|Home | About | Journals | Submit | Contact Us | Français|
The last two years saw the publication of three genome-wide gene expression studies of the fission yeast cell cycle. While these microarray papers largely agree on the main patterns of cell cycle-regulated transcription and its control, there are discrepancies with regard to the identity and numbers of periodically expressed genes. We present benchmark and reproducibility analyses showing that the main discrepancies do not reflect differences in the data themselves, microarray or synchronization methods seem to lead only to minor biases, but rather in the interpretation of the data. Our reanalysis of the three data sets reveals that combining all independent information leads to an improved identification of periodically expressed genes. These evaluations suggest that the available microarray data do not allow reliable identification of more than about 500 cell cycle-regulated genes. The temporal expression pattern of the top-500 periodically expressed genes is generally consistent across experiments, and the three studies together with our integrated analysis provide a coherent and rich source of information on cell cycle-regulated gene expression in S. pombe. The reanalyzed data sets and other supplementary information are available from an accompanying website: http://www.cbs.dtu.dk/cellcycle/. We hope that this paper will resolve the apparent discrepancies between the previous studies and be useful both for wet-lab biologists and for theoretical scientists who wish to take advantage of the data for follow-up work.
The terms ‘cell cycle-regulated’ and ‘periodically expressed’ are used interchangeably in the literature to describe genes that are expressed in a specific stage during the cell cycle. Since the pioneering work in budding yeast (Cho et al., 1998; Spellman et al., 1998), cell cycle-regulated gene expression has been studied at a genome-wide level in bacteria, plants, and mammals (Laub et al., 2000; Ishida et al., 2001; Menges et al., 2002; Whitfield et al., 2002). Recently, three independent groups have used DNA microarrays to identify fission yeast genes that are periodically expressed as a function of the cell cycle (Rustici et al., 2004; Peng et al., 2005; Oliva et al., 2005). For Schizosaccharomyces pombe there are thus now more data available on cell cycle-regulated gene expression than for any other organism. This provides valuable biological information and a rich source for theoretical studies (Tyers, 2004; Bähler, 2005a; Gilks et al., 2005; Wittenberg and Reed, 2005). As for other large-scale data sets (e.g., Cho et al., 1998; Spellman et al., 1998), there is only partial agreement between the three studies with regard to the number and identity of periodically expressed genes; together, the S. pombe studies proposed more than 1300 genes in total to be periodically expressed, but only 360 genes were reported in at least two of the three studies (Oliva et al., 2005). Although such differences probably do not come as a surprise for experts of genomic approaches, they can be disconcerting for biologists who may be confused and lose trust in this type of data. These discrepancies, however, can be explained, and the data are quite consistent with each other when looking beyond a superficial comparison as discussed below. We provide an overview of the data on periodic genes in fission yeast and focus on reconciling these data, and reporting follow-up analyses that compare and integrate all three data sets. We identify the following main reasons for the discrepancies in the reported cell cycle-regulated genes: differences in analysis methods, choices of significance cutoffs, and random experimental noise. Despite their differences, the three data sets are coherent and of comparable quality and, when combined, provide improved detection of periodically expressed genes.
The normalized expression data from the three cell-cycle microarray studies (Rustici et al., 2004; Peng et al., 2005; Oliva et al., 2005) were downloaded from the authors’ web pages (Table 1). All values were converted to log-ratios and technical replicates (if present) were averaged. The expression profiles for each gene in each of the ten experiments were normalized to a mean log-ratio of zero.
To rank genes, we used a scoring scheme that has been shown to be one of the best for finding cell cycle-regulated genes based on microarray data (de Lichtenberg et al., 2005). Briefly, this scheme is based on two p-values that measure the significance of regulation and of periodicity. The p-value of regulation for a given expression profile was calculated as the fraction of 106 random profiles with a standard deviation above that of the observed profile. To evaluate the periodicity, the Fourier score was calculated for a given expression profile: , where ω=2π/T, with T being the interdivision time. The optimal interdivision time for each experiment was estimated based on a reference set of 35 genes shown to be periodically expressed in small-scale experiments (Rustici et al., 2004). The p-value of periodicity was calculated for each gene by comparing its Fourier score to the Fourier scores of 106 random profiles constructed by shuffling the timepoints of the corresponding expression profile. To compensate for interdependencies among timepoints, all p-values were normalized to a median of 1. A combined score was calculated by multiplying the two p-values for a given gene and applying penalty terms to ensure that a low score is only obtained if a gene is both significantly regulated and significantly periodic: . To combine evidence from multiple experiments, the p-values were multiplied to yield a total p-value of regulation and a total p-value of periodicity from which the combined score was calculated.
Within a single experiment, the time of peak expression for a gene is determined by fitting its expression profile with a sine wave. We report this peak time in percent of the cell cycle to compensate for the difference in interdivision time between the experiments. Because different synchronisation methods release cells from different points in the cell cycle, the timescales need to be aligned before peak times can be compared between experiments. To find the optimal alignment, we used a simulated annealing heuristic to minimise the total peak time difference between experiments for the top-500 genes. We arbitrarily defined the zero timepoint as the median peak time of the genes in Cluster 2 (M/G1 phase) of Rustici et al. (2004). For each gene, a combined peak time was calculated as a weighted average (on a circle) of the peak time obtained in each of the ten experiments (see de Lichtenberg et al., 2005) and http://www.cbs.dtu.dk/cellcycle/ for details).
To evaluate the quality of any list of periodically expressed genes proposed based on microarray time series, we constructed three independent benchmark sets, each consisting of genes for which there is independent experimental evidence for cell cycle-regulated expression.
The first set (B1) consists of 40 genes, for which periodicity has been demonstrated in small-scale experiments; slight variations of this list have been used by all three groups to verify their data analyses. From the list of 35 genes used by Rustici et al. (2004), we excluded the gene suc22 as this produces two transcripts of which only one is periodic. We then added five genes that have recently been reported to be cell cycle-regulated (Alonso-Nunez et al., 2005) and the gene uvi31 (Kim et al., 1997).
The second set (B2) consists of genes whose promoters are bound by at least one of the known cell-cycle transcription factors Cdc10p, Res1p, Res2p or Fkh2p based on ChIP-chip experiments in unsynchronized cells (B.T.W., unpublished data). In case of divergently transcribed genes, where binding is observed between the genes, both flanking genes are included in the set. Although false positives will be detected in these experiments, the set should be rich in genes that are truly regulated during the cell cycle. Genes also present in set B1 were excluded to ensure independence between the benchmark sets, leaving 188 genes in set B2.
The third set (B3) consists of genes that are differentially expressed in microarray experiments using unsynchronized strains with genetic perturbations of the genes ace2, sep1, or cdc10 encoding transcription factors as well as S-phase arrested cells (Table 1; Rustici et al., 2004). All genes present in sets B1 and B2 were removed to ensure independence of the benchmark sets, leaving 321 genes in set B3.
Table 1 provides a comparison of experimental platforms and designs of the microarray studies addressing cell cycle-regulated gene expression in fission yeast. All three studies used cells synchronized by centrifugal elutriation (selective synchronization) as well as cells synchronized using the temperature-sensitive cell-cycle mutant cdc25-22 (whole-culture synchronization), with different array platforms and differing numbers of timepoints and biological repeats. The papers also include additional experiments to address the regulation of periodic transcription and/or to analyze specific cell-cycle phases in more detail (Table 1). The three studies propose different numbers of periodically expressed genes: Rustici et al. (2004) suggested 407 genes based on five experiments, whereas Peng et al. (2005) and Oliva et al. (2005) proposed 747 and 750 genes based on two and three experiments, respectively (Table 1 and Figure 1A). When comparing the three proposed sets of genes, a striking and somewhat discouraging conclusion is the poor overlap between the genes reported as periodically expressed in the three studies (Figure 1A; Oliva et al., 2005). For the two papers that reported around 750 periodic genes, the overlap with the other gene lists is especially poor (Table 1; Figure 1A). When redoing this comparison, we noticed that some of the discrepancies arise as a consequence of using different (non-systematic) names for the same genes. Correcting for these gene mapping problems improves the overlap between the studies (Figure 1B). As shown below, however, the main reasons for the poor overlap are due to differences in data interpretation, while the data per se show quite good agreement with each other. To assess these issues, we first evaluate the quality of the published data sets and analyses and then go back to discuss what the different experimental data show when analyzed with the same computational method.
Genes that are periodically expressed as a function of the cell cycle are defined as those that change in expression levels with a period equal to the interdivision time. Various algorithms have been developed for identifying periodically expressed genes, and the choice of method can have a profound impact on the interpretation of cell-cycle microarray data. In budding yeast, for example, widely different sets of genes have been proposed based on analyzing the same microarray data with different computational methods (Zhao et al., 2001; de Lichtenberg et al., 2003; Johansson et al., 2003; Luan and Li, 2004; Ahdesmäki et al., 2005; de Lichtenberg et al., 2005; Willbrand et al., 2005). While single studies identified between 150 and 1000 periodically expressed genes, in total over 1800 different genes have been proposed to be periodic. A recent comparison of the available computational methods showed that some methods simply work better than others in identifying truly cell-cycle-regulated genes and that the better methods yield more reproducible results when applied to different microarray data sets (de Lichtenberg et al., 2005). Thus, a large part of the differences between the lists of periodic genes in the S. pombe microarray studies could be due to differences in how the data were analyzed.
In all three S. pombe studies, the identification of periodic genes was based, in part, on Fourier analysis. Rustici et al. (2004) and Oliva et al. (2005) then calculated probabilities for the oscillations to arise from random fluctuations by shuffling the data for each gene within each experiment, identifying more than a thousand genes each with apparently significant periodicity. Oliva et al. (2005) ranked the genes by their p-values and proposed a list of 750 periodically expressed genes, whereas Rustici et al. (2004) filtered out genes with only subtle changes in expression levels and then visually inspected the remaining profiles to arrive at a smaller, more conservative list of 407 genes. Peng et al. (2005) instead ranked the genes by a CDC score, which combines Fourier analysis with additional terms; their threshold (747 genes) and false-discovery estimates were based on randomly shuffling the data.
To evaluate the different proposed lists of periodically expressed genes, we compared them with independent experimental evidence for cell-cycle regulation using the three benchmark sets described in Materials and Methods. In Figure 2 and Supplemental Figure S1, the number of genes retrieved from a given benchmark set is shown as a function of the number of genes included from each ranked list, whereas each non-ranked list is shown as a single point. Reassuringly, all proposed gene lists show much better than random overlap with the genes from all three benchmark sets. The enrichment over randomness (the slope of the curves) is also strongest for the highest ranked genes that scored best in the original analyses. As one goes down the ranked lists, however, the slopes of the curves eventually become comparable to that of the line representing random expectation. After the first 500 genes or so, there is no further enrichment of genes from the benchmark sets, and selecting more genes from the ranked lists is therefore no better than picking additional genes at random from the genome. Figure 2 can also be used to compare the performance of the three analyses relative to each other: the Rustici et al. list of 407 genes shows a better overlap with the benchmark sets than the highest scoring 407 genes from the lists of Oliva et al. and Peng et al. The benchmark set B3 might slightly favour the Rustici et al. list as it is based on data from the same array platform. At this point, however, it is not clear to what degree these results are influenced by the number of experiments made by each group or by the methods used to measure periodicity in expression.
To better compare the different data sets, we reanalyzed the data from all three groups using the method described by de Lichtenberg et al. (2005). In all cases, our reanalysis performs at least as good as the original analyses published (Figure 2). In brief, our analysis method combines a p-value for regulation with a p-value for periodicity, to ensure that top-ranking genes exhibit both a significant regulation and a periodic pattern of expression. On S. cerevisiae data, this approach has been shown to perform better than other methods for the identification of periodic genes, especially compared to those modelling only the shape of the expression profile without taking into account the magnitude of regulation (de Lichtenberg et al., 2005). The latter could explain the slightly poorer performance of the analysis by Oliva et al. (2005) who ranked the genes based on a score that is independent of the magnitude of regulation. Less or no improvements in performance are observed when reanalyzing the data by Rustici et al. (2004) and Peng et al. (2005) who both used methods that take into account the magnitude of regulation. In accordance with the improved performance on the benchmark sets (Figure 2), our reanalysis also improves the agreement among the three data sets (Figure 1B and 1C). The apparent discrepancies between the data sets are thus in part explained by the use of different and less accurate analysis methods.
The relative performance of the reanalysis of data from the three groups (Figure 2) also shows that the best lists are derived from data sets that include more timecourse experiments (Table 1). This finding is confirmed when applying the de Lichtenberg et al. (2005) analysis method either to all ten experiments individually or to all ten experiments in combination. Reanalysing each of the individual experiments (Supplemental Figure S1) demonstrates only minor differences in performance, which suggests that all timecourse data are of comparable quality. It is therefore not surprising that the best results were obtained when applying our analysis method to all ten experiments in combination (black curves in Figures Figures22 and S1). This is even better than taking the 176 genes included in all three published lists or the 419 genes included in at least two of the original lists (Figures (Figures1B1B and and2).2). This shows that our integrated analysis of all data is superior to simple voting schemes at combining the signals from the ten experiments, which, although being of comparable overall quality, each make independent and complementary contributions and together improve the identification of cell cycle-regulated genes.
Peng et al. (2005) and Oliva et al. (2005) suggested almost twice as many periodically expressed genes as Rustici et al. (2004) (Table 1; Figure 1A). As pointed out before, the microarray expression data reveal no natural, distinct threshold between periodically expressed genes and genes expressed at constant levels throughout the cell cycle (de Lichtenberg et al., 2005; Oliva et al., 2005). Instead, there is a continuum from clearly periodic genes to genes that do not seem to fluctuate as a function of the cell cycle, with a large grey zone in between. This could reflect that many genes are only weakly cell cycle-regulated (<1.5-fold change in expression levels) as well as noise in the microarray data. The transition can be seen in the benchmark analyses as a gradual decrease in the slope of the curves as more genes are included (Figures (Figures22 and S1). The decision on the number of genes that are deemed periodic is thus ultimately based on a somewhat arbitrary threshold. However, the slope of every curve eventually becomes comparable to that of random expectation, from which point on the available benchmark sets cannot justify the inclusion of more genes, and the threshold should therefore be set before this point. Not surprisingly, gene lists based on smaller numbers of experiments reach this limit earlier. In the best-case scenario, where all ten timecourse experiments are combined, the enrichment over random is strong for the first 300 genes, then gradually decreases and is essentially lost altogether beyond the first 500 genes (Figure 2). These analyses thus lend little support to the proposition of ~750 cell cycle-regulated genes, particularly not when based on only two or three experiments. Indeed, both the original lists and the reanalyses of the data sets by Peng et al. (2005) and Oliva et al. (2005) display hardly any enrichment beyond the first 400 genes.
To test if this lack of enrichment is due to limitations of the benchmark sets, we determined reproducibility by comparing the ranked lists obtained from our re-analysis of any two of the ten individual experiments (Figure 3). When selecting the top-300 genes from each list, the average overlap is 121 genes. However, when comparing the next 300 genes (rank 301 to 600), the reproducibility drops dramatically to only 31 genes on average. In comparison, the expected overlap between two randomly selected lists of the same size is 19 genes. From rank 601 to 900, there is essentially no enrichment over random expectation. This demonstrates that only the first about 300 genes are reasonably reproducible between any two of the ten experiments, consistent with the observations made from Figures Figures22 and S1. This drop in reliability for lower ranked genes is also confirmed by visually inspecting Figure 4, which shows the expression profiles of the same three sets of genes used in Figure 3. The top-300 genes show clear periodicity and large amplitudes, whereas these properties are less apparent to the eye in the other two groups. Similar conclusions are reached when comparing the set of 176 genes proposed in all three original studies to those included in at least two studies (243 genes) or those only proposed by one study (863 genes). Only the genes proposed by all three groups show a clear periodic pattern of expression (Supplemental Figure S2).
The gene sets visualized in Figure 4 are sorted by their peak time, whereby the pattern of periodicity stands out very clearly across a group of genes. Although a periodic pattern is seen even for the two bottom panels in Figure 4, this periodicity is not reproducible at the single-gene level when comparing individual experiments (Figure 3). The patterns of periodicity among the lower ranked genes indicate that there are truly periodically expressed genes beyond the highest ranking 300-400 genes, but identification of these requires many independent data sets and even then comes at the price of including an increasing number of false positives as one goes down the ranks.
Together, the analyses shown in Figures Figures22--4,4, S1 and S2 demonstrate that only for the most significant 300-400 genes is the signal strong enough to deem periodic based on a single timecourse experiment; by combining ten timecourses, some 500 periodically expressed genes can be identified with reasonable confidence. Beyond that, regulation becomes weaker, noisier, and/or less reproducible between experiments and therefore more questionable. Notably, many of the profiles of lower ranking genes look, at best, marginally periodic to the eye and would probably not be judged as cell cycle-regulated based on traditional methods (e.g. Figure 8D). A major reason for the poor overlap between the originally reported gene lists is thus that the studies by Peng et al. (2005) and Oliva et al. (2005) suggest many more cell cycle-regulated genes than can reliably be detected from their data. As shown in Figure 1D, the relative agreement between the three studies can be further improved if smaller, more conservative lists of periodic genes are compared. As will be shown below, most of the remaining discrepancies are explained by the general noise level in the microarray data, which leads to different genes that make it into the different top-400 lists, together with the fact that there is a continuum between cell cycle-regulated and non-regulated genes. In fact, differences in the array platforms and experimental protocols only account for a minor part of the apparent discrepancies in Figure 1D (see below). These discrepancies are expected when comparing conclusions from noisy data sets that are each based on only few replicates and should not be interpreted as a lack of congruence between the data from different groups.
Since only a small fraction of the cell cycle-regulated genes have been identified through small-scale studies, it is difficult to assess the number of false positives in a proposed list of genes. In contrast, it is easy to count how many of the known periodic genes are confirmed by microarray analysis. This has lead researchers analysing cell-cycle microarray expression data in different organisms to propose quite inclusive gene lists that have good sensitivity (including most of the known genes) but unknown false positive rate. Peng et al. (2005) and Oliva et al. (2005) employed permutation-based statistical tests and estimated their false discovery rates to be 1.1% and 0.022%, respectively. These exceptionally low error rates are difficult to reconcile with an overlap of only 293 genes between lists of ~750 genes each (Figure 1B).
Peng et al. (2005) and Oliva et al. (2005) suggested higher sensitivity or better cell-cycle synchrony as reasons why they identified more periodic genes than did Rustici et al. (2004), although this is not supported by our reanalyses described above. In fact, when using an automated method, Rustici et al. (2004) identified >1000 ‘significant’ periodic genes with p-values <0.01 in their data but decided to propose a smaller, more conservative list of cell cycle-regulated genes. It is important to realise that random permutation of timecourse data may overestimate the statistical significance of periodicity, and hence lead to an overly optimistic false discovery rate. This is because successive timepoints are not guaranteed to be independent of each other, thereby violating the underlying assumption of the statistical tests (Kruglyak and Tang, 2001). This problem is increased if samples are collected at higher frequency and is particularly true for the data by Peng et al. (2005), who applied Gaussian smoothing to their expression profiles, thus artificially enhancing dependency between neighbouring timepoints. While p-values are useful for judging the relative periodicity of a set of genes (ranking), it is problematic to rely on their absolute values. When reanalyzing the data, we have found that the raw p-values calculated based on random permutations are overestimated by about an order of magnitude, meaning that the false positive rates reported in the three original studies are probably underestimated accordingly. Using statistics alone to set the threshold, two of the groups suggested roughly twice as many genes as their data can support, as judged from the reproducibility between replicate experiments (Figure 3) and consistency with independent sources of evidence for cell-cycle regulation (Figures (Figures22 and S1). The only alternative explanation is that well over a thousand genes are periodically expressed and that each study simply detects a different subset of these, although this would contradict the claim of less than 20% false negatives by Peng et al. (2005). In any case, even if there were many more periodically expressed genes, our analyses show that their profiles are not reproducible between experiments (Figure 3).
In Figure 3, we have subdivided the pairwise comparisons of gene lists from different experiments into four classes, based on whether the two experiments were performed by the same group and based on the same synchronization method. This subdivision demonstrates that experiments performed by the same group tend to be more similar, as do experiments using the same synchronization method. For instance, experiments performed by the same group and using the same synchronization technique on average have 148 genes in common among the top scoring 300 genes, compared to 110 genes among experiments performed by different groups with different synchronization techniques. We speculate that the lab bias is largely due to differences in probe and chip design that may cause some genes to be detected less well on some arrays. One should note, however, that these biases are small and rather insignificant in comparison to the general level of reproducibility of only around 50% between the top-300 genes from any two experiments. Figure 3 thus contradicts the proposition that biases from different synchronization methods give rise to widely different, and spurious, results (Cooper and Shedden, 2003). Instead, the primary source of variation seems to be random, experimental noise rather than systematic experimental biases. Minor variations in the data leading to different genes that make it into the different top-300 lists are the main reason for the small overlap between any two experiments, as the ranking in any single experiment is influenced by subtle differences in periodicity and regulation. We can therefore conclude that the data from the three groups are of similar overall quality (Figures (Figures22 and S1), and they are congruent (Figure 3). These findings also show that the poor overlap observed in Figure 1D is simply a consequence of comparing three lists, which have each been derived from too few experiments to eliminate random, experimental noise. Since many independent experiments are needed to extract the underlying signal from noisy data, it is no surprise that our combined analysis of all ten experiments yields the best results. The differences in synchronization techniques, microarray design and laboratory protocols among the ten experiments therefore make the entire data set more information rich than would have been the case had all the experiments been performed in the same laboratory with the same method.
Agreeing on the cell cycle-regulated genes is one part of the problem; in principle, the time of expression of a gene could still vary between experiments. To examine this in more detail, we assigned a time of peak expression for each periodic gene in a given experiment by fitting its expression profile with a sine wave. These peak times were made comparable across experiments by converting the time scales from minutes to percent of the cell cycle and subsequently aligning the scales with each other (for details, see de Lichtenberg et al., 2005). For the four phase-specific gene clusters defined by Rustici et al. (2004), we calculated the smoothed distribution of peak times for each of the ten individual timecourse experiments (Figure 5). Reassuringly, we find that each gene cluster peaks at roughly the same time and occupies a similar fraction of the cell cycle in all experiments. As expected, the G2 phase constitutes about 60-70% of the cell cycle of fission yeast, in contrast to budding yeast where the four cell-cycle phases are of similar length. Importantly, the different synchronisation techniques lead to similar results, although the distribution of peak times for the S phase genes is slightly delayed for cdc25 block-release experiments compared to the elutriation experiments, indicating that the relative lengths of cell-cycle phases differ somewhat between these types of experiments.
Given the reproducibility of peak times between the different experiments (Figure 5), a single gene-specific peak time can be calculated that summarizes the expression across all ten experiments by weighing the individual peak times relative to each other based on the periodicity of the gene in each given experiment (de Lichtenberg et al., 2005). A nice feature of this scheme is that the average peak time is associated with a standard deviation that quantifies the consistency (or spread) in the temporal expression for each gene. We can thus show that the great majority of the top-500 periodic genes exhibit highly consistent peak times across all experiments (Supplemental Figure S3).
A simple way to globally view the temporal behaviour of gene expression during the cell cycle is to plot the distribution of peak times (Figure 6). This reveals two major waves where gene expression peaks are concentrated, one in M phase and one in early G2 phase, as also observed by Oliva et al. (2005). Although there are genes peaking in expression at all stages of the cell cycle, there is a clear drop in the later half of G2 phase before the largest wave is initiated at the G2/M transition. The numerous genes peaking in early G2 phase are generally much weaker regulated than those peaking during M to S phases (Rustici et al., 2004; Figures Figures44 and and8)8) and show poor reproducibility between experiments (see below); their enrichment in functions such as ribosome biogenesis (Oliva et al., 2005) suggests that this surge in cell cycle-regulated gene expression may prepare the cell for the increased growth during G2 phase (Mitchison and Nurse, 1985). Despite the two stages with enriched periodically expressed genes, the overall timing of peaks is quite continuous across the cell cycle rather than in discrete steps (Figure 4), probably reflecting regulatory fine-tuning and/or differences in mRNA stability.
Based on their estimated p-values, Oliva et al. (2005) proposed that as many as 2000 genes are weakly but significantly periodic. They supported this by showing that when analyzing the 4000 lowest ranked genes in their study, the same two major waves of transcription were observed as for their 750 most regulated genes. When plotting the distribution of peak times for the 2000 least periodic genes according to our combined analysis of all ten timecourses (Supplemental Figure S4), we generally cannot reproduce the distribution seen for the highest-scoring 500 genes (Figure 6). We too observe a tendency for more genes to be assigned to early G2 phase, but late G2 is also rich in expression peaks, which is the opposite of what is observed for the highly scoring genes. Furthermore, we see no sign of a second wave in M phase among the 2000 lowest scoring genes (Supplemental Figure S4). This analysis does therefore not support the periodicity of genes far down the list, but reflects that if one fits sine curves to the profiles regardless of how random they look, the overall pattern shows a tendency for clustering in G2 phase. Although there may be subtle fluctuations among the low ranking genes, the data presented here (Figures (Figures22--4,4, S1 and S2) indicate that these fluctuations do not arise from active regulation of these genes during the cell cycle. The fluctuations are not reproducible at the level of single genes, and the genes that are fluctuating show no significant overlap with any of the benchmark gene sets for which cell-cycle regulation is supported by other sources. Although the phenomenon as such might be interesting, more work would be required to clarify the biological relevance of these subtle oscillations. At this point, it is not even clear whether they should be viewed as a real biological phenomenon or as a bias introduced by the treatment of the microarray data (e.g. normalization).
Despite the poor overlap between the proposed periodically expressed genes, the three cell-cycle studies report a coherent picture of gene expression regulons. All three papers defined groups of genes that behave in a similar way across experimental conditions using different clustering algorithms (Table 1). Whereas the peak times define the timing of expression for each gene (Figure 6), the clustering analyses also take into account the shape of the expression profiles and incorporate additional experiments (e.g. transcription factor mutants). Rustici et al. (2004) describe four large clusters, which together contain almost all periodic genes, while Peng et al. (2005) and Oliva et al. (2005) examined eight smaller clusters each, which together cover only a fraction of the periodic genes. The genes within each cluster peak at a similar time during the cell cycle, reflecting the intuitive notion that peak time of expression is a critical feature of periodic transcription. The different clusters can be divided in three main groups: M/G1-phase, S-phase, or G2-phase. Reassuringly, different clusters within the same group share many genes, while clusters from different temporal groups show little overlap (Figure 7).
The M/G1-phase includes the highest numbers of clusters: Cluster 1 and Cluster 2 (Figure 8A; Rustici et al., 2004), SFF(1), SFF(2), Ace2 and MCB (Peng et al., 2005), and Cdc15, Cdc18 and Eng1 (Oliva et al., 2005). There is good conguence between related clusters (Figure 7). Enrichment of regulatory motifs and genetic experiments agree that the M/G1 clusters contain targets of Forkhead, Ace2p, and MBF transcription factors, which regulate genes for mitosis, cell division, and DNA replication. The data also support a model where a wave of transcription regulated by the Forkhead transcription factor Sep1p precedes and induces an Ace2p-dependent transcriptional wave, as is also emerging from other papers (Martín-Cuadrado et al., 2003; Dekker et al., 2004; Alonso-Nunez et al., 2005; Lee et al., 2005; Petit et al., 2005). Together, these findings define a transcriptional cascade for cell separation in fission yeast (Bähler, 2005b). Besides Sep1p and Ace2p, other regulators such as the Fkh2p forkhead transcription factor may be involved in this pathway (Buck et al., 2004; Bulmer et al., 2004; Rustici et al., 2004; Szilagyi et al., 2005). More work is required to understand how these regulators work together to control periodic transcription during mitosis. Detailed reviews and comparisons with the corresponding regulatory pathways in S. cerevisiae are available (Bähler, 2005a; Wittenberg and Reed, 2005).
The S-phase is characterized by the strongly regulated and tightly co-expressed histone genes (Figure 7; Figure 8B), the regulation of which is not understood. In addition, Rustici et al. (2004) reported a group of genes with lower amplitudes peaking during S-phase, but these were not enriched for any functional category. Oliva et al. (2005) described a small cluster of genes close to telomeres, although most of these are almost identical in sequence, making it difficult to know whether all or just one of them is periodically expressed.
Genes peaking during G2-phase are somewhat different as they show less reproducible and generally much weaker regulation. Accordingly, the overlap between the different G2 clusters is markedly lower than for the M/G1- and S-phase clusters; the only significant overlap is between Cluster 4 from Rustici et al. (2004) and the Ribosome cluster (Rib) from Oliva et al (2005), which is enriched for genes functioning in ribosome biogenesis (Figure 7; Figure 8C). Peng et al. (2005) reported two small clusters containing ribosomal proteins. No promoter motifs were enriched in the Ribosome cluster, and Oliva et al (2005) proposed that global transcriptional repression during mitosis could account for the weak oscillation of these genes. This idea is supported by the observation that this cluster was repressed in nuc2 mutants with condensed mitotic chromosomes (Oliva et al., 2005), although the chromosome compaction in these mutants is stronger than during normal mitosis. Further experiments will be required to substantiate this interesting hypothesis.
Besides genes involved in cell growth, a number of stress genes peak during G2-phase (genes in Cluster 4, the ATF cluster, and the Wos2 and Cdc2 clusters), which are induced in a range of environmental stresses (Chen et al., 2003). Several of these genes seem at best marginally regulated as a function of the cell cycle (e.g. Figure 8D), but more than half of them are present in our top-500 list of periodic genes. Regulation of these genes could be caused by the synchronisation methods, because they showed lower reproducibility across experiments, and some of them were mostly regulated in the cdc25 experiment, which requires a temperature shift. The periodicity of these genes suggests that the cell cycle and environmental stress response are linked, and two recent studies have started to shed light on how these processes are coordinated (Lopez-Aviles et al., 2005; Petersen and Hagan, 2005).
The periodically expressed genes identified in fission yeast have been compared to those reported in budding yeast (Cho et al., 1998; Spellman et al., 1998). All three S. pombe cell-cycle studies agree that although there is a significant overlap in regulated genes, less than 50% of the orthologous gene pairs are periodic with high amplitude in both yeasts. Of our top-500 periodic genes identified by reanalysing all ten experiments, 353 have an ortholog in budding yeast. 102 of these of are also among the top-500 periodically expressed genes in budding yeast microarray studies when applying the same computational method (de Lichtenberg et al., 2005). Distinct regulatory patterns of cell-cycle genes between S. cerevisiae, C. albicans and S. pombe have recently also been reported by Ihmels et al. (2005). Thus, cell-cycle regulation of gene expression is only partially conserved during evolution, although it does show a substantially higher conservation than the regulation of other processes such as meiotic differentiation (Mata et al., 2002).
The three microarray expression studies of the fission yeast cell cycle together provide a wealth of data including ten time series experiments, which are of comparable quality according to our benchmark analyses. Yet rather poor agreement was observed when comparing the three published lists of periodically expressed genes (Oliva et al., 2005). We have revealed four primary causes for discrepancies between the proposed lists: 1) inconsistencies in gene naming; 2) use of different analysis methods for identifying periodic genes; 3) each individual experiment is subject to random noise; and, perhaps most importantly, 4) two of the three studies proposed more periodic genes than can reliably be detected from their data. We could detect only minor systematic differences between data sets produced by different laboratories or using different synchronization techniques. The data themselves are thus congruent, but subject to random experimental noise, which explains the remaining lack of overlap (Figure 1D). As demonstrated by our meta-analysis, the best results are obtained when using a powerful computational method to integrate all available data. The combination of all data from the three independent studies provides an information-rich data set that is superior to the data from any single experiment or laboratory (hence ‘the more the merrier’ in the title). Based on benchmark and reproducibility analyses, we conclude that even in this best situation no more than about 500 periodically expressed genes can be reliably identified based on the available data. Although there may be more genes that are marginally cell cycle-regulated, increasing the list beyond the highest scoring 500 periodically expressed genes will come at a considerable cost of false positives. The temporal expression pattern of the top-500 genes is highly consistent across all ten experiments, which shows that the three studies provide a coherent description of cell-cycle regulated gene expression in S. pombe. Accordingly, there has been good agreement between the three studies with regard to various gene expression modules and their regulation. We hope that our integrated analyses and data sets clarify the reasons for discrepancies between the original studies and that they will be useful for follow-up studies, both experimental and theoretical.
We thank Alvis Brazma and Juan Mata for comments on the manuscript. Research in the Bähler laboratory is funded by Cancer Research UK [CUK], Grant No. C9546/A5262 and by DIAMONDS, an EC FP6 Lifescihealth STREP (LSHB-CT-2004-512143). U.d.L and T.S.J. are supported by DIAMONDS and by grants from the Danish National Research Foundation and the Danish Technical Research Council (Systemic Transcriptomics in Biotechnology). L.J.J. is supported by BioSapiens, an EC FP6 Lifescihealth NOE (LSH6-CT-2003-503265). S.M. is supported by a fellowship from the Swiss National Science Foundation, and B.T.W. is supported by Sanger Postdoctoral and Canadian NSERC fellowships.