Extraction and integration of genetic networks from short-profile omic datasets

TL;DR

This paper introduces two new similarity measures, Mahalanobis cosine and hybrid-Mahalanobis cosine, for inferring genetic networks from short-profile omic datasets, outperforming traditional methods in accuracy.

Contribution

The study proposes covariance-based similarity measures specifically designed for non-Gaussian omic data, improving genetic network inference from ionomics and metabolomics datasets.

Findings

Covariance-based measures outperform Pearson's correlation in recovering gene associations.

Mahalanobis cosine and hybrid-Mahalanobis cosine are effective for different network scales.

The methods are validated on yeast datasets and curated genetic databases.

Abstract

Mass-spectrometry technologies are widely used in the fields of ionomics and metabolomics to simultaneously profile at the genome scale intracellular concentrations of e.g. amino acids or elements. Short profiles of molecular or sub-molecular features are intrinsically non-Gaussian and may reveal patterns of correlations that reflect the system nature of the cell biochemistry and biology. Here we introduce two profile similarity measures that enforce information from the empirical covariance matrix of the data, the Mahalanobis cosine and the hybrid-Mahalanobis cosine. We evaluate the performance of these similarity measures in the task of inferring and integrating genetic networks from omics data by analysing experimental datasets derived from the ionome and the metabolome of the model organism S. cerevisiae, and several large curated databases of genetic annotations. The proposed covariance-based similarity measures can in general recover known and predicted associations between genes better than the commonly used Pearson's correlation and the standard cosine similarity. The choice of which of the two measures to recommend depends upon whether the focus is on extracting genetic associations at a global or local genetic network scale.