Introduction

Reference or so-called “housekeeping” genes are widely used as internal standards in mRNA-based techniques such as reverse transcription quantitative real-time polymerase chain reaction (RT-qPCR) [9, 31] and microarray technologies [35]. Current guidelines recommend the use of reference genes over housekeeping genes to refer to internal controls [10]. A reference gene is one that is constitutively expressed in all nucleated cell types, is involved in basal cell metabolism or survival and is defined as a gene showing the most stable expression across various tissue types and experimental conditions. Quantitative mRNA studies are commonly conducted to compare expression levels of a target gene or gene of interest (GOI) between disease and non-disease and such data need to be normalised to a reference gene.

It is essential when conducting such studies to ensure that all the samples contain the same amount of RNA. Biological parameters such as tissue volume, mass, size or cell number were first employed as normalisation procedures (absolute quantification). However, variability in biological samples rendered this method impractical and unreliable [8]. When comparing the amount of RNA, it was found that normalisation with an internal reference gene was more appropriate (relative or comparative quantification), since it compensates for sample-to-sample variations in total RNA amount. Relative quantification compares changes of mRNA levels between two genes with one of them being the endogenous reference (the normaliser) and will determine the expression of a GOI relative to the reference gene. The most commonly used reference genes are GAPDH (glyceraldehyde-3-phosphate dehydrogenase), HPRT1 (hypoxanthine phosphoribosyltransferase 1), albumin, actins, tubulins, cyclophilin, microglobulins, beclin and ribosomal subunits [48, 49]. Although well established, data normalisation using an internal reference gene is subject to frequent criticism since the quality of normalised expression levels is entirely dependent on the quality of the normaliser [26]. Any variations in expression of the normaliser will result in obscuring real changes and will produce artifactual changes [7]. Hence, to carry out appropriate normalisation in RT-qPCR, suitable reference genes are vital to the generation of biologically relevant results [3, 17].

There are increasing concerns about some of the more conventional reference genes when investigating mRNA expression levels in the human brain [26]. For example, GAPDH is mostly known for its role in glycolysis, but recent reports have demonstrated increased GAPDH activity associated with cell fate and neurodegeneration [14, 43]. Furthermore, increased beclin expression has been associated with neurones and astrocytes at the site of traumatic brain injury [12]. A similar scenario is emerging for actin [40]. Some concerns have also been raised when normalising against 18S ribosomal RNA, which is much more abundant than other mRNA species and, therefore, may not be the best to compare with mRNA species of lower abundance [9, 44]. Moreover, we have demonstrated that not all mRNAs in the CNS, 18S ribosomal RNA being one of them, appear to deteriorate in a similar manner with increasing post-mortem delay based on RIN assessment [19]. These studies would suggest that the identification of a universal human reference gene might prove more challenging than initially thought and that, perhaps, novel organ and tissue-specific reference genes might be more appropriate and useful. Indeed, the literature shows that an ideal and universal reference gene remains to be discovered [6, 7, 42, 48]. There are very few investigative reports to date that have been published trying to establish and validate reference genes for quantitative gene expression analysis of human CNS post-mortem tissue [15, 27].

Several studies have been conducted to isolate stably expressed genes using microarray data. Large scale categorisation of reference genes based on microarray data has generated three lists of reference genes widely cited in the literature, Warrington et al. [51], Hsiao et al. [25], and Eisenberg et al. [22, 46]. Zhu et al. [54] generated an up-dated list of reference genes using more advanced arrays with an increased number of probes. Finally, de Jonge et al. [16] conducted a meta-analysis of 13,629 human gene array samples and found a selection of candidate genes showing a stable expression across various human cell types and experimental conditions. Very few studies to date have been conducted using qPCR to investigate stable gene expression in the CNS. Penna et al. [39] found that CYC1 and EIF4A2 showed stable expression in Alzheimer’s disease only over a small collection of pre-selected reference genes available within a commercially available kit.

We have recently carried out an extensive and unique analysis of genome-wide changes of gene expression in human brain tissue from Alzheimer’s disease (AD), amyotrophic lateral sclerosis (ALS), Huntington’s disease (HD), multiple sclerosis (MS), Parkinson’s disease (PD), schizophrenia (SZ) and their respective region-specific controls using the Illumina whole genome HumanRef8 v2 BeadChip (GEO accession number GSE26927). A total of 117 tissue samples were investigated representing various brain regions including the spinal cord. Since RNA samples were available, we were able to validate with RT-qPCR expression levels of candidate reference genes extracted from the microarray data. Herein, we report a validation exercise by identifying the most suitable reference genes that can be used as internal controls for RT-qPCR studies in human CNS post-mortem tissue.

Materials and methods

Tissue samples

Gene expression levels from a total of 117 samples (111 different cases with 6 replicates) were included in this study covering six neurological conditions and six different brain areas with their respective controls (Table 1). Basic clinical and demographic details can be found in supplemental material 1. RNA samples were kept at −80 °C. A subset was used for the quantification of mRNA expression by RT-qPCR. The subset cohort for RT-qPCR included four cases from each neurological condition and five controls (n = 29).

Table 1 Summary table of tissue samples and number of cases (including replicates)
Full size table

Microarray analysis and selection of reference genes

Gene expression analysis was performed on RNA extracted (RNeasy® tissue lipid mini kit; Qiagen Ltd, Crawley, UK) from snap frozen tissues [19] with the Illumina whole genome HumanRef8 v2 BeadChip (Illumina, London, UK) as described previously [21]. RNA concentration and purity was assessed by spectrophotometry (NanoDrop ND1000; NanoDrop Technologies, Delaware, USA). RNA integrity was further assessed using an Agilent 2100 Bioanalyzer and lab-on-a-chip platform technology (Agilent Technologies UK Ltd, West Lothian, UK). Data normalisation was conducted using the Rosetta error models available in the Rosetta Resolver® system v7.0. Gene expression data from Illumina BeadStudio version 3.0 were loaded into the Rosetta Resolver® Gene Expression Analysis System version 7.0 and an error model applied to each expression intensity from each sample. The basis of this is described in detail in Weng et al. [52], and consists of applying an error model to each expression intensity that provides a p value and error term of the conservative estimate of the gene-independent propagated error as well as the largely gene-dependent scattered error. Inter-array normalisation was achieved by scaling probe intensities by a scaling factor derived from the mean intensity of the array with 10 % outlier exclusion. Expression intensities from probes representing the same transcript were combined using an error weighted average. Annotations were up-dated with the Rosetta Resolver system.

To establish the most stable expressed genes across all samples, coefficients of variance (CV; CV = standard deviation divided by the mean and is expressed as a percentage) were calculated for all genes from the normalised log2-transformed expression data as described by de Jongue et al. [16]. Genes were ranked thereafter based on their CV. The maximum fold change (MFC, the ratio of the maximum and minimum) was also taken into account. In addition, genes showing a ceiling effect (saturated) with their expression level were identified when the mean was higher than the maximum minus 2 standard deviations and were consequently not considered. The genes of unknown function were also removed. A Pubmed search (July 2012) was carried out on a selection of candidate genes to verify their non-involvement in any abnormal neurological condition.

Quantification of mRNA expression by RT-qPCR

The two-step RT-qPCR was performed using the QuantiTect® reverse transcription kit, the QuantiTect® SYBR Green kit and with QuantiTect® primer assays (Qiagen) as previously described [19]. Briefly, RT-qPCR experiments were performed using the Mx3000P™ real-time PCR system with software version 4.01 (Stratagene, La Jolla, USA). The QuantiTect® primer assays are listed in Table 3. For each sample, reactions were set up in duplicate with the following cycling protocol, 95 °C for 15 min, 40 cycles with a 3-step program (94 °C for 15 s, 55 °C for 30 s, and 72 °C for 30 s) and a final melting curve analysis with a ramp from 55 to 95 °C. Samples with no reverse transcriptase reaction to test for contaminating DNA and a negative control containing no RNA template were included in each run. To control for variation in RNA levels due to manipulation across samples, the same cDNA stock was used to investigate expression levels of the 12 reference genes and all RT-qPCR assays for a particular gene were undertaken at the same time in a 96-well plate under identical conditions. Different cDNA stocks were used to assess expression levels of target gene with different normaliser.

To determine expression stability of the reference genes, the quantification cycle (Cq; aka threshold cycle), which is the PCR cycle at which an increase in reporter fluorescence from SYBR green dye can be first detected above noise signal, was determined for each sample. The same threshold level was used for all experiments. Mean Cq from duplicates generated by the software were used to calculate the CV for each gene investigated and then, the reference genes were ranked.

Data accession

All data are available online through Gene Expression Omnibus (GEO accession number GSE26927).

Statistical analysis

The following software packages were used; GraphPad Prism 5.01 (GraphPad Software Inc, La Jolla, CA, USA) and Microsoft® Office Excel® 2007 (Microsoft UK Headquarters, Reading, UK). A parametric t test was used to assess group difference and the Pearson correlation test for a relationship between two variables. To compare variance between groups, a Barlett’s test for equal variances test (multiple groups) and an F test (two groups) was used with p < 0.05 suggesting a significant difference in variances between groups. A paired non-parametric test was used to compare gene rankings. To test whether a group of samples came from a Gaussian distribution, a d’Agostino and Pearson omnibus normality test was conducted. For multiple repeated measures comparison we used the Friedman test with Dunn’s multiple comparison test. For a particular gene, gene expression intensity data and RT-qPCR relative quantity data were divided by the mean of the control group and presented as a ratio. This procedure scaled mean gene expression levels of the controls at one. The ratio of data is presented as mean ± standard error of means.

Results

Detection of constant genes using the microarray data

All gene transcripts from the Illumina whole genome HumanRef8 v2 BeadChip (20,589) were ranked according to their coefficient of variance (CV). The CV ranged from 1.11 % (least variance) to 105.78 % (most variance). The conventional reference genes ranked as follows in the microarray dataset: beclin-1 (rank: 361), GAPDH (408), QARS (1,708), actin-β (4,882), β2-microglobulin (5,334), tubulin-β (6,008), albumin (8,376), cyclophilin A (10,457) and HPRT1 (13,249). CYC1 and EIF4A2 [39] were ranked 259 and 938, respectively. An arbitrary cut-off of <2 % for the CV was used to select a set of the most stable genes. From the 170 genes selected, saturated genes and genes of unknown functions were removed, leaving a total of 131 genes, which can be found in supplementary material 2. We conducted a thorough search using PubMed on the genes at the top of the list to retain only those with no known reports (at the time of the search) of abnormal expression in any CNS disorder until we isolated a number of potential candidates. A total of eight candidate reference genes of various functions were retained. Post-mortem delay (PMD) and age at death were taken into consideration since reports suggest that these variables may impact on the expression of certain mRNA species despite extreme care given to convert the same amount of mRNA from all samples [19]. RIN was not considered since the samples in the present study were pre-selected for their “good” and homogeneous RNA quality, which is one of the prerequisites for conducting a microarray experiment. Consequently, they do not represent the full range of tissue quality available within tissue banks as tests showed that RIN values did not come from a Gaussian distribution, unlike PMD (p = 0.0123) and age at death (p = 0.0097). Despite small gene expression variances, significant shifts in expression due to disease were detected for some genes, which we believe may account for our results, especially for GAPVD1 and BECN1 (see later section; Table 2). Finally, variations in expression between brain regions, groups and disease type (Barlett’s tests for equal variances and F test) were considered. One of the homogeneity of variance violations found for the 12 genes as listed in Table 2 was for GAPDH showing a more stable expression in the superior frontal gyrus grey matter compared to other brain regions (Fig. 1 and supplemental material 3a). No consistent variance in expression between control and disease group across disease states were observed for those 12 genes (supplemental material 3b) and no significant changes were observed across diseases (supplemental material 3c).

Table 2 Correlation with post-mortem delay and age at death and group comparison (control vs. disease)
Full size table
Fig. 1
figure 1

Expression variation across brain regions for GAPDH. The Barlett’s test for equal variances was significant for GAPDH (p = 0.0022) meaning that there was a significant effect in expression variability across brain regions (control + disease samples). More precisely, GAPDH expression was more stable in the superior frontal gyrus grey matter (grey matter lesions from MS patients) than any other brain region (refer to Table 1 for the full brain region listing)

Full size image

Comparison of expression between two hybridisation techniques

Expression stability of the eight novel reference genes, in addition to four more commonly used reference genes (GAPDH, BECN1, QARS and TUBB), was verified using RT-qPCR on a subset cohort (n = 29) from the microarray study. We first verified that selecting a smaller subset cohort of the microarray dataset did not significantly affect the ranking of the selected genes. We compared ranking based on CV. Only minor signal expression variations in ranking were detected between the entire and the subset cohort but were not found to be significantly different (Wilcoxon signed rank test). The four common reference genes that showed the least expression stability over 117 samples remained the least stable in the smaller cohort comprised 29 samples. Cq values from RT-qPCR experiment (n = 29) were also ranked based on CV. Finally, a final average rank was given and all rankings can be found in Table 3 and a summary in Table 4.

Table 3 Ranking of candidate novel reference genes and four commonly used reference genes (in bold) based on microarray and RT-PCR data
Full size table
Table 4 Summary ranking table
Full size table

We then compared the stability in expression of those 12 genes between the two hybridisation techniques. Statistically, there was no difference between all 3 ranking orders of those 12 genes nor was any statistical difference in ranking found between the two hybridisation techniques (Wilcoxon matched pairs test). However, some changes in ranking order were noticed. The most striking disparity in expression between the two techniques was for BECN1 and ATP5E. Expression levels for BECN1 with RT-qPCR were very stable and placed BECN1 at position 1 compared to the microarray data, while ATP5E moved from position 2 in the microarray data to position 12 with RT-qPCR. In summary, the novel investigated genes XPNPEP1, OBSP and AARS showed an overall better or comparable expression stability than the commonly used reference genes.

Performance as normalisers of novel reference genes

Since, RT-qPCR represents the gold standard to validate gene expression changes from cDNA microarrays, we assessed the eight top reference genes from the overall ranking (XPNPEP1, AARS, GAPVD1, ATP5E, OSBP, CSNK2B, BECN1 and QARS; in ranking order; see Table 4) as normalisers for a GOI to determine whether they were able to replicate the increased microarray expression levels of the GOI. Transforming growth factor, beta receptor II (TGFBR2) was chosen as the test-gene of interest. This gene was up-regulated in Huntington’s disease by 1.74-fold by microarray analysis. Except for GAPVD1 and QARS, all normalisers were able to reproduce significantly increased levels of TGFBR2 in HD from the microarray, albeit with a different magnitude of change (Fig. 2a). A good correlation in expression levels between the microarray and the RT-qPCR data was found with the exception of GAPVD1 (Table 5). From the novel reference genes, XPNPEP1, AARS, OSBP and CSNK2B replicated most efficiently the fold change of the test-gene of interest. TGFRB2 expression in HD was 1.0-fold higher with BECN1 (mean = 3.22 ± 1.07) and reduced with ATPE5 (mean = 1.6 ± 2.03) as normalisers. The higher expression of TGFRB2 when using BECN1 as normaliser could be explained due to significant variances of mRNA BECN1 levels in the disease group compared to controls, similarly for GAPDV1 (Fig. 3). XPNPEP1 and AARS were retained as will be discussed thereafter. In addition, using another GOI, TNFRSF14 (tumour necrosis factor receptor superfamily, member 14), which was significantly up-regulated (microarray data: ×1.94; p < 0.001) in PD [21], expression levels were compared using BECN1 and XPNPEP1 as normalisers (Fig. 2b). The significantly increased expression levels of TNFRSF14 were more closely matched to the expression levels from the microarray output when XPNPEP1 was used as normaliser (×2.04; p < 0.01) compared to BECN1 (×1.61; p < 0.05).

Fig. 2
figure 2

qPCR data for TGFBR2 (a) and TNFRSF14 (b) using different normalisers. The black dotted lines represent the expression levels of controls. In the first panel a, except for GAPVD1 and QARS, all normalisers were able to replicate significantly (p < 0.05) increased levels of TGFBR2 in disease (×1.74—microarray data). In b, XPNPEP1 (×2.04; p < 0.01) as normaliser produces a closer match to microarray output (microarray data ×1.94; p < 0.001) than BECN1 (×1.61; p < 0.05). **p < 0.01 and *p < 0.05

Full size image
Table 5 Correlations between the PCR expression data of a GOI, TGFBR2 (×1.74), using different normaliser and the microarray data
Full size table
Fig. 3
figure 3

Expression levels of XPNPEP1 (a), GAPVD1 (b) and BECN1 (c). Despite a smaller coefficient of variance of expression for GAPVD1 (b; CV = 1.56) than XPNPEP1 (a; CV = 1.76) across our samples (control = 55 and disease = 62), there was a significant shift toward an increased expression in the disease group not observed with XPNPEP1 where expression was homogeneous irrespective of disease state. A significant shift towards a decrease in expression was found for BECN1 (c). ***p < 0.001 and **p < 0.01

Full size image

Validation using independent datasets

In the first instance, we used two additional microarray datasets, available from our laboratory, to consolidate that XPNPEP1 is stably expressed across samples and compared its expression to more commonly references genes such as BECN1 and GAPDH. XPNPEP1 showed less variance in expression across samples in both datasets (see Table 6). Also, using a different RT-qPCR protocol and different samples, we found that XPNPEP1 expression was more stable measured by qPCR on Cq values (CV = 7.22; mean = 27.30; SD = 1.974) than GAPDH (CV = 10.60; mean = 21.85; SD = 2.317) in tissue from 10 control white matter samples and 12 MS white matter lesions [20]. We also used an external microarray dataset (GSE13162) from human tissue (frontotemporal cortex) samples (Affymetrix Human Genome U133A 2.0 Array) available from Gene Expression Omnibus (GEO) [13]. The CV of individual probes for XPNPEP1, ACTB, BECN1, GAPDH and HPRT1 show that XPNPEP1 expression was more stable than BECN1 but not GAPDH or ACTB (Table 7).

Table 6 Coefficient of variance comparison between XPNPEP1, BECN1 and GAPDH in two additional microarray datasets of human grey matter tissue from several brain regions
Full size table
Table 7 Coefficient of variance of XPNPEP1 in human microarray dataset (GSE13162) [12]
Full size table

Finally, we investigated further whether the use of XPNPEP1 as a novel reference gene could also be applied to RT-qPCR mouse brain studies. From a listing available at the Allen Institute for Brain Science, we selected three mouse microarray datasets. The first dataset (GSE16496) was published by Kasukawa et al. [29] and used the Affymetrix platform (Affymetrix Mouse Genome 430 2.0 Array). The second study (GSE26024) used the Illumina MouseRef-8 v2.0 BeadChip on 21 samples [5] and the third (GSE9566) used again the Affymetrix platform as mentioned above, but, instead of whole tissue samples, gene expression changes in isolated CNS cells, i.e., neurones, astrocytes and oligodendrocytes were compared [11]. A summary of the results can be found in Table 8. Xpnpep1 was well within the lower CV range amongst the selected common reference genes, Actb, Becn1, Gapdh and Hprt (see supplementary material 4 for individual data), which placed Xpnpep1 as an equal or better reference gene than some common ones currently in use for RT-qPCR studies for human, mouse and single cell CNS samples.

Table 8 Coefficient of variance of XPNPEP1 in mouse microarray datasets (summary)
Full size table

Discussion

The pathogenetic mechanisms underlying the major neurodegenerative disorders are still poorly understood. Our recent microarray gene expression studies aimed to look for common pathways of degeneration underlying disease. In doing so, we wanted to overcome some of the shortcomings previously encountered when using microarray technology, such as the use of different microarray platforms, different analytical software tools or inadequate normalisation. By studying expression data from six different neurological conditions together with their respective controls, we have been able to generate a list of the most stably expressed genes within the CNS not affected by on-going disease processes that can now be used as reference genes. We propose herein XPNPEP1 as a more suitable reference gene for the purpose of normalisation for CNS gene expression studies using post-mortem tissues and also offer a list of potential alternatives.

The expression of the XPNPEP1 [X-prolyl aminopeptidase (aminopeptidase P) 1, soluble] gene showed excellent stability in our analysis. XPNPEP1 is localised on chromosome 10q25.3 and can be found in a wide variety of organisms including mammals, yeasts, and bacteria and a variety of tissues including lung, kidney, brain, adrenal medulla and cell types such platelets, lymphocytes, and astrocytes [36]. Northern blot analysis indicates that this gene product is ubiquitously expressed [50]. It shows amino acid homology with Escherichia coli methionine aminopeptidase, a prototypical member of the “pita bread-fold” family [4]. This novel class of proteases remove the N-terminal methionine from proteins and peptides, a crucial step in the maturation of many proteins towards initiating biological activity, adequate subcellular localisation and eventual degradation [37, 38]. The human XPNPEP1 consists of 623 amino acid residues with a calculated molecular mass of 69,886 Da, and has been shown to hydrolyse bradykinin—a small vasoactive peptide involved in a variety of biological processes [36]. Our findings suggest that increased expression levels of a GOI were better replicated with XPNPEP1 as normaliser than with BECN1. Beclin-1 plays a central role in autophagy [28] and has recently been shown to increase in expression in neurones at the site of traumatic brain injury [12]. Also, neurodegeneration was shown to induce up-regulation of beclin-1 [23]. In the mammalian adult brain beclin-1 is expressed in neurones and astrocytes, and up-regulation of beclin-1 has been reported in injured neurones [53] and in a subpopulation of astrocytes reacting to injured neuronal components [18]. Absolute mRNA expression of beclin-1 will be therefore subject to variances depending on total neuronal cell numbers and disease activity. Thus, caution is required in the use of this reference gene in the study of neurodegeneration.

AARS (alanyl-tRNA synthetase) belongs to a family of 20 cellular enzymes, the aminoacyl-tRNA synthetases, responsible for translating nucleotide triplets in genes and mRNAs. These enzymes are involved in a number of functions such as translational fidelity, tRNA processing, RNA splicing, RNA trafficking, apoptosis and transcriptional and translational regulation [41, 47]. No changes in expression of alanyl-tRNA synthetase (AlaRS) have been associated with any neurodegenerative disorders; however, mutation in the AlaRS protein was shown to lead to Purkinje cell degeneration in rodents [34], to peripheral axonal degeneration and cause autosomal dominant Charcot-Marie-Tooth disease [45]. Our preliminary evidence would suggest that AARS was more invariant than QARS (a common reference gene). We have investigated only one member (AARS) of this group of genes. However, other members may also be good candidates. OSBP (Oxysterol binding protein) is thought to act as a sensor of sterols [24] and overexpression of OSBP has been linked to increased cholesterol and sphingomyelin levels [32]. Although there is no known involvement of OSBP in any CNS disorders as yet, cholesterol metabolism alterations have nevertheless been shown to play a role in the pathophysiology of multiple CNS disorders [1, 30] and therefore would not be an ideal reference gene. Finally, the CSNK2B (Casein kinase 2, beta polypeptide) gene is located on chromosome 6p21, the human major histocompatibility complex [2]. Although not associated with any diseases as yet, it is well established that the human major histocompatibility complex is associated with many autoimmune/inflammatory diseases, and therefore was not retained.

In conclusion, the generation of expression profiles for several neurological disorders, representing various CNS regions allowed us to provide a selection of candidate invariant genes with potential to be useful as novel reference genes for appropriate normalisation of qPCR data from human CNS post-mortem tissue studies. By doing so, we overcame some of the shortcomings encountered with microarray studies. We isolated an initial 131 genes from the microarray data using the coefficient of variance over other methods available [49] and verified expression levels from a selection of genes in a sub-cohort of the microarray study using RT-qPCR. We were able to test several candidate genes as normalisers for a target GOI. XPNPEP1 and AARS performed best in replicating gene expression levels from the microarray data using RT-qPCR. XPNPEP1 performed better than BECN1 in being able to replicate microarray data. This impetus in the search for an appropriate reference gene has emerged due to rising concerns about the validity of the commonly used reference genes [8, 9, 33] and the importance of following stringent standards for molecular expression studies [10]. Although it is advisable to test for the most appropriate reference gene in each experimental setting and to use multiple reference genes, it is nevertheless very useful to have one established reliable tissue-specific reference gene, in particular, for single target gene expression investigations using RT-qPCR in a particular CNS disease. We provide herein one validated novel reference gene, XPNPEP1, useful for RT-qPCR studies of the human CNS and a list of other potential candidates.