Removal of unwanted variation reveals novel patterns of gene expression linked to sleep homeostasis in murine cortex

Abstract

Background: Why we sleep is still one of the most perplexing mysteries in biology. Strong evidence indicates that sleep is necessary for normal brain function and that sleep need is a tightly regulated process. Surprisingly, molecular mechanisms that determine sleep need are incompletely described. Moreover, very little is known about transcriptional changes that specifically accompany the accumulation and discharge of sleep need. Several studies have characterized differential gene expression changes following sleep deprivation. Much less is known, however, about changes in gene expression during the compensatory response to sleep deprivation (i.e. recovery sleep).

Results: In this study we present a comprehensive analysis of the effects of sleep deprivation and subsequent recovery sleep on gene expression in the mouse cortex. We used a non-traditional analytical method for normalization of genome-wide gene expression data, Removal of Unwanted Variation (RUV). RUV improves detection of differential gene expression following sleep deprivation. We also show that RUV normalization is crucial to the discovery of differentially expressed genes associated with recovery sleep. Our analysis indicates that the majority of transcripts upregulated by sleep deprivation require 6 h of recovery sleep to return to baseline levels, while the majority of downregulated transcripts return to baseline levels within 1-3 h. We also find that transcripts that change rapidly during recovery (i.e. within 3 h) do so on average with a time constant that is similar to the time constant for the discharge of sleep need.

Conclusions: We demonstrate that proper data normalization is essential to identify changes in gene expression that are specifically linked to sleep deprivation and recovery sleep. Our results provide the first evidence that recovery sleep is comprised of two waves of transcriptional regulation that occur at different times and affect functionally distinct classes of genes.

Keywords: Circadian; Gene; Microarray; Sleep; Sleep deprivation; Transcriptomics; mRNA.

Fig. 1

Integrated analysis of the effect of 6 h of sleep deprivation in the murine cortex. a Scatterplots of first two principal components (PC, log-scaled, centered intensities) following RMA and RUV normalization. Percent variance explained by each PC in parenthesis. Triangles denote samples from the Maret et al. 2007 study [2] (whole brain), square samples from the Mackiewicz et al. 2007 study [3] (cortex), rhombus the samples from Vecsey, Peixoto et al. 2012 [8] (hippocampus), and circles are samples from this study (cortex). In green, samples following 6 h of sleep deprivation (SD); in orange, time of day matched controls (CC). Samples cluster according to array platform (PC1) and lab (PC2) following Quantile normalization. After applying RUV, samples cluster according to treatment (PC2). b Distribution of unadjusted p-values for tests of differential expression between SD and CC samples following Quantile and RUV normalization. The distribution of p-values following Quantile normalization is not uniform and biased towards 1. RUV returns uniformity to the p-value distribution and increases discovery of differentially expressed genes (genes that have a low p-value). c Volcano plot of differential expression (−log10 p-value vs log fold change) of Quantile and RUV normalized samples. Genes with an FDR <0.01 are highlighted in blue. Positive controls are circled in red; RUV increases the detection of known differentially expressed genes from 0 to 100 %. PCA plots were performed using the R/Bioconductor package EDASeq (v. 2.0.0). RUVs normalization was performed using the R/Bioconductor package RUVSeq (v. 1.0.0). Differential expression analysis was performed using R/Bioconductor package limma

Fig. 2

Analysis of the effect of sleep deprivation and subsequent recovery sleep in the murine cortex. a Scatterplots of first two principal components (log-scaled, centered counts) following RMA and RUV normalization. Percent variance explained by each PC in parenthesis. Triangles denote samples collected following sleep deprivation (SD), circles control samples matched by time of day (CC, number indicates ZT time) and squares samples collected after recovery sleep (RS). Samples only cluster according to treatment after RUV normalization. b Distribution of unadjusted p-values for tests of differential expression between one hour of recovery sleep (RS1) and control samples following RMA and RUV normalization. RUV increases discovery of differentially expressed genes (genes that have a low p-value). c Volcano plot of differential expression (−log10 p-value vs log fold change) of RMA and RUV normalized samples. Genes with an FDR <0.01 are highlighted in blue. Positive controls are circled in red; RUV increases the detection of known differentially expressed genes following recovery sleep. PCA plots were performed using the R/Bioconductor package EDASeq (v. 2.0.0). RUVs normalization was performed using the R/Bioconductor package RUVSeq (v. 1.0.0). Differential expression analysis was performed using R/Bioconductor package limma. The analysis was based only on samples were collected from current laboratory

Fig. 3

The effect of data normalization in differential expression following recovery sleep. a Bar graph displaying the number of up and downregulated probesets detected at each time point relative to time-of-day matched controls following RMA (light grey) or RUV (dark grey) normalization. RUV normalization profoundly affects the detection of differentially expressed genes following various lengths of recovery sleep. Positive controls for genes differentially expressed following 1 h of recovery sleep after were obtained as detailed in Methods. b Heatmap of differentially expressed probesets detected using RUV normalization relative to circadian time-matched controls. In red, upregulated genes. In green, downregulated genes. Clustering based on patterns of gene expression is represented by the dendrogram and color coded. Genes responding within 1–3 h to recovery sleep are indicated by black bars (fast responders), while genes that respond at 6 h are indicated by grey bars (slow responders, grey bar). Genes upregulated by sleep deprivation show two different patterns of response within the first three hours (red and pink clusters, black bar) and two different patterns of recovery at 6 h (green and orange clusters, grey bar). The majority of genes upregulated by sleep deprivation respond slowly with recovery sleep. The majority of genes downregulated are fast responders (mint green, black bar), while a very small proportion recovers within 6 h (lilac, grey bar). SD, sleep deprivation; RS1, sleep deprivation followed by 1 h of recovery sleep; RS2, sleep deprivation followed by 2 h of recovery sleep; RS3, sleep deprivation followed by 3 h of recovery sleep; RS6, sleep deprivation followed by 6 h of recovery sleep

Fig. 4

Patterns of gene expression regulation during recovery sleep. Plots of log-fold change (logFC) of differential expression relative to controls versus time since sleep deprivation. Color coding corresponds to clusters on Fig. 3b. Plots for genes representative of different expression patterns are shown as dashed lines. The ordering of the gene names within genes of the same cluster (same color) reflects the ordering of the plots. The average and standard deviation for each cluster is shown as solid lines and shaded area. Patterns of expression are divided in two classes: ‘fast’ responders (genes that reach basal values within to 1–3 h of RS) and ‘slow’ responders (genes that reach basal values >3 h). Time constants for the change during recovery sleep for average of each cluster are shown

Fig. 5

Enriched functional clusters regulated by recovery sleep. a Functional clusters regulated by 1–3 h of recovery sleep. b Functional clusters regulated by 6 h of recovery sleep. Functional annotation terms from the following databases: Gene Ontology (GO) biological process and molecular function, KEGG pathways and protein information resource keywords, were clustered based on similarity using the Database for Annotation Visualization and Integrated Discovery (see Methods). Clusters of functional terms enriched in down- or upregulated gene lists following SD as compared with the genome as a whole (P value <0.05) are represented as bars. Height of bars represents the enrichment score of each cluster, with the scores of downregulated clusters shown as negative numbers for visualization purposes. Enrichment score was calculated as − log(10) of the geometric mean p-value among all clustered terms. Only clusters with enrichment score >1.5 (average p-value of functional terms within the cluster <0.05) were considered. For details of the functional terms included in these clusters, as well as enriched functional terms that did not cluster with other terms, see Additional file 7