Comparative analysis of integrative classification methods for multi-omics data.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Genomics

Background:

• Advances in omics technologies allow multi-modal data generation from single samples.
• Holistic data integration is crucial for identifying biomarkers and understanding disease mechanisms for precision medicine.

Purpose of the Study:

• To compare supervised multi-omics integration methods for biomarker discovery.
• To evaluate classification performance on simulated and real-world medical datasets.

Main Methods:

• Compared six intermediate integrative methods (matrix factorization, multiple kernel learning, ensemble learning, graph-based methods).
• Used random forest on concatenated and separated data as non-integrative controls.
• Evaluated methods on simulated data (15 scenarios) and real-world data (infectious diseases, oncology, vaccines).

Main Results:

• Integrative approaches performed comparably or better than non-integrative methods on real-world data.
• DIABLO and random forest alternatives outperformed other methods across most simulation scenarios.
• Performance was assessed across various parameters like sample size, dimensionality, and class imbalance.

Conclusions:

• Supervised multi-omics integration methods, especially DIABLO, offer robust classification performance.
• The study provides guidelines for selecting and applying these methods in medical research.
• This work contributes to establishing a gold standard for supervised multi-omics integration.