We monitored genome-wide cell-cycle-regulated expression at 5-minute intervals for up to three cell division cycles, using whole-genome tiling arrays 11. The array is unique in interrogating every base pair in the genome on average six times and providing an 8-bp resolution for strand-specific probes. Two independent synchronization methods were used in order to obtain synchronous cultures (see Materials and methods; Additional file 1). Late G1 phase arrest was induced by exposure of bar1 cells to alpha factor, and by raising the temperature to 38°C for temperature-sensitive cdc28-13 mutant cells. Expression profiles for all genomic regions are provided in a database that is searchable by gene symbol or chromosomal coordinate 32.

To identify all transcribed sequences, we segmented along-chromosome expression profiles, applying an adaptation of the method described by Huber et al. 33 (see Materials and methods). In addition to protein-coding transcripts and infrastructural RNAs, we registered abundant expression of unannotated non-coding RNAs (Additional file 2). These unannotated expressed features comprise 523 antisense transcripts opposite protein-coding regions and 135 intergenic transcripts (Additional file 3). The length distribution of ORFs in these unannotated transcripts is within the range that is expected by chance. Hence, we find no evidence for the unannotated transcripts to be protein coding.

The average segment levels from each time-point were analyzed for periodic expression by two different computational methods 3435, as well as by visual inspection. The aim of this combination of methods was accurate and sensitive detection of cell-cycle-regulated transcripts (see Materials and methods). In order to validate our approach, we compared our gene list of periodic protein-coding genes to a benchmark set that comprised all known cell-cycle-regulated genes identified in single-gene experiments 3536. Our individual cdc28 and alpha-factor datasets were each better than most of the available ones 282930 (Additional file 4). Furthermore, our combined list of periodic protein-coding genes, despite being based on just two experimental datasets, performed almost as well in identifying the benchmark set of genes as that of Gauthier et al. 37, which integrated all available genomic datasets of cell-cycle-regulated genes performed to date (Additional file 4). Thus, our dataset and analysis method reproduced the previous data on cycling protein-coding genes.

Altogether, 598 periodic mRNAs, 37 cycling antisense RNAs, and 11 cycling intergenic transcripts were identified and ranked according to their peak time of expression (Figure 1; Additional file 5). Non-coding periodic transcripts were expressed in all cell-cycle phases (Figure 2; see Additional file 6 for the determination of the boundaries of the cell cycle phases). Overall, the peak times of antisense periodic expression were consistent with the waves of expression of periodic protein-coding genes 38. To characterize the newly discovered periodic ncRNAs, we overlapped them with regions of conserved RNA secondary structure 39. Despite their cell-cycle-regulated expression, the unannotated intergenic and antisense ncRNAs had little secondary structure (Additional file 6). Conversely, infrastructural ncRNAs, comprising tRNAs, rRNAs, small nuclear and small nucleolar RNAs, were highly structured but were not periodically expressed.

Studies in mammalian cells have suggested that antisense RNAs could regulate gene expression of their sense counterparts, whereby sense and antisense transcripts often exhibit expression correlation patterns 4041 and overlap in opposite directionality 42. We thus analyzed antisense RNAs in the context of the sense-antisense pairs (SAPs) of which they are a part. We categorized the pairs into four classes based on their expression coupling: 13 periodic antisense transcripts opposite periodic sense transcripts; 24 periodic antisense transcripts opposite non-periodic sense transcripts; 43 non-periodic antisense transcripts opposite periodic sense transcripts; and 443 non-periodic antisense and sense transcript pairs (Additional file 7).

The 13 periodic antisense transcripts opposite periodic sense transcripts were further subdivided based on the relative timing of expression of the sense and antisense transcripts. Considering the absolute difference between their expression peak times, two pairs (ALK1 and HSL1) cycle in-phase, whereas seven (CTF4, FAR1, HMS2, TAF2, TIP1, YNL300W and YPL162C) show anti-correlated expression (Additional file 8). Expression profiles of the other four SAPs (PRY3, YLR050C, YMR253C and YPL230W) had phase shifts between 0 and π.

Remarkably, several genes encoding important cell cycle regulators fall within the categories listed above (Figure 3a-c). Among them, FAR1 is important for mating pheromone-induced growth arrest and, together with cyclins CLN2 and CLN3, plays one of the key roles in the G1/S transition 43. FAR1 is expressed at the M/G1 transition and needs to be shut down in late G1 for the cell to pass the G1/S checkpoint. Its antisense RNA peaks starting from the late G1 phase and throughout the G1/S transition, when Far1 protein should not be present. TAF2, which is involved in transcription initiation, is expressed in late M and early G1 phase; its antisense transcript peaks in late G1 and further into S phase. The sense and antisense transcripts of CTF4, which shapes and maintains chromatin structure to ensure the passage through the S-phase checkpoint 44, are expressed in an anti-correlated manner, peaking in the G1/S and G2/M transitions, respectively. The CTF4 sense transcript appears to be transcribed from a bidirectional promoter shared with the antisense transcript of the neighboring gene, MSS18 (Additional files 6 and 9). Together these expression patterns suggest that some of the antisense transcripts may play a role in cell-cycle regulation.

We analyzed Gene Ontology (GO) categories of genes overlapped by antisense transcripts. Most of the protein-coding messages opposite the 37 periodic antisense transcripts (13 + 24) fall into GO categories linked with the process of cell division, including cell wall and organelle organization and biosynthesis, regulation of transcription, signal transduction and protein modification, carbohydrate metabolic processes, and cell cycle (Additional file 10). Surprisingly, 15 of the 37 sense transcripts are of unknown function. We carried out a similar analysis for the 43 non-periodic antisense transcripts opposite periodic sense transcripts. As expected, most of these cycling sense messages fall into cell-cycle-related GO categories, including genes involved in bud site selection and polarization (BUD9, GIC1), daughter cell separation from the mother (DSE2, CTS1), cell wall proteins, and so on (Additional file 7). Analysis of GO categories for the remaining 443 non-periodic SAPs did not show enrichment in any particular category, although almost a quarter of the genes have unknown function (Additional file 11).

We observed a statistically significant correlation (P < 0.002; 5 × 4 contingency table; χ² test) between the overlap patterns of the sense and antisense transcripts and the relationship of their expression profiles (Additional file 12). Altogether we distinguished five types of overlap within a given SAP: antisense transcript contains the transcribed message of its sense counterpart; the antisense transcript is contained within the sense transcript; the antisense transcript overlaps either the 3' or the 5' end of its sense partner; and the antisense transcript overlaps two distinct sense transcripts. The following patterns of overlap were over-represented compared to what was expected by chance. In 8 out of 13 periodic antisense transcripts opposite periodic sense transcripts, the antisense transcript is mainly contained within the protein-coding message; 2 of these 8 cycle in-phase, and 6 display opposite-phase expression. For 5 of 24 SAPs in which only the antisense transcript cycles, the antisense transcript contains the complete sense message, and for another 5, it overlaps 2 sense transcripts. In 15 of the 43 pairs in which only the sense message is cell cycle regulated, the antisense transcript overlaps the 5' end of the mRNA and in many cases extends further upstream.

To investigate sense and antisense expression in more detail, we also searched for putative TF binding sites (Additional file 6) and supported these predictions with the existing ChIP-chip data. TF binding site analyses are inherently non-strand-specific; however, our data on the temporal expression of the sense and antisense transcripts yield clues to the regulation of strand-specific expression. For example, ChIP-chip data and our motif analysis for FAR1 suggest binding of both the M-phase TF Mcm1 45 and the G1/S TF SBF 46 within the region spanned by 600 bases before and after the transcript. This evidence for SBF regulation of FAR1 contradicts the timing of expression of the sense transcript since FAR1 is expressed at the M/G1 transition and needs to be shut down in late G1. Our data show late-G1-specific expression of the FAR1 antisense transcript, thus providing a putative explanation for the presence of the TF binding site for SBF. Overall, our analyses indicate that the cycling unannotated transcripts have binding sites for the same set of TFs that drive sense transcription during the cell cycle (Additional file 6).

Altogether, 135 unannotated intergenic transcripts were detected in our dataset. Of these, 11 oscillate with mitotic progression (Additional files 5; Additional file 13c). As for the antisense transcripts, their peak in expression follows the waves of excitation in mitotic progression observed for protein-coding genes 38. To elucidate the role of these intergenic transcripts in cell cycle regulation, deletion strains for 10 of the 11 unannotated periodic transcripts were generated in both strain backgrounds. Growth curves of the deletion strains did not show significant lagging in cell doubling time after asynchronous growth in rich media for 28 hours at 30°C and 37°C. Lack of phenotype is consistent with our previous observations for the unannotated intergenic transcripts detected from asynchronous culture 11. This suggests that their deletion phenotypes have more subtle effects than those of many protein-coding genes.

Previous studies have identified a large number of annotated periodic transcripts. Compared to the integrated dataset of Gauthier et al. 37, our list contains 223 additional periodic protein-coding genes, of which 109 were also not identified by Pramila et al. 29 and Spellman et al. 30 (Figure 4; Additional file 14). Only 3 of the 109 have been shown to be periodically expressed in small scale experiments 47. GOslim analysis 48 showed that the biological function is unknown for 35 of these 109 genes, whereas 41 perform functions directly or indirectly associated with the regulation of the cell cycle, such as organelle organization and biogenesis, cytoskeleton organization and biogenesis, ribosome biogenesis and assembly, and so on (Additional file 15).

Of the 598 periodically expressed protein-coding genes, just 7 contain an intron according to the Saccharomyces Genome Database annotation: CIN2, MOB1, PMI40, RFA2, SRC1, TUB1, and USV1. This is due to the fact that many of the budding yeast introns reside within genes that encode ribosomal proteins 48. In addition, none of the introns in periodically expressed genes show signs of phase-specific splicing; hence, in contrast to meiosis in budding and fission yeast 4950, we see no evidence for a regulatory role of splicing in the mitotic cell cycle of budding yeast.

W101 (50 ml; MATa ade2-1 trp1-1 leu2-3, 112 his3-11, 15 ura3 can1-100 [psi1]) background temperature-sensitive cdc28-13 mutant S. cerevisiae strain K3445 (YNN553) was grown for approximately 8 to 10 hours in rich yeast-extract/peptone/dextrose (YPD) in a shaking water bath at 25°C and diluted in 3 × 1.6 liter cultures for overnight growth in an air incubator at 25°C. The following morning the cultures of OD600 approximately 0.2 were mixed together, distributed into 45 × 100 ml samples and arrested in late G1 at START by shifting the temperature from 25°C to 38°C. After 3.5 hours, the cells were transferred back to permissive temperature to re-initiate cell division and samples were collected every 5 minutes for 215 minutes (equal to more than two complete cell cycles). The cultures were centrifuged and snap-frozen in liquid nitrogen. The degree of synchrony was monitored by assessing the number of budding cells and measuring the bud size (Additional file 1). Nuclear position was determined by Hoechst staining with fluorescence microscopy (Additional file 16).

To arrest bar1 strain DBY8724 (MATa GAL2 ura3 bar1::URA3) 30 in G1 at START, alpha-factor pheromone was added to a final concentration of 600 ng/ml. After 2 hours of arrest, cells were released by washing and recovered in fresh preconditioned medium to facilitate initiation of mitosis. Samples were collected every 5 minutes for 200 minutes (equal to three cell cycles). The degree of synchrony was monitored by assessing the number of budding cells. Nuclear position was determined by Hoechst staining with fluorescence microscopy.

Total RNA was isolated from the culture corresponding to each time-point by the standard hot phenol method 11. Poly(A)-RNA was enriched from 1 mg of total RNA by a single passage through the Oligotex Oligo-dT Column (Qiagen, Hilden, Germany). Poly(A)-RNA was treated with RNase-free DNaseI (Ambion's Turbo DNA-free Kit, Foster City, CA, USA) for 25 minutes at 37°C according to the manufacturer's instructions and subsequently reverse transcribed to single-stranded cDNA for microarray hybridization. Each 200 μl reverse transcription reaction was carried out in duplicate and comprised 6 μg of poly(A)-RNA, 3 μg random hexamers (RH6), 1 μl of 6 mg/ml Actinomycin D (ActD), 0.4 mM dNTPs containing dUTP (dTTP:dUTP = 4:1), 40 μl 5× first strand synthesis buffer (Invitrogen, Karlsruhe, Germany), 20 μl 0.1 M dithiothreitol (Invitrogen), and 1,600 units of SuperScript II (Invitrogen). The synthesis was carried out at 42°C for 1 h and 10 minutes, followed by reverse transcriptase inactivation at 70°C for 10 minutes. Poly(A)-RNA and RNA in heteroduplex with cDNA were digested by a mixture of 3 μl of RNAseA/T cocktail (Ambion) and 3 μl of RNAseH (Invitrogen) for 15 minutes at 37°C followed by inactivation of the enzymes for 15 minutes at 70°C. Replicate cDNA samples were further applied to the Affy Clean-up column (Affymetrix, Santa Clara, CA, USA), eluted together in 30 μl DEPC-H₂O and quantified. Purified cDNA (3.3 μg of each 5-minute time-point sample) was fragmented and labeled with WT Terminal Labeling Kit (Affymetrix) according to the manufacturer's instructions and then hybridized to tiling arrays.

For DNA hybridization, both strains were grown in YPD media overnight to saturation in three biological replicates and whole-genomic DNA was purified using the Genomic DNA Kit (Qiagen). Genomic DNA (10 μg) was digested to 25 to 100 base fragments with 0.2 U of DNaseI (Invitrogen) in 1× One-Phor-All buffer (Pharmacia, Munich, Germany) containing 1.5 mM CoCl₂(Roche, Mannheim, Germany) for 3.5 minutes at 37°C. After DNaseI inactivation by boiling for 10 minutes, the sample was 3' end-labeled in the same buffer by the addition of 1.5 μl of Terminal Transferase (25 units/μl; Roche) and 1.5 μl 10 mM biotin-N6-ddATP (Molecular Probes, Karlsruhe, Germany) for 2 hours at 37°C, and hybridized to the tiling array.

The array was designed in collaboration with Affymetrix (PN 520055), as described in David et al. 11. Probe sequences were aligned to the genome sequence of S. cerevisiae strain S288c (Saccharomyces Genome Database of 7 August 2005). Perfect match probes were further analyzed.

The log-base 2 perfect match (PM) probe intensities from each array were background corrected and calibrated using the DNA reference normalization method described in Huber et al. 33, which was applied separately to both datasets, cdc28 and alpha-factor.

To determine the transcript boundaries in the combined dataset, a piece-wise constant model was fitted to the normalized intensities of the unique probes ordered by genomic coordinates. The basic model described in Huber et. al. 33 was modified to allow time-point-dependent levels. The normalized intensities (z_jk) were modeled as:

where μ_sk is the array-specific level of the s-th segment, ε_jk are the residuals, j = 1, 2,., n indexes the probes in ascending order along the chromosome, k indexes the time-point (array), t₂,., t_S parameterize the segment boundaries (t₁ = 1 and t_S+1 = n + 1) and S is the total number of segments. Model 1 was applied separately to each strand of each chromosome. For each chromosome, S was chosen such that the average segment length was 1,250 nucleotides. Change-points were estimated using a dynamic programming algorithm implemented in the tilingArray package 33.

After segmentation, the average of the probe signals within the segment boundaries was calculated for each time point. A table of segment levels is available from the supplementary materials webpage 32.

To estimate a threshold for expression, the average level over both datasets was calculated for each segment. Segments not overlapping annotated, transcribed features were used to estimate the background level as follows. A normal distribution was fit in order to determine a threshold at which the estimated false discovery rate was 0.1% 11. For the mean of the normal distribution, we used the midpoint of the shorth (the shortest interval that covers half of the values), for the variance, the empirical variance of the lowest 99.9% of the data. Segments whose level fell below this threshold were considered not expressed.

Segments were then assigned to different categories depending on how they overlapped with annotated features as described in David et al. 11, with the difference of re-naming the unannotated isolated features to the unannotated intergenic. Expression values for each annotated feature were calculated as weighted averages of the overlapping segments on the same strand.

We used a combination of three approaches to identify periodically expressed segments and annotated features based on the cdc28 and alpha-factor datasets: the method of Ahdesmaki et al. 34, which calculates P-values for a robust nonparametric version of Fisher's g-test 5455, the permutation-based method of de Lichtenberg et al. 35, which scores genes based on both the magnitude of regulation and the periodicity of profile, and by systematic visual inspection. For the two computational methods, score cutoffs were determined based on comparison with existing benchmark sets of 113 known cycling genes identified in single-gene studies 47. A combined list of cycling transcripts was compiled that contains all transcripts identified as cycling by at least two of the three methods. The peak time of expression for each transcript was calculated as percentage of the cell cycle duration as previously described 35. To determine the length of the cell cycle in each experiment, the period length was optimized to fit the expression profiles for selected genes from the benchmark set.

To test if the ncRNAs are likely to be novel protein-coding genes, we extracted all ORFs within unannotated antisense and intergenic transcripts and compared their length distributions to what would be expected by chance. The length of an ORF was defined as the distance between a stop codon and the most upstream ATG codon. Two separate background distributions were used for antisense and intergenic transcripts, to take into account that these two types of ncRNAs have different sequence properties (k-mer frequencies), because the former are located opposite of protein-coding genes whereas the latter are located within intergenic regions. For antisense transcripts, a set of sequences with the same length distribution was sampled from the genomic regions opposite other protein-coding genes. Opposite genomic regions with matched length distribution and sequence properties were used as a background for the unannotated intergenic RNAs. The ORF length distributions observed for the antisense and intergenic transcripts were not statistically significantly different from their respective background distributions according to the Kolmogorov-Smirnov test.

We used the TAMO suite 56 to identify the TFs that preferentially bind to regulatory regions of periodic non-coding transcripts. We systematically searched for binding motifs that were significantly overrepresented for the region, spanning from -600 bp upstream up to +600 bp downstream of 37 periodic unannotated antisense and 11 intergenic transcripts of interest, relative to a background set composed of all transcripts detected in the alpha-factor experiment. A benchmark set comprised 113 genes whose transcription was reported as cell cycle regulated in single-gene studies previously 47, whereas the lowest scoring 252 non-periodic antisense transcripts from the alpha-factor induced arrest dataset served as a negative control. We also performed de novo motif discovery on these sequences, using the combination of methods contained in the TAMO software suite. This analysis revealed no significantly overrepresented sequence motifs. We then searched for the putative TF binding sites that matched the position-specific score matrices from MacIsaac 5758.

We investigated the overlap between transcripts and genomic regions with conserved secondary structure 39. We used Steigele et al.'s 39 regions for cutoff 0.5. The regions were remapped to the current genome assembly using Exonerate (requiring 100% identity). The regions are strand-specific and overlap with these regions was also considered in a strand-specific way.

We generated deletion strains with the help of PCR-based technology as described on the Stanford Yeast Deletion webpage 59 using a set of up- and downstream primers flanking the defined periodic unannotated sequence listed in Additional file 5. The growth of deletion strains was monitored in liquid media using GENios automatic microplate readers (TECAN).