We developed an algorithm for combining sequence and expression data in order to identify cycling genes 12. The algorithm uses probabilistic graphical models, and in particular Markov random fields, to combine these data sources. Genes are represented as nodes in the graph and are connected by edges to other genes (in the same species and all other species), based on their sequence similarity as determined by a BLAST score (Figure 1a). Each node (gene) is assigned an initial cycling score that is determined from expression data using a method from de Lichtenberg and coworkers 13. Starting with this score, we propagate information along the edges of the graph until convergence. Thus, if a node with a medium to high score is connected to a set of nodes with high scores, then the information from the neighboring nodes can be used to elevate our belief in the assignment of this node, and vice versa. This method allows us to identify several cycling genes that can be missed in an analysis focused on a single species as a result of expression noise (Figure 1 and Additional data file 1 [Supporting Figures 1 to 3]). Similarly, genes with marginal scores that are only connected to low scoring genes can be filtered out of the cycling gene lists. Once the algorithm converges each gene is assigned a posterior cycling score between 0 and 1. For comparison reasons, we select for each species a set of genes with roughly equal size to those used in the original reports (although the identity of these genes is different), remove all other genes from the graph, and consider only the subgraph induced by the selected genes. This graph is analyzed to identify multidomain homology cliques 14 (Figure 1b-e). Each of these cliques is then analyzed to determine the set of species included. These findings are reported as three cell cycle conservation (CCC) sets with conservation across two (budding and fission yeast), three (yeasts and human cells), or all four species.

See Materials and methods and Additional data file 3 for further details on our graph-based algorithm and on clique analysis. Also see our supporting website 15 for a complete list of genes identified using our algorithm.

Our method combines expression and sequence data. This raises an obvious question; is the quality of our lists comparable to the quality of previous lists that relied on expression data alone? In other words, does our method sacrifice the accuracy with respect to the set of cycling genes in each species in order to obtain a larger set of conserved genes?

A possible way to assess the quality of such lists is by comparing them with other high-throughput data sources 13. For example, protein-DNA binding data are available for nine budding yeast transcription factors that are known to be involved in cell cycle specific transcription 16. It is expected that many cycling genes would be bound by these factors. When comparing the genes in our list with the original list 2, we find that both exhibit a threefold enrichment for these interactions compared with a random gene list (Figure 2a). Stationary phase expression experiments yield similar results (Figure 2b). Similarly, both our list and the original list 1 of cycling human genes are enriched for binding of known cell cycle factors (Nrf1 and E2f2; Additional data file 1 [Supporting Figure 4d]). Genes on both lists exhibit lower expression levels in nonproliferating tissues (Figure 2c) and higher expression levels in cancer cells (Additional data file 1 [Supporting Figure 4c]). Expression data for fission yeast and Arabidopsis support our list for these species as well (Figure 2d-e).

Combined, these results indicate that the species-specific lists derived using our method are comparable in quality to those of previously reported cell cycle gene lists. Additional data file 2 (Supporting Tables 1 to 3) presents the percentage overlap between the lists of cycling genes identified using our method and previously reported cycling gene lists for the four species.

Figure 3a presents the number of conserved genes for the different evolutionary distances represented in our datasets. About 21% of the budding and fission yeast cycling genes reside in cliques containing genes from these two species (CCC2). When adding human genes, roughly 10% of cycling yeast genes and 8% of cycling human genes are included in such cliques (CCC3). Finally, between 5% and 7% of cycling genes in all four species are conserved in sequence and expression (CCC4). Additional data file 2 (Supporting Tables 4 to 10) presents the list of genes assigned to CCC4 and CCC3 for each of the species. We note that although our original sequence similarity criterion was based on BLAST e-values, following the clique analysis the resulting sets are in very good agreement with curated homology databases 17. For example, 82% of budding yeast genes in CCC2 have a curated fission yeast homolog in CCC2. Similarly, 82% of fission yeast genes in CCC2 have a curated budding yeast homolog in CCC2. See our supporting website 15 for complete homology references.

To test the agreement of our conserved lists with complementary high-throughput datasets, we have repeated and extended our analysis discussed above but focusing only on genes included in CCC3 and CCC4. As Figures 2 and 3 and Additional data file 1 (Supporting Figure 4) show, CCC3 and CCC4 genes exhibit much stronger cell cycle characteristics when compared with the original set of cycling genes for each species. For example, in a protein-protein interaction dataset for budding yeast 1819, genes in CCC3 are involved in ten times more pair-wise interactions when compared with a random set of similar size from the full set of cycling genes (Figure 3c). This indicates that these genes have long been involved in the same function. Similarly, the percentages of human genes bound by two cell cycle transcription factors are much higher for the CCC3 and CCC4 sets (Nrf1 and E2f2 20; Additional data file 1 [Supporting Figure 4d]). Also, for humans the CCC3 and CCC4 sets are much more repressed in several nonproliferating tissues when compared with the full set of cycling human genes (Figure 2c). Similarly, CCC4 genes exhibit stronger cell cycle characteristics in Arabidopsis and fission yeast expression experiments (Figure 2d-e).

We have also repeated our analysis by comparing our lists with subsets of cycling genes with high amplitude in each of the four species. As shown in Additional data file 1 (Supporting Figure 5), high amplitude genes exhibit similar cell cycle characteristics to the CCC3 and CCC4 sets for human and plants. However, for the two yeasts these high amplitude genes are more similar to the full set of cycling genes. This indicates that expression analysis alone cannot be used to identify this core set of genes.

To further validate our findings of a large overlap between the cycling genes in the two yeast species, we turned to motif analysis. Several transcription factors are conserved between budding and fission yeast 21. A possible explanation for expression conservation (or lack thereof) is in the conservation (or lack of conservation) of a binding motif for these cycling genes.

We started by looking at genes bound by the budding yeast factor Swi6, which regulates transcription at the G1/S transition 22. We extracted three lists for this factor. The first, denoted BY6, contained cycling budding yeast genes in CCC2 determined to be bound by Swi6 23. The second list, denoted FY6C, contained fission yeast genes that both were in CCC2 and had homologs in BY6. These genes were determined to be cycling and conserved by our method. The third list (FY6NC) contained noncycling fission yeast genes with cycling budding yeast homologs bound by Swi6. This latter list serves as a negative control because it contains genes that have lost their cycling status between the two species. Four motif finders were run on each dataset; SOMBRERO 2425, BioProspector 26, Consensus 27, and AlignACE 28 (see Materials and methods, below, for details). All four motif finding algorithms were able to identify the Swi6 motif in BY6 and FY6C, indicating that this motif is conserved between the two species, at least for some of the conserved cycling genes (Additional data file 1 [Supporting Figures 6 and 7]). In sharp contrast, none of these motif finders was able to identify the Swi6 motif in the upstream regions of genes in FY6NC.

We have extended the motif analysis discussed above to study ten additional transcription factors that were determined to play a key role in regulating cycling genes in budding yeast 316. For each of these factors we extracted all cycling budding yeast genes determined to be bound by this factor 23 and their fission yeast homologs. As we did for Swi6, we further divided the fission yeast genes into two sets; the first contains fission yeast genes in CCC2 and the second (a negative control list) contains noncycling fission yeast homologs of cycling budding yeast genes. Next, we ran the four motif finders on each dataset.

The results are presented in Table 1 and Additional data file 1 (Supporting Figures 6 to 17). In Table 1 we report on the number of motif finders that identified the correct motif for each factor and on the percentage of genes in the set that contained this motif. Similar to the results obtained for Swi6, the other two G1/S factors, namely Swi4 and Mbp1, exhibit the optimal motif conservation pattern; the expected motifs are found in both the fission yeast cell cycle genes and the positive control of conserved budding yeast cell cycle genes, but are not found in the negative control set of noncycling fission yeast genes. Motif scan analysis (Additional data file 2 [Supporting Table 11]) confirms the results for these factors. For G2/M, the Fkh2 sets display similar, although less significant, pattern (two of four motif finders identified the correct motif for the cycling set). However, Fkh1 and Fkh2 motifs also appear, although less strongly, in the negative control sets. In total, FKH-like motifs are present in eight of the 11 negative control datasets. The M/G1 phase analysis is complicated by small dataset size. This may result from the lack of conservation between the two species for this phase 21. As a result, motif match for this set is either weak (Swi5) or nonexistent (Mcm1 and Yox1).

To further validate that genes in CCC3 and CCC4 are core cycling genes, we studied their importance using deletion data. Surprisingly, only 15% of cycling yeast genes are essential in rich media conditions 29, which is roughly equal to the overall percentage of essential yeast genes (18%). However, as Figure 4 shows, 35% of budding yeast genes in the CCC3 list and 46% of the genes in the CCC4 lists are essential. To test whether similar result could be obtained using only sequence data (without expression data for the other species), we extracted from the full list of cycling budding yeast genes those with homologs in all other species, without taking into account their cycling status in these other species. Although this increased the percentage of essential genes (to 27%), these percentages remained well below those achieved for CCC4, which uses the expression data.

We have also carried out similar analyses for human genes using data from RNA interference (RNAi) experiments 30. In these experiments 24,373 genes were knocked down using RNAi and assessed for phenotypic influence on cell growth. For 1,152 (4.7%) of the genes, the resulting knockdown cells presented phenotypic growth defects. As Mukherji and coworkers 30 note in their report, roughly 6% of cycling human genes reported by Whitfield and colleagues 1 are included in this list. Similar to the process we conducted in yeast, we considered sequence data only and extracted from the Whitfield list those genes with homologs in the other three species. For this list, the percentage of genes increases to 10%. Again, the most enriched lists are obtained when using the CCC3 and CCC4 sets. For these, the percentage climbs to 16% (CCC3) and 17% (CCC4). These findings highlight the importance of the conserved set and support our conclusion that it contains key cycling genes.

To determine cell cycle regulated protein complexes conserved between these species, we searched for protein complexes with one or more subunits in the CCC3 set using high-throughput protein-protein interaction data. This type of data is thus far only available in budding yeast 1819. Additional data file 2 (Supporting Table 12) and Additional data file 1 (Supporting Figure 18) present some of the protein complexes that we identified. Some of these complexes are known to regulate important events in the cell cycle. For example, the origin recognition complex (ORC) is a well conserved complex that is involved in the initiation of DNA synthesis 31. Other examples are the cohesin complex, which is responsible for binding the sister chromatids during mitosis after S phase 32, and the ribonucleoside-diphosphate reductase (RNR) complex, which is involved in the maintenance of the cellular pool of dNTPs 33.

The CCC3 list gives us our first look at the conserved core of periodically transcribed genes across evolution. Even though CCC3 contains relatively few genes (0.4% to 1.3% of the total number of genes for each species), many of these genes play a role in key processes required for growth. Using Gene Ontology (GO) analysis 34, we identified categories that were enriched in this set. For budding yeast these categories include cell cycle (P = 3 × 10^-15), DNA replication (P = 2 × 10^-13), and mitosis (P = 1 × 10^-7). Similar enrichments were found for human conserved cycling genes and for fission yeast. For example, cell cycle (P = 5 × 10^-17), DNA replication (P = 7 × 10^-14) and cell division (P = 2 × 10^-9) are enriched in humans, and cell cycle (P = 10^-9) and chromatin assembly/disassembly (P = 10^-9) are enriched in fission yeast. Figure 3 and Additional data file 2 (Supporting Tables 13 to 21) present P values for the various GO categories.

Some categories were more enriched in the CCC3 set than in the full list. For these categories, CCC3 contains a disproportionate number of genes when compared with the overall percentage of cycling genes. This indicates that many of the genes associated with these functions have been conserved in cyclic expression between the species. These include categories related to DNA metabolism (P = 5.7 × 10^-12 for CCC3 and P = 1.1 × 10^-6 for the full list) and chromatin assembly (P = 3.7 × 10^-5 versus P = 0.01). In contrast, there are a number of categories that are much more enriched in the full list, indicating that they have probably evolved, or at least greatly expanded, in the individual species. These include categories such as mitosis for fission yeast (P = 1.6 × 10^-4 versus P = 3.6 × 10^-7) and the cell wall category, which exhibits a great deal of species-specific variation between the budding yeast, the fission yeast, and metazoans 35. For the human list, DNA repair and chromosome segregation were more significantly enriched in the full set (P = 5.9 × 10^-4 versus P = 7.3 × 10^-5, and P > 0.1 versus P = 9.0 × 10^-8, respectively). Although these functions are conserved across organisms, our analysis indicates that many of these genes are cycling only in human cells, perhaps indicating that these functions have been adapted to accommodate the longer cell cycle.

Partial functional knowledge is available for all but one (YPL247C) of the 72 budding yeast genes on CCC3. Sixteen of these genes encode products that are involved in DNA replication and another 23 are involved in chromosome organization and biogenesis. These include structural components (Mcms, tubulins, and histones) as well as regulatory proteins (cyclins, Cdc20, and Cin8). The mcm2 (cdc19) and mcm6 genes were previously known to be cyclic subunits of the highly conserved Mcm pre-replication complex in fission yeast 536. Our combined analysis indicates that two other genes (mcm3 and mcm5) may also be periodic, similar to the budding yeast and human Mcm subunits. Another large class of conserved cyclic genes is involved in chromosome segregation (ASE1, KIP1, NUM1, and STU2) and cytokinesis (MOB1, HOF1, KEL2, and IQG1). In addition, the list includes factors that affect transcription globally (ARP7 and TUP1) and specifically (ACE2, FKH2, and HCM1). Interestingly, the S phase specific transcription factor Hcm1 has a conserved cyclic transcript, as do 22 of its predicted targets 3. The fact that nearly 30% of the budding yeast CCC3 genes are potential targets of Hcm1 is consistent with the known role of Hcm1 in regulating genes involved in chromosome dynamics 337.

There is only a small number of genes in CCC3 that are not obviously involved in cell cycle specific processes. These genes include three involved in metal homeostasis (SMF2, SMF3, and CTH2), some cell wall proteins (FIG2, AGA1, and SED1) and alkaline phosphatase (PHO8). These gene products could be involved in unknown aspects of the cell division cycle, or they could be evolutionarily related to other cell cycle proteins.

We applied a probabilistic graphical model to combine microarray expression data and sequence data for identification of cycling genes, as described in Lu and coworkers 12. We used microarray expression data reported by Spellman 2, Rustici 4, Oliva 5, Peng 6, Whitfield 1, and Menges 7 and their coworkers. We downloaded protein sequences from the National Center for Biotechnology Information website 41. The method starts by using gene specific expression data to compute a cycling score based on both the amplitude and periodicity 13. We run BLASTALL 42 to calculate bit scores between all pairs of sequences, as was done by Sharan and coworkers 43. We use a Markov random field to model the joint likelihood of the data. The (hidden) cycling status of each gene is represented by a node in the graph, and two nodes are connected by an edge if the bit score for the two genes is above a threshold. We define potential functions on nodes to capture information from the cycling scores, where we assume the scores of cycling and the noncycling genes follow a mixture of extreme value distributions, and define potential functions on edges to capture the correlation of cycling statuses between similar genes. The posterior beliefs of the cycling status of the genes are estimated using loopy belief propagation algorithm. Finally, we rank the genes by their posterior and use the name number as were used in the original papers (500 for Arabidopsis, 800 for budding yeast, 600 for fission yeast genes, and 1,000 human genes). See Additional data file 3 for complete details.

Genes identified as cycling in each species were used to identify conserved sets of cycling genes. This is done using the Markov clustering algorithm (MCL) 14 as follows. First, we start with the graph of all cycling genes. Edges in the graph are defined based on the bit score cut-off, as mentioned above. Second, for any connected subgraph in this graph, we use MCL to break it into smaller subgraphs if it has more than 30 nodes. Third, repeat the previous step until all connected subgraphs have at most 30 nodes.

Next, we assign genes to different conserved sets based on the other species represented in the subgraph to which they belong. The numbers of genes in the conserved sets are shown in Figure 3, in which the sets are organized as a tree reflecting the evolutionary relation between the four species.

For each gene in the lists, the appropriate intergenic region was extracted from the budding or fission yeast genome. Four motif finders were run on each dataset: SOMBRERO 2425, Consensus 27, BioProspector 26, and AlignACE 28. Both SOMBRERO and BioProspector require a background model, and the background was constructed from all intergenic regions (in the appropriate genome) for both cases. SOMBRERO was run using default settings, and simultaneously using all known yeast motifs as an appropriate source of prior knowledge. Consensus and BioProspector were run using default settings requiring the top ten motifs to be reported. AlignACE was run using default settings (using a seed motif length of 10), and provided with the background intergenic GC content (31.45% for fission and 35.3% for budding). See Additional data file 3 for further details.

GO enrichment P values were computed using STEM 44, which relies on hyper-geometric distribution. Corrected P values were computed by permutation analysis using STEM.