There are now numerous reports describing the unreliability of commonly used housekeeping genes for the normalization of QPCR data (see Introduction). The expectation of identifying an "ideal" housekeeping gene that is stably expressed under all conditions was not met in this study. Rather, all eight housekeeping genes evaluated here show unstable expression under more than one experimental condition. Nonetheless, the degree of instability is important to note. We find that gapdh is one of the most unstable and elfa and 18s among the most stable genes during development (2 – 120 hpf) and across tissue types. Our results are consistent with the earlier zebrafish study in which five of our eight housekeeping genes were measured between 2 and 72 hpf and in pooled male-female tissues . Based on the agents tested in our treatment series, however, we would eliminate 18s as a suitable housekeeping gene. Instead, our choices for studies using zebrafish embryos to test hormones and endocrine disrupting chemicals that interact with ER and AhR l would be elfa and bactin1.
It is important to note that the rank order of gene stability reported in the present study cannot be generalized, without further testing, to other fish species, tissue types, life stages, or treatment paradigms. An examination of the fish literature reveals many contradictory findings [cited in ]. To illustrate, a QPCR study of eight housekeeping genes in adult fathead minnows reported that 21 d exposure to ethinyl E2 (EE2) markedly suppresses hepatic expression of four genes, including elfa, bactin1, gapdh and g6pd (but not 18s) . A similar study in adult medaka also found a suppression of bactin1 expression in the liver with a 21 d exposure to EE2 but in contrast found the EE2 exposure increased hepatic expression of gapdh . In zebrafish embryos, on the other hand, we find that a 3 d exposure to E2 suppresses g6pd, upregulates 18s but has no effect on elfa or bactin1. Similarly, results showing E2 and EE2 effects on gadph and bactin1 expression in zebrafish by microarray analysis are inconsistent with our QPCR results, indicating that it may be incautious to compare findings using different methods of mRNA analysis [20, 21].
Although changes in housekeeping gene expression in response to EtOH, DMSO or low doses of chemicals appear small, even small differences can add significant error to target gene expression during normalization. The example of cyp19a1b expression following ICI treatment demonstrates the impact on a modest gene response. When housekeeping genes affected by ICI are used for normalization the expected down-regulation is negated. If this situation were to occur when testing a novel chemical, or when using QPCR to verify results of microarray analysis, a real effect could be overlooked. In the case of cyp1a, a gene that is robustly induced by AhR ligands, a difference in housekeeping gene expression results in a more than 3-fold exaggeration in upregulation. This kind of overstatement could have real implications for data interpretation, for example, when comparing dose-response characteristics of different chemical agents, or when screening environmental samples for bioactivity. To avoid unforeseen errors in normalization, for example, by the presence of unknown agents in complex mixtures that affect reference gene expression, the stability of the chosen housekeeping gene can be routinely monitored by recording changes in Ct values.
Given that many of the classical reference genes have proven unreliable [22–28], alternative normalization strategies have been proposed. One typical approach is to normalize to total RNA levels. While this avoids the difficulties of selecting and validating a housekeeping gene, the shortcoming of this method is that it does not control for errors introduced at the reverse transcription step of PCR reactions. In addition, it primarily measures ribosomal RNA (rRNA) whereas qPCR aims to determine mRNA expression. Furthermore, normalization for total RNA assumes that the rRNA: mRNA ratio is the same in all groups, which might not always be the case [2, 29, 30]. Finally, the high abundance of rRNA compared to mRNA makes it difficult to subtract the baseline value in qPCR analysis [31, 32]. For the same reasons, markers of rRNA such as 18s or 28S rRNA might be suboptimal as normalization factors in many settings [2, 32, 33]. Also, as our data shows, it cannot be assumed that rRNAs are stably expressed under all conditions.
Another promising method for normalization is the use of statistical software to determine the most stably expressed gene . By using the geometric average of multiple control genes, geNorm software provides accurate normalization of qPCR data . When geNorm was applied in this study, the recommended housekeeping genes were indeed those with minimal Ct changes and overall stable expression. Other statistical programs (BestKeeper, Norm Finder) have been developed to determine the most appropriate reference gene for a given experimental condition [31, 35]. They use different algorithms to analyze the variation in the expression of reference genes, which could result in different recommendations for the most suitable reference gene. The disadvantage of using statistical programs is that considerable effort and cost is expended to generate data for analysis, but similar constraints apply to any strategy that requires the validation of multiple housekeeping genes.
The intent of this study was to provide a database that helps zebrafish researchers to identify a shortlist of candidate housekeeping genes for specific experiments. For example, although it has been a relatively popular housekeeping gene for zebrafish research, gapdh clearly has large variability in its expression under all experimental conditions tested in our study and so would not be recommended for normalization. Studies by Tang et al in zebrafish  and Filby and Tyler in fathead minnows  also found gapdh to be unsuitable for data normalization. The gene with the least variability across all the conditions assessed in this study was elfa and so may be an appropriate initial selection for normalization.