We applied our method (Figure 1) to 113 transcription factors using the ChIP data of Lee et al. 10 and obtained over-represented motifs for each transcription factor (Table 1 and see also Table A at 19). First we took the intersection of these target genes and cell-cycle genes (Figure 1). Next the method enumerated all possible 6-mer to 9-mer motifs and selected only motifs that were over-represented in the upstream regions of the intersection genes. Only 21 of the 113 TFs had over-represented motifs at our statistical criteria (see 19); most of the other transcription factors do not have a cell-cycle specific role. However, our criteria were very stringent (see Materials and methods) to avoid false positives.

For many of the transcription factors in Table A at 19, there is existing knowledge about their mechanisms of DNA binding. From this knowledge, we can see that the relationship between a TF and its over-represented motifs fall into three categories: direct binding; piggy-back binding; or cross-binding (Figure 2). For example, it is known that Fkh2 (and to a lesser extent Fkh1), together with Mcm1, recruits Ndd1, and they control the transcription of G2/M genes 1420 in the yeast cell cycle. GTAAACAA is known to be the direct-binding motif of both Fkh1 and Fkh2 21. Our results show that GTAAACA[A] ([A] is either A or N) is indeed over-represented in Fkh1 and Fkh2 ChIP data (Table 1). But GTAAACA is also over-represented in the ChIP data for Ndd1. Ndd1 does not directly bind to DNA but interacts with Fkh1 or Fkh2, both of which bind directly to DNA 20. Thus, GTAAACA is a direct-binding motif for Fkh1 and Fkh2, but a piggy-back binding motif for Ndd1 (Figure 2a,b).

Mcm1 is involved in several biological processes including cell cycle control 14202122. The transcription factor can form a homodimer and this is reflected in the dyad symmetry of its binding motif, TTACCNAATTNGGTAA 23, which is often referred to as the ECB 24. Because we searched for 6-mers to 9-mers, we did not extract this 16-mer, but we did extract sub-sequences. For example, our motif TTTCCTAA (Table 1) is exactly half of the motif (TTTCCTAATTAGGAAA) found by Liu et al. 15 and is almost identical to half of the ECB 23. Interestingly, the motif ATAATTA was associated with Mcm1 in M/G1 phase. This motif is likely to correspond to (T/C)AATTA, the binding site of the proteins Yox1 and Yhp1, which are recently characterized binding partners of Mcm1 in M/G1 25. Thus, ATAATTA may be a cross-binding motif of Mcm1 via Yox1 and Yhp1 (Figure 2c). Yox1 and Yhp1 were not among the 113 transcription factors assayed by ChIP 10, and so are not in the set of transcription factors for which we determined over-represented motifs. However, Horak et al. 26 did ChIP experiments for Yox1 and Yhp1. We searched for over-represented motifs for their ChIP targets, and in addition, for targets determined on the basis of mutagenesis experiments by Pramila et al. 25. While we found the putative binding motif in the dataset of Pramila et al., we could not find it, or any similar motifs, in the dataset of Horak et al. Although this is disappointing, it is consistent with the fact that Pramila et al. and Horak et al. largely disagreed on the genes regulated by Yox1 and Yhp1; Horak et al. defined 320 targets, whereas Pramila et al. defined 28 targets. Only two of these targets overlapped, whereas the expectation from picking random genes is an overlap of 1.5.

SBF, a complex containing Swi4 and Swi6, predominantly controls the expression of budding and cell-wall genes, and MBF, a related complex composed of Mbp1 and Swi6, functions in DNA replication 13. The binding motifs of SBF and MBF (called SCB and MCB motifs) 27 are CRCGAAA and ACGCGT 3, respectively. In our results (Table 1), the most prominent motifs of the DNA-binding proteins Swi4 and Mbp1 27 are CGCGAA and ACGCGT, consistent with the known motifs. We find that Swi6, which is a non DNA-binding cofactor of Swi4 and Mbp1 27, has the same motifs, CGCGAA (see 19) and ACGCGT. We interpret these as piggy-back binding motifs for Swi6 via Swi4 and Mbp1. We also find the novel variant CGCGTC, which is associated with as many as nine TFs, including SBF and MBF (see 19). Conlon et al. 28 also pointed out this GCGTC motif in cell-cycle genes. We describe further studies of the CGCGTC motif in additional data provided at 19. As expected, our integrated approach has substantially enhanced the signal-to-noise ratio. Such integration may not be necessary for the analysis of yeast, but will be crucial for analysis of higher eukaryotes. Among the nine TFs, Ash1 is thought to be a regulator of mating-type switching 29. Hence this points to a possible connection between cell cycle and mating-type switching through the combination of SBF/MBF and Ash1. The binding motif of Ash1 is known to be YTGAT 29, but we found [A]GGCAC[C] and GCGGCA. Probably, these putative motifs are indirect binding motifs of Ash1, which suggests that Ash1 may cooperate with an unknown factor (or factors) through these motifs and in the end with SBF/MBF as well (see also 19).

SCB- and MCB-like motifs were also found as over-represented motifs from the ChIP data for Stb1 and Ste12 (Table 1). Stb1 binds to Swi6 in vitro and is thought to interact with the Swi6 subunit of SBF and MBF to regulate transcription in vivo 30. In our results, the most prominent motifs of Stb1 were CGCGAAAA and ACGCGA, which closely resemble the SCB and MCB, respectively. Together, these results imply the presence of the complexes Stb1+Swi6+Swi4, and Stb1+Swi6+Mbp1, which are different from the standard complexes SBF (Swi6+Swi4) and MBF (Swi6+Mbp1). These Stb1 motifs are piggy-back binding motifs via SBF and MBF. In the case of Ste12, in contrast, there is no evidence that Ste12 binds to SBF or MBF. Furthermore, many of the genes that have both the Ste12 direct-binding motif and the SCB- and MCB-like motif (CGCGTC) are bound by Ste12, SBF and MBF (see below). Thus, we suggest that the SCB- and MCB-like motif is a cross-binding motif of Ste12; that is, there is a group of genes where Ste12 binds to its direct binding site and SBF or MBF binds to its direct binding site in the same promoters.

Ste12 is involved in pheromone response and filamentous growth 223132. The known binding motifs of Ste12 are ATGAAAC and TGAAACA (called the PRE) 3334. Our results match these known motifs perfectly (Table 1). In addition, we find that Dig1 is associated with these motifs. Dig1 is an inhibitor of Ste12 35. Thus, the PRE motifs are presumably piggy-back binding motifs for Dig1 via Ste12. This result suggests that the Dig1+Ste12 complex binds DNA, despite the fact that Ste12 activity is inhibited, consistent with the results of Olson et al. 35.

Ace2 and Swi5 are transcription factors that function at the M/G1 boundary 1436. The direct-binding motifs of Ace2 and Swi5 are variously stated as ACCAGC 3738 or RRCCAGCR 1 or CCAGCA 39 (for Swi5). Our results are consistent with these motifs (Table 1).

Met4 and Met31 are involved in sulfur amino-acid metabolism 40 and may have a transcriptional role in cell-cycle control 12. The transcriptional mechanisms differ between targets but two main mechanisms have been suggested 40. First, the DNA-binding protein Cbf1 binds to its motif TCACGTC, and tethers a Met4+Met28 complex to this site. Second, the DNA-binding protein Met31 (or Met32) binds to its motif AAACTGTG, and likewise tethers a Met4+Met28 complex to this site. In our results, TCACGTG appears from the Met4 ChIP data, which matches the known binding motif of Cbf1, suggesting that TCACGTG is a piggy-back binding motif of Met4 via Cbf1. The motif ACTGTGG also appears in the Met4 ChIP data. This motif is similar to the known binding motif of Met31/Met32 and the motif AAACTGTGG of Spellman et al. 1. Thus this motif is presumably a piggy-back binding motif of Met4 via Met31/Met32. Using the Met31 ChIP data, we find the over-represented motif TGTGGC, which overlaps the motifs of Spellman et al. (AAACTGTGG) and Tavazoie et al. 2 (AAANTGTGGC), and represents the direct-binding motif of Met31. Note that in the studies of Spellman et al. and Tavazoie et al., which are based on expression clustering and motif searching, one has to try and identify the TF from other kinds of data after finding an over-represented motif. In our case, the ChIP data for the relevant TF gives this information directly.

Finally, we found that Mth1, which is involved in glucose signal transduction 41, binds to several cell-cycle genes, including some of the histone genes, and genes involved in budding and polarized growth. It was somewhat surprising to find Mth1 as a controller of cell-cycle genes. It is thought to act as a co-factor with other transcription factors involved in glucose signal transduction, and is not known to bind DNA directly. We found the motifs CAGCAG and CGCGTC over-represented in Mth1 ChIP data (see further investigation at 19). We presume that Mth1 is a piggy-back or cross-binding transcription factor for these motifs. Possible candidates for the direct-binding factor are Swi5/Ace2 (for CAGCAG), or SBF or MBF (for CGCGTC), or Rgt1, as Rgt1 has a GC-rich binding site 42 and is also involved in glucose signal transduction 41. As glucose accelerates the growth of yeast cells, and therefore accelerates the cell cycle, it is possible that Mth1 is used to control a rapid cell cycle in response to the glucose growth signal. Indeed, Heideman and co-workers 43 have shown that expression of CLN3, a major activator of the cell-cycle program, is controlled in part by the availability of glucose.

The serial regulation of the yeast cell cycle is thought to occur as follows 14: MBF (Mbp1+Swi6) and SBF (Swi4+Swi6) bind to the motifs ACGCGT 3 and CRCGAAA 3, respectively, to control the expression of late G1 genes 27; Fkh1/Fkh2 and Mcm1 bind to GTAAACAAA 21 and TTACCNAATTNGGTAA 23, respectively, and recruit Ndd1 to control G2/M genes 20; and Mcm1 and Ace2/Swi5 bind to the ECB motif and RRCCAGCR 1, respectively, to regulate M/G1 genes 242536. This general model is supported by many experiments 110131444. Transcriptional control in S and S/G2 phases is less well characterized, but some studies suggest the involvement of SBF and Fkh1/Fkh2 101445.

Because we are interested in combinatorial control, we used the procedure shown in Figures 1b-d to search for combinations of TFs and motifs that are specific to each of the cell-cycle phases. For G1, we confirmed the regulatory role of MBF and SBF: Mbp1 and Swi6, and Swi4 and Swi6 are predicted to bind to ACGCGT and CGCGAA (and variants) respectively (Table 2). However, we further predict that Stb1, Mbp1 and/or Swi4 are associated with ACGCGA. Taken together with the results above, this suggests that putative complexes of Stb1+Swi6+Mbp1, and/or Stb1+Swi6+Swi4 bind to this motif and regulate some G1 genes. Lee et al. also found a significant number of targets bound by both Stb1 and Swi4 in G1 phase 10.

G1 phase also gave us the combinations {SBF, Ste12} and {MBF, Ste12} (see also data at 19). Examples of genes in these categories include PCL2, GIC2, MSB2, CRH1 and SRL1. At least some of these are genes involved in a normal G1 phase, but are also involved in mating and the pheromone response. For instance, GIC2 is involved in polarized growth, which is needed for normal budding, but is also needed for mating. Perhaps surprisingly, some (for example, PCL2, GIC2) of these genes are strongly induced by alpha-factor (which acts via Ste12), while others (MSB2, SRL1) appear to be repressed.

In S phase, we find the novel combination {SBF, Fkh2, Hir1} (Table 2). A few genomic analyses 1045 have indicated the involvement of SBF and Fkh1/Fkh2 in this phase, and one of them suggested that this combination may be associated with the regulation of histone genes 45. To clarify our results, especially with regard to histone regulation, we divided the target genes of this combination into histone genes and other genes (which are involved in budding, cell-wall synthesis, microtubules and the spindle-pole body). This division did not conflict with expression coherence, because scores (the average standard deviation scores) measuring expression coherence for both sets became lower as a consequence of this division. For the latter (that is, non-histone) genes, we found that the over-represented TFs (see Materials and methods) were Swi4, Mbp1 and Fkh2, but not Hir1 (p-value <10^-3). Thus SBF and Fkh2 most probably bind to these motifs, which suggests that both these transcription factors probably regulate budding, cell-wall synthesis, and spindle-related genes in S phase.

The nine histone genes are the other major class of S-phase regulated genes. These are organized into five transcription units consisting of four divergently transcribed pairs (HTA1-HTB1, HTA2-HTB2, HHT1-HHF1, HHT2-HHF2) and HHO1. Histone mRNAs are regulated in at least three ways. First, there is cell-cycle regulated mRNA stability, such that the message is only stable during S-phase. Second, there is a negative transcriptional element that represses transcription at inappropriate times 46; this repressive system involves the HIR genes - HIR1, HIR2 and HIR3. Third, there is a positive transcriptional element that induces histone mRNA synthesis during S-phase 47. Despite the fact that histones were the first cell-cycle regulated genes discovered in yeast (and perhaps in any organism), the positive regulatory element and its transcription factor remain poorly characterized. The positive regulatory region includes repeats of the sequence GCGAAA 47, which closely resembles the SBF-binding site. However, histone mRNA abundance continues to oscillate through the cell cycle even in Swi4, Mbp1 and Swi6 single mutants 4849, which argues that SBF is not essential for regulation, possibly because of the other two modes of regulation (mRNA stability, repression). Furthermore, as we suggest below, there might also be other redundant activators.

For instance, we identified Met4, in addition to Swi4, Hir1 and Hir2, as an over-represented TF for the histone promoters. The association of Met4 was completely unexpected because Met4 is thought to be a regulator of amino-acid metabolism 40, and the involvement of Met4 in histone gene regulation has never been reported. Furthermore, we found that the binding motifs (detected at the motif finding step) for Met4 also exist in some histone promoters. Thus Met4 may be a novel regulator of histone expression.

There have been three genome-wide ChIP experiments directed at the targets of MBF and SBF 101314. All five histone transcription units have been associated with SBF (and/or MBF) in at least one of these studies, and the HHO1 and HTA1-HTB1 units were found in all three studies. The SCB-like motif (GCGAAA) is clearly present in four of the five transcription units. Thus the ChIP and motif evidence for regulation of histones by SBF is very strong. Two of the genome-wide ChIP studies also looked for Fkh1 and Fkh2 targets. The ChIP data of Simon et al. 14 show two of the five histone transcription units as targets of Fkh2 (p-value in ChIP data <0.02), and two more as possible targets (p-value <0.07) (with the exception being the HHT2-HHF2 unit), and all four of these units have clear Fkh1/Fkh2 motifs. On the other hand, the data of Lee et al. 10 show only the HHT1-HHF1 transcription unit to be an Fkh2 target (p-value <0.01), so the evidence for involvement of Fkh1/Fkh2 is suggestive but not conclusive. Finally, our analysis confirms the binding of Hir1 and Hir2, the repressive factors 46. In summary, we suggest that the positive transcription factor for the histone genes is SBF (and to some extent MBF), probably in combination with Fkh2.

Our results are consistent with the result of genomic analyses 1014 for S/G2 and the standard model for G2/M. In S/G2, we confirmed both Fkh2 and its binding motifs as an over-represented TF and over-represented motifs (Table 2). In G2/M, we found {Fkh1/2, Mcm1, Ndd1}, in agreement with the standard model (Table 2). It has been suggested that Fkh2 has a more prominent role than Fkh1 in G2/M transcription 20. Our analysis agrees, as the p-value (5 × 10^-7) of Fkh2 was much more significant than that (4 × 10^-3) of Fkh1 for over-represented TFs in promoters having both GTAAACA (the Fkh motif) and TTCCTAA (part of the Mcm1 motif).

In M/G1, we found Mcm1 and its motifs (Table 2), in agreement with the standard model. We also found some new or unusual combinations. First, we found the combination of Mcm1 and Swi4 (and Yox1, see above), targets of which include SWI4, UTR2 (involved in cell-wall organization and polarized growth) and AGA1 (encoding a cell-wall protein). The M/G1 interval is a crucial time for the cell wall, because it is then that the bud separates from the mother. It appears that at least some cell-wall genes are under the dual control of the M-phase regulator Mcm1 and the G1-phase regulator Swi4. The dual regulation of the gene for the critical cell-cycle transcription factor Swi4 by Mcm1 and Swi4 has been shown previously 24 and is quite intriguing, because it creates two feedback loops 25: induction of SWI4 by Swi4 is a positive feedback loop, and induction of YOX1 by Swi4 negatively regulates Mcm1 activity and so is a negative feedback loop.

In M/G1, we also found a novel combination {Swi5, Ste12, Dig1}. Whereas Swi5 is involved in the cell cycle, Ste12 regulates mating and pseudohyphal growth 35. M/G1 is a critical phase for these processes. Targets of this combination are TEC1 and CHS1, and perhaps also AMN1, KAR4, GFA1, SST2 and AGA1. Many of the genes listed (for example TEC1, KAR4, SST2) are important for mating or pseudohyphal growth, and the other genes listed may also be involved.

Full methods are described in 19. This first step is represented briefly in Figure 1a. Among 4,339 nonredundant promoter sequences, we identified target genes (promoters) of a TF from each of 113 ChIP datasets 10 with p-value <10^-2. Next we took the intersection of these genes with each of six gene sets of Spellman et al. 1, that is, all the cell-cycle regulated genes and subclasses of G1, S, S/G2, G2/M and M/G1 phase genes. We defined each of these intersection sets as a foreground set. As a background (control) set, we chose the intersection of non-target genes of a given TF (with p-value in ChIP data >0.8), and non-cell cycle genes (genes other than all the cell-cycle regulated genes). From the 113 × 6 foreground-background pairs, we excluded pairs where the foreground set had fewer than 10 genes or the background set had fewer than 500.

For each foreground-background pair, we searched for over-represented motifs (6-9 bp) by a word-counting strategy 45253. The main differences of our approach from others are to use a set of non-target genes from ChIP data as a background set, and to use the contingency table test for manipulating the background set, whose size can be small. For all possible motifs, we calculated the statistic with Yates correction of the 2 × 2 contingency table test as follows:

where N is the sum of a, b, c, and d; a and b are the occurrences of a given motif and other motifs than the given motif (non-motif) in a foreground set, respectively; c and d are the same in a background set. We counted these occurrences in both strands of the 600-bp upstream sequences. Then we calculated the p-value (without multiplicity correction) to see if each motif is significantly over-represented at the p-value threshold of 10^-8. We also required that: rank, according to p-value, of motifs in each data pair must be within the top 10; the number of upstream sequences with a motif must be greater than 25 in all the cell-cycle regulated genes; and a is 10 or more. We also excluded simple repeat motifs like AAAAAA, ATATATA and so on. Furthermore, for the list of motifs obtained from each foreground-background pair, we merged similar motifs into extended ones (see 19).

For this second step (Figure 1b), we took each of the phase-specific gene sets (the G1, S, S/G2, G2/M and M/G1 genes as classified by Spellman et al. 1) as a foreground set, and the intersection of the non-cell-cycle genes and genes with constant expression profiles (see 19) as the background set. We then searched all possible order-1 combinations (single motifs), order-2 combinations (pairs), and order-3 combinations (triples) of motifs found. For each of the combinations, we calculated the statistic of Equation (1), where a and b are now the numbers of upstream sequences with and without a given combination in a foreground set, respectively; c and d are the same in a background set. Then we calculated the p-value (without multiplicity correction) to see if each combination is significantly over-represented at the p-value threshold of 2 × 10^-15 and 2 × 10^-5 for G1 and the other phases, respectively. We also required at least 60 occurrences of an upstream sequence for the G1 set, and at least eight occurrences for the other phase sets. Finally, for the list of obtained combinations, we merged associated combinations into extended combinations (for example merging (M1, M2) and (M2, M3) into (M1, M2, M3), see 19). The p-value threshold for finding an associated combination pair was set to 2 × 10^-5.

This procedure (Figure 1c) checks the coherence of expression profiles over time for genes that have a given motif combination in their upstream sequences. For measuring the coherence, the average standard deviation score is used:

Score = E_i(σ_g(X_i,g)),

where X_i,g is the normalized expression level of gene g at time i, σ_g is the standard deviation over genes, and E_i is the average over time. The lower the score, the closer the expression profiles are to the average. We calculated the score for genes having a motif combination in the upstream sequences for the data of Cho et al. 44. We compared the score of genes having a motif combination with that of a phase-specific gene set. We selected only combinations whose genes had smaller scores than the phase-specific gene set.

This algorithm extracts TFs that are bound to the promoters with a motif combination (Figure 1d), based on the hypergeometric model:

where K is the number of all promoters used and T is the number of promoters that are bound by TF_i (p-value in ChIP data <10^-2) among the K promoters, and k is the number of promoters with a motif combination from a phase-specific gene set and t is the number of promoters that are bound by TF_i among the k promoters. For each of 113 TFs in ChIP data, this algorithm calculates the p-value

and then outputs TFs binding to a significant number of the promoters. We call such TFs over-represented TFs.

After checking the coherence of expression patterns, we used the above algorithm to find the set of over-represented TFs (p-value <10^-2; more stringent in G1, <10^-7) in the promoters with a motif combination comparing to binding TFs in all the promoters. From the procedure for finding over-represented single motifs, the set of TFs possible to bind the component motifs of a combination can be inferred. For each motif combination and its component motifs, we took the intersection of TFs from these two sets. We also kept those TFs for which there is additional experimental evidence even if they belonged to only one set. Thus we assigned each over-represented TF to each component motif. After arranging similar motif combinations (see 19), we obtained the final results as shown in Table 2.

Full descriptions of the methods, detailed investigations for the transcription factor Mth1 and the binding motif CGCGTC, lists of putative histone regulators, lists of putative target genes of motif combinations, and lists of genes whose promoters have a complicated motif structure are all available at our website 19. All the files are hierarchically structured and accessible on a web browser by tracking the hyperlinks.