A framework for block-wise missing data in multi-omics.

Area of Science:

• Multi-omics data integration
• Bioinformatics
• Statistical modeling

Background:

• High-throughput technologies generate vast omics data, necessitating integration for improved predictive models and biomarker discovery.
• Managing multi-omics data presents challenges including heterogeneity, noise, high dimensionality, and particularly block-wise missing data patterns.

Purpose of the Study:

• To address challenges of high dimensionality and block-wise missing data in multi-omics integration.
• To develop and evaluate a regularization and constrained-based approach for robust multi-omics analysis.

Main Methods:

• Implementation of a regularization and constrained-based methodology in the R package 'bwm'.
• Application to binary classification and continuous response variable tasks.
• Validation on breast cancer and exposome multi-omics datasets, including scenarios with extensive missing data.

Main Results:

• The proposed model achieves strong performance (86-92% accuracy, 68-79% F1 for classification; 0.72-0.76 correlation for regression) even with missing data across all omics.
• Performance shows a slight decline with increasing missing data percentages, but surpasses single-omic missing data scenarios when block-wise missingness affects multiple omics.
• Feature selection consistency is observed across different omics, with block-wise missing data potentially enhancing model robustness through diverse observation profiles.

Conclusions:

• The developed R package 'bwm' effectively handles high dimensionality and block-wise missing data in multi-omics integration.
• The approach demonstrates robustness and strong predictive performance, offering a valuable tool for biomarker discovery and analysis of complex biological systems.
• Further investigation into the mechanisms driving improved performance in multi-omic missing data scenarios is warranted.