Use of signal thresholds to determine significant changes in microarray data analyses

Abstract

The use of a constant fold-change to determine significant changes in gene expression has been widely accepted for its intuition and ease of use in microarray data analysis, but this concept has been increasingly criticized because it does not reflect signal intensity and can result in a substantial number of false positives and false negatives. To resolve this dilemma, we have analyzed 65 replicate Affymetrix chip-chip comparisons and determined a series of user adjustable signal-dependent thresholds which do not require replicates and offer a 95% confidence interval. Quantitative RT-PCR shows that such thresholds significantly improve the power to discriminate biological changes in mRNA from noise and reduce false calls compared to the traditional two-fold threshold. The user-friendly nature of this approach means that it can be easily applied by any user of microarray analysis, even those without any specialized knowledge of computational techniques or statistics. Noise is a function of signal intensity not only for Affymetrix data but also for cDNA array data, analysis of which may also be benefited by our methodology.

Targeted Gene Deletion Using DNA-Free RNA-Guided Cas9 Nuclease Accelerates Adaptation of CHO Cells to Suspension Culture

“…However, in the case of low expression genes, fold change-based analysis generates unacceptably high false positives. 37 To avoid choosing a false positive as an engineering target, the first (adherent) and last (suspension) time points in the RNA-seq data were visualized by MA plots. This revealed that low abundance transcripts appear on the left x-axis, whereas high abundance transcripts are located right x-axis (Figure 2A).…”

“…Conventional analysis of RNA-seq often uses fold change values to describe change in expression of individual genes. However, in the case of low expression genes, fold change-based analysis generates unacceptably high false positives . To avoid choosing a false positive as an engineering target, the first (adherent) and last (suspension) time points in the RNA-seq data were visualized by MA plots.…”

Predicting the targets of IRF8 and NFATc1 during osteoclast differentiation using the machine learning method framework cTAP

“…Likewise, gene transcripts expressed at a higher fold change in the test sample compared to the control sample may not necessarily indicate an upregulation even if such a decision of differential expression was obtained with statistical significance after background noise removed. The article by [46] points out that merely using fold-change to determine significant changes in gene expression does not reflect signal intensity and can result in a substantial number of generating false positives and false negatives. The design behind cTAP is that if we take into account the collective behaviors of related genes in a cohort of gene expression data sets of the same or comparable context, this potential problem of false positives and false negatives could be abated.…”

Blood RNA expression profiles undergo major changes during the seventh decade

“…The probes presenting a FDR < 0.05 were considered statistically significant and kept for further analysis. Due to the large number of probes passing the p -value threshold (Figure 3 and Supplementary Table S2 ), a second filter set a threshold on the absolute value of the fold change (FC) [ 25 ]. There is no consensus on a FC threshold, but literature shows that an absolute FC ≥ 1.2 renders the results more likely to be reproducible [ 16 ].…”