Beta Distribution Network Modeling Improves Biological Integration of Multi-Omics Data

TL;DR

A new network model using Beta distribution improves clustering of multi-omics data, revealing biologically meaningful cross-platform associations.

Contribution

A novel Beta distribution-based network model enhances biological integration of multi-omics data by standardizing correlations and identifying outlier relationships.

Findings

The Beta distribution method outperformed WCNA in clustering accuracy and purity on synthetic data.

The approach achieved better clustering of proteins, metabolites, and lab analytes in real human cohort data.

Higher performance was observed at higher signal-to-noise ratios and fewer reference clusters.

Abstract

Correlation based clustering of multi-omics data typically results in platform-specific modules rather than biologically meaningful cross-platform associations, as technical variation and data structure differences between omics platforms dominate the correlation patterns.We introduce a novel network model that fits a Beta distribution to analyte-analyte correlations. First, correlations across different platforms are standardized by aligning Beta distribution shapes uniformly. Our approach then constructs an analyte relationship network by identifying outlier correlations against a null model background. We generated synthetic data which mimics realistic multi-omics profiles, where intra-omics correlations are generally stronger than inter-omics correlations. Next, we evaluate clustering performance using synthetic datasets, comparing our method to the standard Weighted Correlation Network Analysis (WCNA), which constructs networks based on a power-law distribution. Our method demonstrated higher clustering accuracy (p = 1.3e-13) and purity (p = 2.22e-16) compared to WCNA. The improvement was particularly notable at higher signal-to-noise ratios and fewer reference clusters. We further tested our method on human multi-omics cohort data (Arivale Wellness Study), our strategy outperformed the standard WCNA, achieving a higher average silhouette score and improved entropy, leading to better clustering of proteins, metabolites, and laboratory analytes. This approach is powerful for integrating multimodal molecular data in a biologically meaningful way, enabling the discovery of cross-platform relationships that reveal functional connections across biological layers