Significance tests for comparing digital gene expression profiles

TL;DR

This paper introduces two exact significance tests, one frequentist and one Bayesian, for comparing digital gene expression profiles, especially effective for low expression tags, improving accuracy over traditional asymptotic methods.

Contribution

Develops two novel exact significance tests for digital gene expression comparison, addressing limitations of asymptotic tests for low counts and variable sample sizes.

Findings

Tests outperform chi-square for low frequency tags

Frequentist test uses tag-specific significance levels

Bayesian and frequentist tests are mathematically linked

Abstract

Most of the statistical tests currently used to detect differentially expressed genes are based on asymptotic results, and perform poorly for low expression tags. Another problem is the common use of a single canonical cutoff for the significance level (p-value) of all the tags, without taking into consideration the type II error and the highly variable character of the sample size of the tags. This work reports the development of two significance tests for the comparison of digital expression profiles, based on frequentist and Bayesian points of view, respectively. Both tests are exact, and do not use any asymptotic considerations, thus producing more correct results for low frequency tags than the chi-square test. The frequentist test uses a tag-customized critical level which minimizes a linear combination of type I and type II errors. A comparison of the Bayesian and the frequentist tests revealed that they are linked by a Beta distribution function. These tests can be used alone or in conjunction, and represent an improvement over the currently available methods for comparing digital profiles.