A first consideration in the analysis of our results is that the prospective reference genes that we have analyzed had already been selected in previous studies because of their utility for this function [25, 28, 30, 32], they have relative stable expression and we did not expect large differences between them. Also, these genes are from different functional families and not known to be coregulated. This implies that each one provides independent and complementary information, which is an important requisite for geNorm analysis. A second important consideration is that we have made efforts to minimize every known source of experimental variation by DNAse digestion, adjusting the amount of input RNA, using two-step reverse transcription polymerase chain reaction (RT-PCR) and by correcting raw results for PCR efficiency [19–21, 24, 33]. These steps are a requisite for the assumption that observed results reflect true gene expression. We found specially necessary to include a DNAse digestion step because human genome has many processed pseudogenes inserted by retrotransposition that are amplified in PCR even with intron-spanning primers .
The systematic selection of the best reference genes for real-time qPCR has been approached with different methods. All of them look for stability of the expression levels, by either absence of differences between clinically relevant groups , or relative stability in relation to other reference genes [30, 34] or in relation to clinically relevant groups . These analyses have been facilitated by the free availability of programs and by the description of their principles and use [30, 32, 34]. There are no definitive reasons to prefer one method over the others as their relative strengths depend on the circumstances and we have used three of the best-grounded: a conventional statistical test to compare clinically relevant groups and the geNorm and NormFinder softwares. Results from the three were broadly similar, though the conventional statistical analysis lacked sensitivity. The independence of our results from the analysis method gives credence to the conclusions.
The most striking result was the poor performance of some commonly used reference genes. A special case was GAPDH that is widely used in many areas of research  and is one of the best reference genes in many tissues . Nevertheless, there have been also previous examples of this gene leading to wrong results due to its lack of stability in specific experimental conditions . In our study, GAPDH was not among the best reference genes in any of the analyses done. Other two commonly used reference genes, ACTB and 18S RNA , performed better in our tests but they were not among the more stable genes in most comparisons. These results confirm, once more, the need to evaluate the reference genes in each experimental setting. A particularly striking example in this regard, is the contrast between our results and the reported in prostate cancer tissue, where HPRT1 was the most stable gene, and RPL13A and ACTB were the most unstable . In our experiments, their ranks were reversed, i.e. HPRT1 was the most unstable and RPL13A one of the most stable genes.
Best reference genes in articular cartilage from elderly subjects were among the less commonly used: TBP, RPL13A and B2M. Best results will be obtained by combining two or three reference genes as emphasized by several authors [30, 32]. We propose that for general studies of cartilage from elderly subjects a combination of TBP and RPL13A could be a good starting point, with the inclusion of B2M if practical. For specific comparisons other combinations could be more appropriate.
Finally, it is necessary to take into account some limitations of our study. First, we have included a limited array of prospective reference genes. Other genes have been proposed for use in qPCR, and it is possible that some of them are better candidates for articular cartilage studies in elderly subjects. Microarray data from cartilage, that now start to be published [5, 10] will provide clues for the identification of the best candidates. Second, our results only apply directly to articular cartilage with a focus in OA of large joints. In particular, collection of samples from surgical procedures dictated that all donors were older than 60 years and that OA samples were of an advanced disease stage. This mimics most of the studies in cartilage in OA. However, it is unclear how well our results could be extended to other joint areas, patients with different ages or OA at early stages. Nevertheless, our study can serve as a guide for any kind of cartilage study, and reference genes could be used once tested for low M values. It is also unclear to what extent results obtained with SYBR Green quantification will be applicable to other relative qPCR techniques.