Statistical modelling of an outcome variable with integrated multi-omics.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics
• Metabolomics

Background:

• Dimensionality reduction is crucial for multi-omics data analysis.
• Univariate and multivariate methods exist for omics data integration.
• Multivariate approaches offer advantages in capturing joint structures and reducing noise.

Purpose of the Study:

• To describe and evaluate univariate and multivariate methods for multi-omics data integration.
• To compare the performance of these methods in outcome modeling.

Main Methods:

• Description of one univariate and two multivariate methods.
• Performance evaluation using simulations with correlated multivariate normal and categorical omics datasets.
• Assessment using root mean squared error (RMSE) for outcome modeling.

Main Results:

• Multivariate methods generally perform well, especially with more integration components.
• Multivariate methods outperform univariate methods with two normal omics datasets.
• Comparable performance observed with one normal and one categorical dataset.
• Similar performance across methods in real-world metabolomics and metabolomics-genetic data for body mass index modeling.

Conclusions:

• Multivariate methods are valuable for summarizing multi-omics data into low-dimensional components for outcome modeling.
• These methods offer a promising alternative to high-dimensional univariate approaches, even with non-normal data.