Phenotype prediction using biologically interpretable neural networks on multi-cohort multi-omics data.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Integrating multi-omics data is crucial for advancing precision medicine.
• Interpretable predictive models are needed to understand complex biological mechanisms.
• Current methods often lack transparency in multi-omics data analysis.

Purpose of the Study:

• To develop interpretable neural networks, termed "visible networks," for multi-omics data analysis.
• To assess the performance, interpretability, and generalizability of visible networks for predicting traits.
• To compare multi-omics visible networks with single-omic networks.

Main Methods:

• Utilized genome-wide RNA expression and CpG methylation data from the BIOS consortium (N=2940).
• Employed neural networks informed by prior biological knowledge to create visible networks.
• Performed cohort-wise cross-validation to evaluate diagnostic performance and interpretation consistency.

Main Results:

• Visible networks achieved high accuracy in predicting smoking status (AUC=0.95) and inferring age (error=5.16 years).
• Multi-omics visible networks demonstrated improved performance, stability, and generalizability over single-omic networks.
• Identified key genes (e.g., AHRR, GPR15, COL11A2) consistently predictive across cohorts.

Conclusions:

• Visible neural networks offer a powerful, interpretable approach for integrating multi-omics data.
• These models provide novel insights into biological mechanisms underlying traits and diseases.
• Visible networks show significant potential for enhancing precision medicine through robust multi-omics analysis.