Although the use of FISH is now widespread and its application varied, the method has proven limitations. Bacterial counts obtained by FISH with a universal probe seldomly equal the counts obtained with nucleic acid labeling. Our literature review shows that on average, 56% of the all cells present in a sample is detected using the probe EUB338. Furthermore, the detection of target cells with FISH varies enormously, from 1% to 100%. Although this range reflects the variability across 51 studies covering a wide range of study sites and methodological differences, substantial variability has also been reported within individual studies (e.g. [1,4,25]). This variability is often reported, but not fully investigated. In the sections below, we discuss the main factors that seem to contribute to this variability, as well as other factors that were not significant in the statistical analysis, but are nevertheless relevant.
4.1Methodological factors linked to FISH performance
The type of fluorochrome bound to the probe explained a significant portion of the variation in %Eub cells detected using FISH. In the early nineties, the most commonly used dyes for FISH in microbiology were fluorescein and rhodamine-derivates (e.g. FITC, FLUOS and TRITC; expending formula in Table 1). Their low fluorescence intensity per mole fluorochrome, and their sensitivity to pH and bleaching limits their application, especially in the case of cells with low abundance of rRNA targets . Fluorescent dyes with high quantum yields and extinction coefficients such as the cyanine series Cy (expending formula in Table 1) are increasingly being used to overcome these problems. The latter are superior to the classical fluorescein dyes, because they result in significantly brighter staining and are very stable to photobleaching [27,28]. These differences are reflected in our analysis, because the choice of fluorochrome significantly affected the performance of the FISH protocol, with Cy3 yielding the highest average %Eub.
Our results strongly suggest that the use of different fluorochromes will yield widely different results in the same samples, but the actual patterns obtained with a given fluorochrome within a study will probably still be robust. We also caution against the use of probes conjugated with different fluorochromes (e.g. [29,31–36]) within the same study. There seems to be a general agreement on the benefit of fluorochrome with high quantum yield, since 61% of the studies used Cy3, against 27% that used FITC, FLUOS or TRITC. However, the use of high-performance fluorochrome such as Cy3 has a tendency to increase the background signal caused by cross-reaction with other specimens, inorganic particles and detritus (T.B. and P.d.G., pers. obs.). Improving the hybridization and wash conditions and the adjustment of stringency can reduce this drawback.
Both formamide and sodium chloride (in the hybridization buffer and the wash solution, respectively), significantly influenced the performance of FISH. These two chemicals are used to adjust the stringency conditions of the hybridization and post-hybridization steps for proper annealing of the oligonucleotide probes to the target sequence . Formamide decreases the melting temperature by weakening the hydrogen bonds, thus enabling lower temperatures to be used with high stringency. The salt immobilizes hybrid molecules and is used instead of formamide in order to reduce the amount of toxic waste . Optimizing the hybridization stringency conditions is a trade-off between an optimal probe sensitivity and specificity. Increasing the formamide concentration increases the specificity of the hybridization, but further addition results in a drastic drop of bound probe and signal intensity [39,40]. In theory, the optimal formamide concentration depends mostly on the structure of the oligonucleotide sequence and the hybridization temperature, and this should not vary greatly among studies that use the same sequence and hybridization conditions. Yet, this is not the case at all. For example, many of the studies reviewed here used a hybridization temperature of 46°C, but the formamide concentration used ranged from 0 to 35%. Likewise, there was a range from 0 to 900 mM NaCl in the wash solution used in these same studies. The results presented here suggest that these methodological differences among studies result in significant differences in the detection of target cells and render the results less comparable among studies.
The other methodological factors such as the hybridization temperature, the fixative and the counting methods, did not have a significant effect on the proportion of target cells hybridized by EUB338. This lack of significant effect must be interpreted with caution, however. Studies that have examined these factors individually have reported important effects on the hybridization efficiency [15–17]. Our analysis suggests that the type of fixative, for example, may be secondary to other methodological aspects, such as formamide concentration, or alternatively, that there may be insufficient data to perform a meaningful comparison, as is the case with the counting method. The lack of a significant effect of these methodological factors in a combined analysis such as the one we carried out, should not be interpreted as a suggestion that these factors are unimportant and therefore be overlooked when developing a FISH protocol.
4.2Environmental factors that influence FISH performance
The ecosystem type was the factor that explained the largest amount of variability in %Eub. We aggregated the data into 10 Ecosystem types, which vary greatly in trophic state, from the hyper-eutrophy of an enriched culture to the oligotrophy of drinking water. It is conceivable that if researchers tended to apply a certain variation of the protocol in specific ecosystems, the strong influence of the latter may reflect differences in the protocol. An inspection of Table 1, however, reveals that there are no systematic patterns in terms of protocols used within a given system.
The influence of ecosystem type becomes obvious when the extremes are analyzed. Systems that have little resource limitation, such as laboratory cultures, biofilms and sewage sludge, tend to yield a high percentage of hybridized cells. Not surprisingly, these are the systems where the first successful applications of FISH in natural aquatic systems were performed [30,42–44]. But other aspects of the patterns in %Eub across different systems are more difficult to explain simply on the basis of productivity or nutrient enrichment. For example, the %Eub in continental shelf waters is, on average, higher than in nearshore marine waters, which are typically more productive. Freshwaters on average appear to yield somewhat lower %Eub than marine samples in general, although there are several eutrophic lakes, reservoirs and ponds included in the data set (Fig. 1A). Finally, coastal bays and marshes, as well as marine and freshwater sediments, tend to yield lower %Eub than all other systems, although the latter also tend to have high levels of metabolism . The low detection in sediments (and to a lesser extent in other samples with high particulate loads, such as some estuarine and marsh samples, Fig. 1A) could be partially attributed to the difficulty in extracting the cells and distinguishing them among the particles , and there are reports on problems associated to the detection of cells even in sediments characterized by a high bacterial activity (Fig. 2A) . However, the differences between marine and freshwater systems, for example, are more difficult to explain simply on a methodological basis. The absence of a clear relationship between the perceived productivity of a system and the FISH signal may result from the fact that a system's trophic status and the observed bacterial growth rate and biomass are often relatively uncoupled [20,25,50]. Alternatively, these differences could suggest either an almost complete lack of relationship between the level of bacterial activity and %Eub, or alternatively, that these ecosystem effects may be driven by intrinsic differences in phylogenetic composition across ecosystems .
Different molecular approaches are now converging to show that there are major differences in the phylogenetic composition of the bacterial assemblages among different aquatic systems [2,46]. In this regard, the dominant phylogenetic group was marginally not significant in explaining the variability in the %Eub cells detected with FISH. It is possible that intrinsic differences in cell activity or rRNA contents between phylogenetic groups could influence the capacity to detect these cells using FISH, and explain the strong effect of ecosystem type on %Eub. Such differences, for example, have been shown between Sphingomonas sp. and Vibrio sp., two species that are abundant in marine waters  and that belong to two different phylogenetic groups, i.e. the α- and γ-subclass of the Proteobacteria. These two species have different numbers of copies of the rRNA operon, leading to intrinsic differences in the number of potential targets . There is also emerging evidence that different phylogenetic groups may be functionally distinct [3,49] and have intrinsically different levels of single-cell activity [11,19,20,21,50], which could theoretically impact their response to FISH. For example, a strong relationship was found between the abundance of β-proteobacteria, the abundance of highly active cells, and bacterial growth rate in a temperate estuary , suggesting that this bacterial group may be intrinsically either more active or more responsive to FISH than are other estuarine populations. However, most of the evidence to date linking phylogenetic composition to cellular activity and macromolecular contents is largely circumstantial, and the importance of the phylogenetic dominance as a factor driving the variability in FISH is one aspect that has profound implications and that needs to be further addressed.
Regardless of whether there is a link between phylogenetic composition and the response of cells to FISH, it is likely that the detection sensitivity will depend in part on the physiological state of the target cells. Faster-growing or highly active cells tend to have more ribosomes, and hence bind proportionately more probe molecules, resulting in cells hybridized with a stronger fluorescent signal . This relationship has been demonstrated on bacterial strains in steady-state laboratory cultures (e.g. [8,51,52]), and is often used to explain variation in the proportion of cells detected with FISH in natural water samples [20,23,35,43,50,53–57]. However, Kerkhof and Kemp  showed that, in contrast to bacteria grown in steady state, there was no relationship between rRNA content and specific growth rate with bacteria in non-steady-state, and there are virtually no studies that have assessed the relationship between bacterial activity and FISH response in natural bacterioplankton communities. Gasol et al. , for example, showed that there was a strong correlation between the number of high-DNA cells (assumed to be more active) and the number of hybridized freshwater riverine and reservoir Bacteria, but could not find any correlation at all with the %Eub.
We attempted to assess the possible relationship between %Eub and the level of cellular activity with data extracted from six papers that reported %Eub and also bacterial growth rate determined from the incorporation of leucine or thymidine [11,19–23]. Individually, none of these data sets showed a correlation between specific rates of leucine or thymidine incorporation and %Eub. When the data sets are combined, however, there is a suggestion of a positive relationship between substrate incorporation rate and %Eub when only bacterioplankton are considered. This relationship seems much stronger for studies that have used thymidine as a tracer of bacterial production (Fig. 2B). What Fig. 2 does show, is that in assemblages characterized by high bacterial growth rates, FISH detection is also high and rapidly approaches 100%, but when community growth is slower, %Eub is not only lower on average, but also much more variable. The studies that focused on bacterioplankton ([19,21–23], Gasol pers. comm.) employed almost identical protocols, so not all the variability observed in Fig. 2 is driven by methodological differences. On the other hand, the large departure of the biofilm data  from all others may be an indication of methodological differences, or may suggest that different types of communities, i.e. bacterioplankton versus biofilm, may have intrinsically different relationships between specific bacterial activity and the capacity to detect cells using FISH.
The variability in the effectiveness of FISH for a given level of bacterial activity might be expected if we consider that the bulk of the ribosome pool is not required for protein synthesis, and that the ribosomes are not the limiting factor contributing to a low rate of bacterial growth. Cells with low activity might have rRNA at a sufficient concentration to yield a fluorescent signal detectable with FISH. For example, Fukui et al.  and Kramer and Singleton  concluded from their experiment on Desulfobacter and Vibrio species that between 20% and 30% of the maximum value of intracellular rRNA still persisted in starved cells, and Fukui et al.  showed that the limit of detection of cells (after FISH with EUB338 labeled with FITC) was reached when the rRNA content of the cells was less than 8% of the rRNA content of growing cells. Interestingly, a different pattern of physiological response to starvation has been demonstrated between the γ-proteobacteria and the Cytophaga-Flavobacterium group , perhaps with consequences on taxon-specific responses to FISH. In addition, it has been suggested that, besides the actual physiological state of bacteria, the physiological history of the cells appears to have a dramatic effect on the rate at which cells cease to be detected using rRNA-targeted probes .
The absence of a strong relationship between %Eub and cell-specific activity could be partly due to the fact that the latter may not be a good reflection of the level of activity of the cells that are actually growing and being detected by FISH. In natural bacterioplankton assemblages there is coexistence of cells in various stages of growth, dormancy and starvation, and bulk measurements of bacterial activity provide no information regarding the distribution of this activity among cells at all. It is conceivable that the relationship between %Eub and bacterial activity would be much stronger if the former were compared to the bulk activity scaled to the fraction of cells responsible for most of the community metabolism, or to the growth rate of the dominant ribotypes, rather than to the entire assemblage.
There is little question that the type of protocol used will also affect the apparent relationship between bacterial growth or metabolic rate and %Eub. For example, Alfreider et al.  report a relatively constant proportion of hybridized Bacteria in the slush layers of a mountain lake, in samples where bacterial production or other growth proxies varied by more than an order of magnitude. They hypothesize that this was partially the consequence of improved methodology, especially the superior signal strength of the fluorescent dye Cy3 they used. Another example of how methodology may influence the relationship between cell activity and %Eub is provided by Pernthaler et al. . In this study, the authors compared the performance of the common EUB338 probe conjugated with CY3 with a polynucleotide probe with multiple fluorescein or CY3 conjugates, in samples from the North Sea. The latter probes were shown to be significantly superior in terms of detection of target cells relative to EUB338. Interestingly, the polynucleotide probes appeared to be much less sensitive to changes in the level of bacterial activity (that most likely occurred during the seasonal cycle) than the conventional EUB338 probe, which seems more dependent on seasonal changes in cell activity. The conclusion here is that, the more effective the protocol is to detect target cells, the less effective the protocol may become to assess the physiological state of the cells.