Motivation: Comparisons of gene expression levels within and between species have become a central tool in the study of the genetic basis for phenotypic variation, as well as in the study of the evolution of gene regulation. DNA microarrays are a key technology that enables these studies. Currently, however, microarrays are only available for a small number of species. Thus, in order to study gene expression levels in species for which microarrays are not available, researchers face three sets of choices: (i) use a microarray designed for another species, but only compare gene expression levels within species, (ii) construct a new microarray for every species whose gene expression profiles will be compared or (iii) build a multi-species microarray with probes from each species of interest. Here, we use data collected using a multi-primate cDNA array to evaluate the reliability of each approach.
Results: We find that, for inter-species comparisons, estimates of expression differences based on multi-species microarrays are more accurate than those based on multiple species-specific arrays. We also demonstrate that within-species expression differences can be estimated using a microarray for a closely related species, without discernible loss of information.
Supplementary information: Supplementary data are available at Bioinformatics online.
Studies of gene expression levels have been revolutionized by the advent of DNA microarrays, which enable the characterization of transcription profiles of tens of thousands of genes simultaneously (Bowtell, 1999; Eisen and Brown, 1999). This technology has been applied to a wide range of problems, including transcriptional profiling of developmental processes, comparison of gene expression levels between cases and controls, and drug response studies (e.g. Scherf et al., 2000; Shapiro et al., 2004). In particular, it has recently been used to elucidate variation in expression within a species [e.g. between sexes of Drosophila melanogaster (McIntyre et al., 2006)], and the evolution of expression levels between species (e.g. Khaitovich et al., 2005; Ranz et al., 2003). In the latter applications, an important limitation is that microarrays are available only for a small number of model organisms. As a result, studies of gene expression levels in different species or strains rely on cross-species hybridization, i.e. hybridization of target RNA to gene probes designed for other, closely related species (Gilad and Borevitz, 2006).
Cross-species hybridization studies can be divided roughly into two sets. In the first, studies aim to estimate gene expression differences within species. For example, D.melanogaster microarrays were used to study gene expression profiles at different developmental stages of D.simulans (Rifkin et al., 2003). Similarly, human microarrays were used to study gene expression differences between diseased and healthy chimpanzee liver tissues (Bigger et al., 2001), and to study patterns of gene expression following SIV infection in rhesus macaques (Vahey et al., 2003). These studies implicitly assume that a microarray designed for one species can be used to interrogate differential expression in another species without substantial loss of information (Bigger et al., 2001, 2004; Diez-Tascon et al., 2005; Donaldson et al., 2005; Huff et al., 2004; Vahey et al., 2003). In other words, they assume that, although there are sequence mismatches between the target RNA and the probes on the array, the foreign RNA will hybridize sufficiently well to the microarray, such that within-species gene expression differences can be reliably identified.
To date, this assumption has not been tested explicitly, as previous studies did not have a way to compare the hybridization properties of perfect and mismatched RNA without confounding technical effects. Instead, most studies used a number of indirect approaches to assess the suitability of cross-species hybridization to detect differentially expressed genes, with somewhat conflicting results. Vahey et al. (2003), for example, reported no detectable difference in the dynamic range of absolute intensity levels between human and rhesus macaque hybridizations, when human microarrays were used for both species. Similarly, Walker et al. (2006) concluded that human sequence-based DNA arrays can be used effectively to detect differential gene expression in closely related primates, based on their observation that similar numbers of genes were detected as present or absent on Affymetrix human arrays when RNA from different species was used. In contrast, Bar-Or et al. (2006) suggested that using arrays designed based on other species may be problematic. They investigated the reliability of cross-species hybridizations by studying gene expression patterns in nematode-infected potatoes, using either potato or tomato custom microarrays. When results from the two approaches were compared, Bar-Or et al. (2006) found that estimates of differential expression using cross-species hybridization were inconsistent with results from species-specific hybridizations. They concluded that the use of cross-species hybridization ‘may bias gene expression profiling of a biological process’. A caveat, as the authors point out, is the uncertainty associated with the homology of some of the tomato/potato probes for orthologous genes. In addition, the authors note that inconsistencies across platforms might be due to ‘scanning protocols, dye labeling orientation, laboratories, etc …’, as they could not account for these technical differences between the cross-species and species-specific experiments (Bar-Or et al., 2006). In summary, while several indirect studies suggested that cross-species hybridizations can be used to reliably detect differential expression (Rifkin et al., 2003; Vahey et al., 2003; Walker et al., 2006), a more direct examination suggested that it may be problematic (Bar-Or et al., 2006). However, the latter study relied on two distinct platforms. Thus, the extent to which cross-species hybridization reliably detects gene expression differences remains to be tested explicitly, using a platform that enables one to control for technical artifacts.
The second class of cross-species hybridization studies aims at studying gene expression differences between species, with the general goal of elucidating the evolution of gene regulation (Enard et al., 2002; Gilad et al., 2006; Khaitovich et al., 2004; Nuzhdin et al., 2004; Ranz et al., 2003; Saetre et al., 2004). This aim introduces an additional challenge to the analysis, as both expression levels and sequences may differ between species. As previous work has shown, when inter-species studies are performed using a microarray designed for a single species, sequence mismatches between target RNA and the probes on the array affect the estimates of gene expression, leading to biased estimates (Gilad et al., 2006; Sartor et al., 2006).
In order to circumvent this problem, two approaches have been used. The first approach is to mask all probes with a sequence mismatch between the studied species (Khaitovich et al., 2004, 2005; Khaitovich et al., 2004; Kirst et al., 2006). However, when gene expression is compared across more than two species, a large number of probes have to be masked, resulting in exclusion of some genes and reducing the precision of measurement of gene expression for others. For example, Affymetrix human arrays measure gene expression using sets of 11 short (25 mers) probes for each gene. If expression between human, chimpanzee, and rhesus is compared using this platform (given mean divergence values of 1% between human and chimpanzee and 5% between human or chimpanzee and rhesus), an average of fewer than three probes per gene are expected to be a perfect sequence match for all three species. This small number of probes may greatly reduce the precision with which gene expression is estimated.
An alternative approach to control for the effect of sequence mismatches is to use probes derived from the various species of interest (Gilad et al., 2005). Gene-probes for individual species can be used in at least two possible study designs: (i) multi-species microarrays, in which probes from multiple species are included on each array, or (ii) species-specific microarrays, where each array includes probes from only one species. In both configurations, sequence mismatches do not affect estimates of gene expression. The use of species-specific arrays does not rely on cross-species hybridization, while the use of multi-species arrays provides an estimate of gene expression levels without the confounding effects of sequence mismatches (Gilad et al., 2005). Indeed, in a multi-species array, RNA from each species is hybridized to probes from the same species as well as to probes from other species. For example, human and chimpanzee RNA are co-hybridized to both human and chimpanzee probes. Loosely, the sequence mismatch effect is then cancelled out by averaging the intensity ratios estimated from the human and chimpanzee probes (Gilad et al., 2005).
Given that single-species and multi-species arrays both circumvent the problem of sequence mismatches, the question arises as to which platform configuration is preferable for a given number of hybridizations. Microarrays for individual species clearly present a great advantage over multi-species arrays, regardless of the type of probes used (cDNA or short oligo based), as multi-species arrays have to be designed and constructed specifically for each combination of studied species. However, in the case of species-specific microarrays, a potential problem is that expression levels are compared among microarrays constructed for the different species. The probes on each species array differ, and may therefore bias the estimates of gene expression differences.
To examine these issues in more detail, we generated expression data using a multi-species cDNA array and assessed the reliability of cross-species hybridizations when gene expression differences are estimated within and between closely related species. First, we considered the case where microarrays of one species are used to examine differential expression within a different, but closely related species. Second, we investigated the relative performance of species-specific and multi-species platforms for the study of gene expression profiles across species, for a given number of RNA samples.
2.1 Microarray data
We used two datasets. For the analysis of cross-species hybridization, we used a comparison of gene expression in a liver cell line (HEP) before and after a 48-h treatment with RNAi for the FOXO1A transcription factor. Three biological replicates were used for the FOXO1A RNAi treatment. RNA from each biological replicate was hybridized to the multi-species cDNA array along with RNA from the untreated cells as reference, in two technical replicates (Supplementary Figure 1).
For the analysis of probe effects, we used a comparison of gene expression between RNA extracted from the liver of one adult human and one adult rhesus macaque. We used a two-dye system and performed the following hybridizations: (i) self–self hybridization of a human liver RNA sample, (ii) self–self hybridization of a rhesus liver RNA sample and (iii) a co-hybridization of the human and rhesus samples. Each configuration was repeated in two technical replicates with a dye swap (Supplementary Figure 3).
2.1.1 Hybridization and data collection
Hybridizations were performed as previously described (Gilad et al., 2005), using the multi-species cDNA array (Gilad et al., 2005). This microarray contains probes for ≈1000 orthologous genes from human, chimpanzee, orangutan and rhesus. Probes for eight of those genes were independently amplified more than once using the same primers, and spotted on the array in different locations.
Hybridization intensity data were extracted using Genepix 6 (Molecular Devices, Sunnyvale, CA) with the morphological background estimation. Background correction was performed using the “normexp” procedure in limma (Smyth, 2005). After background correction, the measurements from the Cy5 channel (Rp) and the measurements from the Cy3 channel (Gp) for each probe p were converted to M and A values for normalization, where is the log ratio and is the average log expression value.
2.2 Analysis of cross-species hybridizations
Intensity-dependent print-tip lowess normalization was performed for all arrays (Yang et al., 2002). Normalized M values were analyzed using a probe-wise linear model with an effect for the RNAi treatment using the limma package (Smyth, 2005). The blocking introduced by the technical replication was handled by pooling the variance across replicates, using the method of Smyth et al. (2005). This analysis produced estimates of the log fold change (FC) in expression for each probe. Discrepancies estimated between probes from different species were calculated by subtracting the log fold difference estimated using cross-species hybridization from the log fold change estimated from the human probe, i.e. (where s = c, o, r for results from the chimpanzee, orangutan or rhesus macaque probes, respectively). Tests of differential expression were conducted using empirical Bayes-moderated t-tests, which ensure stable inference, even with small sample size (Smyth, 2004).
2.3 Analysis of probe effects
Normalization of the self–self hybridizations was performed using print-tip lowess normalization. For the co-hybridization of human and rhesus, a lowess curve was fitted separately using either the rhesus or the human probes. Assuming that the sequence mismatches affecting the human and rhesus probes are, on average, equal in magnitude but opposite in direction, the two lowess curves are expected to be symmetric about zero (Gilad et al., 2006). Therefore, we normalized the arrays by adjusting the log differential expression measures by the average of the two lowess curves. This procedure yields normalized M values. In addition, since we analyzed single-channel measurements, we performed a quantile normalization on the A values (Yang and Thorne, 2003). After background correction and normalization, the data were converted back into separate channel estimates.
We assumed that the measured intensity (I) of a gene from the target t on a probe p is proportional to the abundance of the target RNA in the sample (Ct) (Irizarry et al., 2003). Hence, if the probe has a perfect sequence match to the transcript (e.g. measuring the expression of a human gene using a human probe), then the measured intensity will be the result of the binding properties of probe p (Pp) multiplied by the transcript abundance (Irizarry et al., 2003).2005). The measured intensity on a particular probe will therefore be 2006). All the parameters are now on the log scale, and so they become additive. Here, μt refers to the transcript abundance, κtp is the sequence mismatch effect between target species t = h, r and probe species p = h, r, which will be zero if the target and probe species are the same. We assume that the sequence mismatch effects are symmetric in target and probe such that ktp = kpt. πp is the probe effect and αs is a random effect for the two highly correlated measurements of each channel taken from each spot s, which are assumed to be independent with variance . Finally, εtps is the residual error, assumed to be independent with variance σ ɛ2. The fit was performed using a REML estimation procedure implemented by the mixedModel2 function in the statmod package in R (Smyth, 2006).
To test the hypothesis that the relative probe effects are significantly different from zero, we fitted a new model to the data, in which we omit the probe effect term. In this case, we fitted both models using maximum likelihood and used a likelihood ratio in order to test whether the full model fits significantly better than the restricted model. P-values were estimated under the assumption that the likelihood ratio is distributed as χ2 on one degree of freedom. Multiple testing adjustment was performed by using a false discovery rate approach (Benjamini and Hochberg, 1995).
In order to study the different components of the probe effect, we used the data from the eight genes that were extracted and spotted more than once on the array. We modeled the extraction effect as a random effect, and the composition effects due to changes in sequence as fixed effects. The model is therefore expanded to include an extra random effect term for the extraction, βe.2000).
3.1 Intra-species comparison of gene expression using cross-species hybridization
Our first goal was to test the reliability of cross-species hybridization to assess differential expression. To do so, we used hybridization results from a multi-species cDNA microarray. This microarray contains species-specific probes for 907 orthologous genes from human, chimpanzee, orangutan and rhesus macaque (Gilad et al., 2005), corresponding to average nucleotide divergence values of 1–5% (Chen and Li, 2000; Consortium, 2005; Wall et al., 2003).
The data analyzed are from a comparison of gene expression levels in a human liver cell line (HEP) with or without a treatment with RNAi for the FOXO1A transcription factor. Specifically, three biological replicates of the RNAi treatment were co-hybridized to the multi-species cDNA array along with a reference control, in two technical replicates (Paola De-Candia, AO, and YG; manuscript in preparation). The choice of a comparison of gene expression levels with or without a transcription factor knockdown is particularly useful for our investigation, because we expect many genes to be differentially expressed in this dataset. Using this study design, we were able to compare estimates of differential expression obtained from the human probes on the array to differential expression estimates from the chimpanzee, orangutan or rhesus macaque probes that correspond to the orthologous genes (Supplementary Figure 1). Because probes from the four species are on the same array, technical variation in sample preparation (due to RNA sampling, labeling, hybridization, etc…) does not contribute to the difference between species-specific probes. Thus, by comparing the estimates of differential expression across probe sets from different species, we were able to directly compare the results obtained from the perfect match probes (i.e. the human probes) to the results from cross-species hybridizations at varying divergence levels (from approximately 1–5%).
We generated background-corrected and normalized log ratios for each probe on each array. We then fitted a probe-wise linear model to the data from all the arrays, and estimated expression differences between treatment and control from each probe (see methods). In order to assess the reliability of cross-species hybridization to detect differential expression, we first compared the magnitude (fold change) of the expression difference estimated from the human probes to the corresponding estimates from probes of the other species. To do so, we estimated for each gene , where FC is the fold change in expression between treatment and control, estimated from a human probe (h) or an orthologous probe (s = c, o or r, for chimpanzee, orangutan, or rhesus macaque probes, respectively). We then compared Δ, which can be thought of as the effect of cross-species hybridization on the fold change estimates, to standard errors of the estimated differential expression from the human probes only, which can be thought of as an estimate of technical variance. As can be seen (Fig. 1), the difference in estimates of differential expression across probes from different species is only rarely larger than the technical variance estimated from replicate experiments using perfectly matched probes. This observation suggests that the effect of cross-species hybridization on estimates of differential expression is minimal. We further explored the effect of cross-species hybridization by assessing gene expression differences between treatment and control using moderated t-statistics (Smyth, 2004), and comparing the results across probe sets for different species (Fig. 2). Given the low values of Δ estimated above, we expected that differentially expressed genes will be reliably identified using probes from all species. Indeed, we found that the correlations from all probe sets are high: r = 0.83, 0.80 and 0.80 for comparisons of t-statistics from the human probe set to that of the chimpanzee, orangutan, and rhesus, respectively. In particular, we find that 471/590 (80%) of genes that are identified as differentially expressed (at an FDR < 0.05) using the chimpanzee probes are also identified as differentially expressed using the human probes. The overlap is 79% when using the orangutan or rhesus macaque probes (Table 1), indicating that there is no increase in the false discovery rate with divergence. Conversely, of the genes identified as differentially expressed using the human probes (at an FDR < 0.05), the proportion identified using the chimpanzee, orangutan or rhesus probes is 0.80, 0.76 and 0.73, respectively (Table 1). This observation may reflect a certain loss of power to detect differentially expressed genes as divergence increases. However, importantly, the concordance between the lists of differentially expressed genes using data from probes for different species is comparable to that observed for data generated using different probes for the same species (Shi et al., 2006). In addition, many of these expression changes are small and, thus, their identification using a statistical cutoff may be highly sensitive to experimental error. To address this, we used the data from the human probes to identify 103 differentially expressed genes with a minimum absolute fold change of 1.5, as well as at an FDR <0.05. Of these 103 genes, the proportions identified using the chimpanzee, orangutan or rhesus probes (at an FDR <0.05) are 0.93, 0.93 and 0.91, respectively.
|Number of DE genes using the human probes||Number of DE genes using the chimpanzee probes||Number of DE genes using the orangutan probes||Number of DE genes using the rhesus probes|
|Overlap with results from the human probe||471||443||426|
|Number of DE genes using the human probes||Number of DE genes using the chimpanzee probes||Number of DE genes using the orangutan probes||Number of DE genes using the rhesus probes|
|Overlap with results from the human probe||471||443||426|
Finally, we find no increase in the values of Δ, or a reduction in the correlations of the t-statistics, as overall sequence divergence between species increases (Figs 1 and 2). Put together, these observations indicate that cross-species hybridization at the studied range of divergence values have a modest, if any, effect on the ability to detect gene expression differences. An analysis of results from probes with different divergence values also supports this conclusion. Specifically, when we considered the discrepancy between estimates obtained from probes for different species, we found that it did not correlate with the estimated sequence mismatches between probes (Supplementary Figure 2). Assuming sequence mismatch effects are a proxy for sequence divergence between target RNA and the probe (Gilad et al., 2005), this again suggests that estimates of expression differences are not affected by sequence divergence.
3.2 Inter-species comparison of gene expression
Our second goal was to investigate the differences between species-specific and multi-species microarrays when applied to study inter-species expression differences. While on a multi-species array, each RNA type is hybridized to probes from multiple species, when species-specific platforms are used, RNA is measured and compared using different probes. For example, gene expression differences between human and rhesus macaque are estimated using species-specific arrays by separately hybridizing human and rhesus macaque RNA to human-specific and rhesus-specific microarrays, respectively. In this case, the expression of each individual gene is measured using independent human and rhesus macaque orthologous gene probes. Thus, the binding properties of specific probes (henceforth referred to as the ‘probe effect’) may affect the measurement of gene expression and hence the estimated difference in expression between species.
The dataset that we used in order to estimate the size of the probe effects consists of the following two-dye hybridizations on the multi-species cDNA array: (i) a self–self hybridization (i.e. one sample labeled twice independently) of a human liver RNA sample, (ii) a self–self hybridization of a rhesus macaque liver RNA sample and (iii) a co-hybridization of the human and rhesus macaque samples. Each configuration was repeated in two technical replicates with a dye swap (see the Methods Section).
Again, we obtained background-corrected and normalized-intensity estimates for each sample from both human and rhesus macaque probes on the array. Given our study design, we have measurements of the human target on the human probe the human target on the rhesus macaque probe, and the rhesus macaque target RNA on both the human and rhesus macaque probes, for every gene (Supplementary Figure 3). These four measurements allow us to estimate four parameters per gene: two species-specific expression levels (from human and rhesus macaque), the sequence mismatch effect and the relative human to rhesus macaque probe effect. These effects were estimated using a gene-wise linear mixed model on the single-channel measurements (see the Methods section for more details on the modeling of these effects).
Figure 3 shows the distribution of the probe effect estimates for all genes on the array. If there were no differences in probe binding properties between the human and rhesus macaque probes, we would expect the probe effect to be zero. However, we find that a large fraction (59%) of probe effects are significantly different from zero (FDR < 0.05), and that this difference can be as large as 8-fold (Fig. 3). We note that the distribution of probe effects is symmetric around zero, indicating that, as expected, neither the human nor the rhesus macaque probes have consistently better (i.e. stronger affinity) binding properties, when sequence mismatches are taken into account.
3.3 Investigating the probe effect
A number of factors might result in binding property differences between cDNA probes, not all of which will be related to the sequence composition of the probe. In particular, it is recognized that differences in the concentration of probes might affect the measured intensity (Yang et al., 2002). If so, the probe effects shown in figure 3 would reflect different sequence composition across species (i.e. a composition effect) as well as differences in extraction and spotting of the cDNA probes (i.e. an extraction effect). To test this, we took advantage of eight genes for which probes were extracted and spotted on the array more than once for each species (see the Methods section), and for which it is therefore possible to distinguish the extraction effect from the composition effect. While the number of genes with which we could perform this analysis is small, it appears that both extraction and composition effects contribute to the overall probe effect (Fig. 4).
4.1 Cross-species hybridization
We tested whether cross-species hybridization affects the ability to measure differences in gene expression within species using hybridization results obtained from a multi-species cDNA array. Our observations suggest that cross-species hybridizations do not result in a discernable loss of information. In particular, we found that estimates of differential expression from probes of different species do not generally differ more than technical replicates of the same experiment, using perfectly matched probes. As a result, we observed a high (∼0.8) correlation of differential expression across probes of closely related species. Our correlation results are similar, regardless of the level of sequence mismatches between the probes. We also did not observe a reduction in the slope of the regression when estimates of differential expression were compared across probes with increasing levels of sequence divergence (Fig. 1). This observation suggests that the relative precision with which differences in gene expression are estimated is not affected by the level of sequence divergence (given the range of divergence examined). The observation that sequence differences between the target and probe do not markedly affect the ability to detect differential expression has also been shown using Affymetrix arrays. Cope et al. (2004) demonstrated that mismatch probes, which contain a single nucleotide mismatch in their probe sequence, can be used to detect differential expressions and perform as well as the perfect match probes for nearly all but the most weakly expressed genes.
Considered together, our results indicate that gene expression studies within species can be performed using surrogate microarray platforms for a closely related species. In particular, our observations suggest that as long as orthologous genes are known, a human microarray can be used to study within-species gene expression differences in apes and old-world monkeys, without the need to construct a specific microarray for each species of interest.
4.2 Species-specific or multi-species microarrays?
We have previously demonstrated that the use of a microarray designed for one species to study inter-species gene expression differences is problematic, because of the effect of sequence mismatches on hybridization intensity (Gilad et al., 2006). Either multi-species or species-specific microarrays can be used to circumvent this problem. However, gene expression estimates from species-specific arrays may be confounded by probe effects, which may lead to differences in hybridization intensities even when the gene is not differentially expressed. The effect of probe-binding properties on hybridization intensity has been observed previously. On Affymetrix platforms, for example, a large variation in hybridization intensity was observed across different probes for the same gene (Draghici et al., 2006). In typical microarray experiments, probe effects can be generally ignored (when multiplicative error is assumed) because RNA samples are compared using the same probe or set of probes across arrays, and the probe effects cancel out (Draghici et al., 2006). However, this is not the case when gene expression across species is compared using species-specific microarrays, since the expression of different orthologous genes is measured by species-specific probes. Unlike different Affymetrix probes for the same gene, orthologous probes are highly similar. Hence, it might be expected that the probe effect from species-specific arrays would be minimal. For example, on the multi-species array that we used, orthologous probes from all species were amplified using the exact same PCR primers (Gilad et al., 2005). Nevertheless, we found that 59% of probe effects are significantly different from zero. Thus, when we used the multi-species cDNA array, and mimicked a species-specific array analysis by considering human and rhesus probes separately, probe effects confounded the estimates of gene expression, leading to spurious detection of differentially expressed genes (Supplementary Figure 4). As an illustration, only 21 genes were shared among the top 100 genes (ranked by P-value) that were identified as differentially expressed using either the multi-species array or the species-specific data.
For a small number of genes, we were able to deconstruct the probe effect to its extraction and composition components. While the extraction component may be mainly relevant to spotted microarray platforms, the composition effect is expected to be even more severe in microarray platforms that make use of much shorter probes and which are therefore more sensitive to sequence differences. As discussed above, inter-species expression analysis using multi-species arrays side-steps the problem of probe effects since expression estimates are measured using the same probes across arrays.
One study design that would allow the use of single-species arrays while overcoming the problem of probe effects is to hybridize RNA from all the studied species to all the species-specific array types. While inefficient with respect to the use of biological material and microarrays, this study design would allow separate estimates of differential gene expression, sequence mismatch effects and probe effects, and thus provide unbiased estimates of gene expression profiles.
We have shown that the ability to detect differentially expressed genes within a species is not reduced by the use of an array designed for closely related species. This observation suggests that it should be sufficient to construct single-species microarrays for only a few model organisms, assuming that other organisms of interest are not too highly diverged. In addition, we have demonstrated that, given a limited number of RNA samples (or arrays), multi-species rather than species-specific microarrays should be used to reliably estimate inter-species gene expression differences.
We thank J. Borevitz, M. Przeworski and T. Speed for helpful discussions and comments on the manuscript. This research was supported by grant GM077959–01 to Y.G. and by NHMRC grant 257529 to G.K.S. and A.O.
Conflict of Interest: none declared.