Quantification of gene expression in diseased tissue can determine genes which are involved in the development or progression of disease [1], suitable for genomic evaluation [2] or gene based prognostics [3] or treatment [4]. Real-time reverse transcription polymerase chain reaction (real-time RT-PCR) is the most commonly utilised method of measuring gene expression in biological systems. The method provides accurate quantification of transcript number, good sensitivity over a wide range of transcript expression levels, and increasing high throughput capabilities. Several factors contribute to errors of variation in gene expression measurement, including issues relating to sample starting cell number and sample cell types, mRNA extraction protocol and handling techniques [5], mRNA quality [6, 7], method of reverse transcription [8] and analytical detection chemistry method [5].

To accommodate these differences in RNA sample preparation and analysis, the measurement of relative expression of transcript has evolved a means to control these variables employing a process that is termed normalisation [9]. Normalisation of real-time RT-PCR data is classically performed through the selection of a calibrant internal control gene, known as a reference gene or "house-keeping" gene. Conceptually, an ideal gene selected as an internal reference control should have a constant level of expression across the tissue or cell samples used throughout the experiment, and should not exhibit altered expression with diseased or, control tissues, or indeed experimental conditions [10]. Initially, ubiquitously applied reference genes were sought that could be applied across tissue and experimental types [3]. However, recent studies have shown that the expression stability of some of the commonly used reference genes, such as B2M, GAPDH and ACTB is not constant for all tissues or disease states [10, 11].

Current studies identify reference genes that are validated for each tissue or cell type and disease or experiment. These reference genes can be selected by evaluating data from real-time RT-PCR statistical algorithms, such geNorm [11], Global Pattern Recognition [12], Bestkeeper [13], Normfinder [14] or equivalence tests [15]. The principal of the geNorm algorithm is that from an initial group of candidate reference genes tested across all the types of tissue studied and the experimental conditions, the expression ratio of the two reference genes that display the most similar expression identified these genes as the best choice to monitor variation in test gene expression [11]. Global Pattern Recognition is a statistical algorithm which compares the expression or each gene to every other gene used in the comparison, similar to analysis of variance (ANOVA) but with exclusion of nonsensical data (e.g. threshold cycle (C_T) values of 40, where no amplification has taken place) [12]. The Bestkeeper algorithm measures the geometric mean of reference gene crossing point values, to determine the optimal reference gene for use in a samples set. [13]. Equivalence testing is the mathematical determination of the standard deviation of differences in expression values between samples being compared [15]. The Normfinder algorithm uses a model-based approach to the estimation of expression variation, which takes into account variation across sub-groups and avoids the artificial selection of co-regulated genes [14].

Osteoarthritis (OA) is a condition characterised by the destruction of articular cartilage, resulting in pain and dysfunction of the affected joint. OA is a prevalent disease of mammalian joints, which affects up to 20% of the canine population at large [16] with the hip, stifle or elbow joints most commonly affected. The prevalence of OA on radiographic evaluation of dog populations is much higher, with estimates of up to 73% of individuals in a single breed having radiographic evidence of OA of the hip, or a disease (hip dysplasia) which can lead to OA. [17]. OA can be experimentally induced in canine joints through surgical procedures, such as cranial cruciate ligament transaction [18], and OA associated with naturally occurring cranial cruciate ligament rupture is identified in dogs [19].

To date, the majority of molecular research into OA has investigated the mechanisms involved in the catabolism of articular cartilage. However, OA is not solely a disease of articular cartilage, as there are changes in other articular tissues, such as fat [20] and ligaments [21]. The quantification of gene expression in all articular connective tissues, such as cartilage [22], cranial cruciate ligament [23], and infrapatella fat [20] will help to determine the molecular pathogenesis of OA.

In this paper, we identify the best reference genes for use in real-time RT-PCR experiments investigating gene expression in canine articular connective tissue studies. The study draws upon reference genes used in studies evaluating dys-regulation of gene expression in human tissue [11], and employs them to determine if these have similar application in canine studies. The geNorm algorithm is employed to investigate the expression stability of 9 commonly used reference genes (glyceraldehyde-3-phosphate dehydrogenase [GAPDH], beta-actin [ACTB], beta-2-microglobulin [B2M] hydroxymethylbilane synthase [HMBS], hypoxanthine guanine phosphoribosyl transferase [HPRT], ribosomal protein L13a [RPL13A], succinate dehydrogenase flavoprotein subunit A [SDHA], TATA box binding protein [TBP] and tyrosine 3-monooxygenase/tryptophan 5-monooxygenase activation protein, zeta polypeptide [YWHAZ]) [11] in normal and diseased canine connective tissues. We hypothesised that some of the genes selected would demonstrate stable expression in the different canine connective tissues investigated and could therefore be applied as reference genes to normalise future real-time RT-PCR studies that evaluate gene expression in canine on connective tissues.

The gene expression levels (threshold cycle, C_T) for each sample group were averaged with standard deviation was plotted for each group of samples (Figures 1 and 2).

The ideal number of reference genes required to reduce pairwise variation (V) to <0.20 was 4 for canine cell lines, 5 for cartilage, 7 for cranial cruciate ligament and 8 for fat tissue (Figure 3). Addition of further reference genes below the threshold V value of 0.2 did not greatly reduce pairwise variation any further. The increase in V values observed following the hypothetical utilisation of a total of eight reference genes (canine cell lines) or nine genes (fat and cartilage) was consistent with previous reports [11], and has been ascribed to the additional reference gene(s) having the worst expression stability characteristics [11].

A strong correlation between transcript quantity and reference gene stability was only identified for one experiment (LPS stimulation of DH82 cells, R = -0.617). All other experiments where determined to have low correlation between reference gene stability and transcript quantity (r range = 0.238 to -0.280).

At present there is no consensus as to which stability algorithm should be used to optimise reference gene stability. The model based approach is the best for analysing genes whose function is poorly defined (such as new potential genes from microarray data), and therefore may have the potential to be co-regulated as pairwise measurements of stability will artificially tend to select co-regulated genes [14]. Comparisons of the different methods of candidate reference gene selection tend to identify the best and the worst reference genes consistently [24–26] when evaluating reference genes whose functions are well defined. As there is no "gold standard" technique for determining the rank orders of candidate reference genes. We used the geNorm algorithm because in provided both measures of individual gene stability, and the measures of pairwise variation for groups of genes. Furthermore accepted threshold measures of gene stability and pairwise variation are well described with this method [11], and the genes we evaluated have well characterised biological functions, and for whom no co-regulation has previously been reported.

To date, limited information has been published on the selection of appropriate reference genes for use in the quantification of gene expression in mammalian articular connective tissues or canine tissues in general. Reports describing the optimisation of reference genes for use in canine mammary tumours are published [27], as well as those for use in prostate, kidney, mammary gland, and left ventricle of five to nine dogs with or without undefined diseases [28]. We investigated the stability of reference genes in a number of different tissues affected by disease (i.e. tissues from which meaningful comparisons of gene expression would be made), and therefore for which the comparison and optimisation of different reference genes is important.

The expression levels of different reference genes (elongation factor 1-alpha, GAPDH, actin) have been compared in human articular cartilage using a crudely quantitative method, and displayed that the expression level of each reference gene was raised in OA articular cartilage [29]. To our knowledge, only one of the reference genes (B2M) we evaluated to be consistently expressed has been reported as being differentially expressed in OA articular cartilage [30]. Yet despite the lack of information on suitable reference gene selection, the majority of studies quantifying gene expression in connective tissues use GAPDH as the reference gene. GAPDH did not demonstrate stable expression across all tissues in this study, or in previous studies of reference gene stability in different canine tissues [27, 28]. Indeed, no single reference gene was identified as having stable expression when analysing the data from normal and diseased tissues separately and combined, which is also consistent with previous reports evaluating canine [28] and human [11] tissue.

The relationship between the stability value of a reference gene and the C_T value of that gene was measured, as genes which are less abundantly expressed will be more susceptible to errors in measurement due to small variations from the calculated efficiency values. However, we found the stability of individual reference genes (M Value) appears to be unrelated to the level of expression (C_T) which was consistent with a previous report [28], with the exception of one experiment which was probably a type II statistical error.

The justification of using the selected genes as reference genes in canine tissue is based on published work defining their stability as for use as reference genes with human tissue [11, 31], or canine tissue [28]. The function of these genes in canine cells is assumed to be the same as that reported for humans. The expression stability of the two sets of reference genes within the canine cell lines was evaluated, using LPS stimulation to mimic biological variation seen between normal and diseased tissue. Interestingly, the expression stability of the reference genes in the cell lines was much higher than those reported with tissue samples, (lower M values shown in Table 1) which reflects the benefits of working with cell cultures compared to clinical tissue specified. Cell cultures should allow more control over the heterogeneity of cell type, sample handling, storage and mRNA recovery, thereby minimising the level of degradation frequently identified with pathological clinical tissue [32], which may in turn affected gene expression measures [6].

A pair wise variation (V) of 0.15 has been recommended as an arbitrary cut-off point below which the inclusion of additional reference genes expression was not required [11], although this degree of reference gene expression stability could not be achieved for fat tissue or cranial cruciate ligament. If the selected V value was increased to 0.2, then pairwise stability could be achieved using between 4 genes (cell lines) and 8 genes (fat tissue). Alternative measures of disease status, such as histological grading [33] to further select tissue specimens in comparable stages of disease may have helped reduce the variability in reference gene stability, and thus the number of reference genes required for each experiment.

The selection of candidate genes based on ontological function, combined with the evaluation gene expression microarray experiments [25] may reduce the number of reference gene required to obtain acceptable expression stability between samples, as clearly the number of genes suggested are too large to be used in practice. However, until these have been identified, this experiment gives an indication of which genes can be used for the normalisation of gene expression measures in canine OA tissues, and which genes are not suitable for normalising real-time RT-PCR data from fat (ACTB) or cartilage (YWHAZ). Ultimately, we found that reference gene optimisation has to be performed on a tissue and disease specific basis, even when evaluating tissue from the same or similar diseases.

No single reference gene was identified as having stable expression in all canine articular tissue types. The combinations of reference genes required to demonstrate stable expression in each tissue were also identified. The number of genes required to achieve stability for comparing normal and diseased infrapatella fat and cranial cruciate ligament were inappropriate for routine application. This study underlines the necessity to select reference genes on the basis of tissue and disease specific expression evaluation and highlights the requirement for the identification of new reference genes with greater expression stability for use in canine articular tissue gene expression studies.