Estimating differential expression from multiple indicators

Abstract

Regardless of the advent of high-throughput sequencing, microarrays remain central in current biomedical research. Conventional microarray analysis pipelines apply data reduction before the estimation of differential expression, which is likely to render the estimates susceptible to noise from signal summarization and reduce statistical power. We present a probe-level framework, which capitalizes on the high number of concurrent measurements to provide more robust differential expression estimates. The framework naturally extends to various experimental designs and target categories (e.g. transcripts, genes, genomic regions) as well as small sample sizes. Benchmarking in relation to popular microarray and RNA-sequencing data-analysis pipelines indicated high and stable performance on the Microarray Quality Control dataset and in a cell-culture model of hypoxia. Experimental-data-exhibiting long-range epigenetic silencing of gene expression was used to demonstrate the efficacy of detecting differential expression of genomic regions, a level of analysis not embraced by conventional workflows. Finally, we designed and conducted an experiment to identify hypothermia-responsive genes in terms of monotonic time-response. As a novel insight, hypothermia-dependent up-regulation of multiple genes of two major antioxidant pathways was identified and verified by quantitative real-time PCR.

Figure 1.

(A) Power analysis of Fisher’s exact test. For each combination of input parameters, the average of 1000 simulations is plotted. (B) Distribution of gene-specific probe counts. (C) Performance analysis based on DE estimates from simulated gene-expression data. For each combination of analysis parameters the average of 100 simulations is plotted.

Figure 2.

Performance analysis of DE analysis pipelines on the MAQC data. Performance of each pipeline is presented as a radial plot, which includes results from three microarrays (A) or two distinct RNA-seq datasets (B) based on six complementary performance indicators. Abbreviations: AUC, MCC, TPR, FPR, TNR and FNR.

Figure 3.

Performance analysis of DE-analysis pipelines on data from mouse embryonic fibroblasts exposed to 1% O₂ for 24 h. The plot depicts the combined detection rate of GO categories ‘cellular response to hypoxia’ (GO:0071456) and ‘glycolysis’ (GO:0006096) as indicators of hypoxia response. The data is plotted as mean ± standard error of all possible comparisons between subsets of size N of the hypoxic and normoxic groups (original N = 4).

Figure 4.

Large-scale analysis of temporal dynamics of transcription in mouse embryonic fibroblasts exposed to mild hypothermia. (A and B) Temporal profiles of probe expression levels of selected genes during mild hypothermia (A) and normothermia (B). The solid blue line indicates a linear fit to the data points and the gray shadowing represents standard error of the fit.

Figure 5.

Temporal expression profiles of selected genes during mild hypothermia (32°C) and normothermia (37°C) as reported by quantitative real-time PCR. Expression level of Ywhaz was used as reference. The mean of three replicates and standard error has been plotted.