarkers from integrated omics data.

Purpose of the Study:

• To develop an unsupervised framework for prioritizing shared biomarkers across multiple omics data layers.
• To identify cross-layer molecular hubs and nonlinear dependencies.
• To improve biomarker discovery and disease stratification using integrated omics data.

Main Methods:

• A multivariate random forest (MRF) framework was developed.
• Inverse minimal depth (IMD) importance was used to rank shared biomarkers across omics layers.
• Three IMD-based selection strategies and an optional IMD power transform were implemented.
• The method was evaluated through extensive simulations and applied to TCGA and ADNI datasets.

Main Results:

• MRF-IMD matches linear methods (SPLS/CCA) in linear settings and outperforms them in nonlinear settings.
• Univariate ensemble learners underperformed in the unsupervised, multivariate context.
• Applied to TCGA data, MRF-IMD identified cancer-relevant pathways and improved survival stratification.
• In pan-cancer and ADNI analyses, MRF-IMD features yielded more coherent clusters and better disease-progression stratification.

Conclusions:

• The MRF-IMD framework provides a scalable and interpretable approach for multi-omics biomarker discovery.
• It effectively captures nonlinear, cross-layer dependencies crucial for understanding complex diseases.
• This method advances reliable biomarker identification in integrated omics studies.