Tellipsoid: Exploiting inter-gene correlation for improved detection of differential gene expression

TL;DR

Tellipsoid leverages inter-gene correlation in microarray data to enhance differential gene expression detection, significantly increasing statistical power by sharing information across tests using Mahalanobis distance.

Contribution

The paper introduces a novel method that exploits inter-gene correlation to improve differential analysis, outperforming existing approaches in power.

Findings

Outperforms published methods in prostate cancer data

Uses Mahalanobis distance for optimality

Increases statistical power in gene detection

Abstract

Motivation: Algorithms for differential analysis of microarray data are vital to modern biomedical research. Their accuracy strongly depends on effective treatment of inter-gene correlation. Correlation is ordinarily accounted for in terms of its effect on significance cut-offs. In this paper it is shown that correlation can, in fact, be exploited {to share information across tests}, which, in turn, can increase statistical power. Results: Vastly and demonstrably improved differential analysis approaches are the result of combining identifiability (the fact that in most microarray data sets, a large proportion of genes can be identified a priori as non-differential) with optimization criteria that incorporate correlation. As a special case, we develop a method which builds upon the widely used two-sample t-statistic based approach and uses the Mahalanobis distance as an optimality criterion. Results on the prostate cancer data of Singh et al. (2002) suggest that the proposed method outperforms all published approaches in terms of statistical power. Availability: The proposed algorithm is implemented in MATLAB and in R. The software, called Tellipsoid, and relevant data sets are available at http://www.egr.msu.edu/~desaikey