To identify origins in S. pombe, S. octosporus, and S. japonicus, we mapped all sites with increased DNA copy number in early S phase 11 24 . We synchronized cells in G2 by elutriation, allowed them to enter S phase in the presence of the ribonucleotide reductase inhibitor hydroxyurea (HU), harvested genomic DNA from the S-phase arrested cells, sheared the DNA to about 200-bp fragments, and used deep sequencing to count the frequency of each region of the genome (Figure 1). As a control for systematic biases in DNA preparation and sequencing, we also sequenced DNA from G2 cells, which have uniform DNA copy number.

We employed Helicos single-molecule sequencing technology because it economically provides large numbers of reads without amplification and with significantly lower GC-bias than alternative sequencing technologies 25 26 . For S. pombe, we collected three pairs of datasets (each pair corresponding to S and G2 cells): two from wild-type cells with different synchronization protocols and one from a different strain background (see below). For S. octosporus, we collected one pair of wild-type datasets. For S. japonicus, we collected two wild-type S-phase datasets and one wild-type G2 dataset. For each sample, we collected an average of 18 million mappable sequence reads, which represents >200-fold sampling of copy number at a 200-bp resolution (Table 1).

To produce replication profiles for each species, we normalized the read frequency in the S and G2 datasets, subtracted the G2 background from the S-phase signal, and smoothed the resulting curves (see Materials and methods). The replication peaks of S. pombe and S. japonicus are about 10 kb wide, consistent with previous estimates of fork progression in HU arrested cells (Figure 1) 27 . The replication peaks of S. octosporus are more than twice as wide, perhaps reflecting a more heterogeneous arrest in HU.

We found that many of the peaks in the S. pombe Sp1 and Sp2 datasets were notched, presumably due to loss of bubble-shaped replication intermediates during our sample preparation. To compensate for the notched peaks, we identified peaks in all of our replication profiles using a template fitting approach that predicts where the peaks would be if they were not notched (Figure 1; Figure S1 in Additional file 1). To estimate the sensitivity of our peak detection, we compared two independent S. pombe G2 samples from the datasets Sp1 and Sp2 (Figure 1b). The difference between the two datasets, which reflects the experimental noise in the data, has a distribution of 1.00 ± 0.145. We therefore restricted our analysis to peaks of heights greater than 1.30, or about 2 standard deviations above the baseline. To confirm the reproducibility of our peaks, we compared independent datasets from both S. pombe and S. japonicus and found a strong correlation between peak locations in the datasets (P < 10^-10; Figure S2 in Additional file 1).

To validate our replication peaks, we compared our data to confirmed origins. For S. pombe, we examined 37 origins that were previously shown to be active in vivo by two-dimensional gel analysis (collated in 15 ) and have mappable reads in our datasets. Of the 37 origins, 22 overlap with an identified replication peak in our data sets; 8 are less than 1 kb away and 4 are between 1 and 2.5 kb away; for 3 we see no replication peak in the region (Figure 1a; Table S2 in Additional file 2). Overall, the enrichment of origins near replication peaks is highly significant (P = 4.41 × 10^-09). We also compared our replication peaks to AT islands, intergenes of high AT content, which are strong predictors of origin function in S. pombe 17 . We find a significant (P < 10^-10), but not perfect, overlap of AT islands with our replication peaks (Figure 1a), consistent with our sequence analysis of origin motifs (see below). We also see comparable overlap between our replication peaks and published S. pombe origin maps (see Materials and methods).

For S. japonicus, we compared our data to 11 ARSs isolated from a genomic library 28 . Six of the ARSs encompass a replication peak and four others were within 2.5 kb of a replication peak (P = 0.012); for one ARS, we found no associated replication peak (Figure 1d; Table S3 in Additional file 2). These results suggest that our replication peaks correspond to in vivo origins of replication.

We compared the locations of replication peaks within syntenic regions among the three species and found no significant conservation of origin location (P = 0.22), consistent with the lack of conservation of origin locations among budding yeast species of similar evolutionary distance 29 .

Nonetheless, genome-scale patterns of origin firing are conserved between S. pombe and S. octosporus (Figure 2). The replication peaks on chromosomes 1 and 2 of these species are greatly enriched in the central, centromere-containing regions of the chromosomes and much diminished on the arms; in contrast, replication peaks are uniformly distributed across chromosome 3. Early firing of S. pombe centromeric origins has been reported 30 . This early firing is due to an interaction between the Dfp1-dependent kinase (DDK; the S. pombe homolog of the Cdc7-Dbf4 replication kinase) with the Swi6 heterochromatin-binding protein (the S. pombe homolog of the HP1 heterochromatin-binding protein) 31 . The Swi6-binding motif in Dfp1 is conserved in S. octosporus, which also shows a bias towards chromosome centers, but not in S. japonicus, which does not (Figure 2b). To test if the Dpf1-Swi6 interaction is required for the centromere bias, we mapped origins in an S. pombe strain carrying a mutant Dfp1 that does not interact with Swi6 31 . The mutant does not affect the centromere bias (Figure 2a), consistent with the observation that the bias extends hundreds of kilobases from the centromeres, much farther than heterochromatin does, and encompasses all of chromosome 3.

The central regions of chromosomes 1 and 2 have been previously reported to have more efficient origins than their arms 11 13 15 , but the effect is not as strong as the bias seen in our datasets. We investigated experimental factors that might account for this difference in the distribution of origin firing. Previous experiments had synchronized cells in G2 by inactivating a temperature-sensitive (ts) allele of the Cdc25 mitotic activator and then releasing the blocked culture into a synchronous cell cycle. This protocol arrests cells for up to 4 hours in G2 and causes them to grow to as much as twice their normal size before dividing. In contrast, we had used centrifugal elutriation to select a synchronous population of G2 cells from an asynchronous culture, a protocol that does not involve arresting the cell cycle. We repeated our origin mapping using a cdc25-ts block-and-release approach, and confirmed that origin firing is much more uniformly distributed across the genome (Figure 2). Nonetheless, the same origins are used in the two conditions (Figure S2 in Additional file 1). This result suggests that a prolonged G2 arrest causes origin firing potential to be more uniformly distributed.

Although origin locations have been identified in a wide variety of eukaryotes, the sequence characteristics that confer origin activity to these loci have only been elucidated in S. cerevisiae and S. pombe 1 . We used two computational approaches to analyze the sequence characteristics of the origins we identified in the three Schizosaccharomyces species. The first approach used a support vector machine (SVM), a machine-learning algorithm that tests the ability of k-mer sequences (k = 1 to 6) to predict origin function (see Materials and methods) 32 . The results of such tests can be analyzed with a receiver operating characteristic (ROC) curve, which plots true-positive rate (sensitivity) as a function of false-positive rate (1 minus specificity). The area under the ROC curve (auROC) demonstrates the performance of the SVM model, with 1.0 being perfect and 0.5 being no better than random. auROC thus serves as a metric of the predictive power of the model.

We performed SVM analysis on all data sets (Figure 3 and Table 1; Tables S4 and S5 in Additional file 2). For all three species, we are able to predict origins with high sensitivity and specificity, as demonstrated by auROC scores ranging from 0.76 to 0.89 (Table 1). To determine which sequences are the most predictive of origin function, we examined the auROC scores for individual k-mers. As expected for S. pombe, the AT richness is a strong predictor of origin function, as demonstrated by the high predictive value of the A nucleotide (auROC = 0.62 to 0.79; Table S4 in Additional file 2). However, simple AT richness is not as strong a predictor as polyA tracts (auROC up to 0.86), consistent with the fact that the AT-hook motif, of which S. pombe Orc4 has nine (Figure S4 in Additional file 1), binds polyA. Interestingly, the sequence character of the origins used in the cdc25-ts-synchronized culture is less biased towards strictly polyA and is less predictive of origin function than that of the unarrested, wild-type culture (Table 1; Table S4 in Additional file 2). This change in the sequence character of the origins used in the cdc25-ts-synchronized cells is consistent with a wider range of less well-defined origins being used in these experimental conditions. S. octosporus origins have similar sequence character, with polyA tracts dominating the SVM signal; however, the auROC scores are significantly lower than those for S. pombe, possible due to the lower resolution of the data set (Figures 2 and Figure 4a, Table 1; Table S4 in Additional file 2).

Surprisingly, AT content is not a positive predictor of origin function in S. japonicus. In fact, it is a strong negative predictor of origin function (auROC = 0.15). Instead, S. japonicus origins are characterized by high GC content and a variety of GC-rich k-mers. Interestingly, A is a much stronger negative predictor of origin function than 6A, which is a weakly positive character (auROC = 0.54). This result suggests that although S. japonicus origins are depleted in AT content, they maintain a higher relative polyA content, consistent with the existence of AT-hooks on S. japonicus Orc2 and Orc4 (Figure S4 in Additional file 1).

As expected, SMVs trained on one data set from a species performed well on other datasets from that species (Table S5 in Additional file 2). Conversely, the SMVs trained on S. pombe perform poorly on S. japonicus, which has very different origin characteristics, and vice versa. Interestingly, the S. pombe SVMs also performed poorly on S. octosporus, and vice versa. Therefore, it appears that even though S. pombe and S. octosporus have qualitatively similar origin characteristics, there are quantitative differences in their origin sequences.

Our second approach to characterizing origin sequence used the ab initio motif discover algorithm MEME, which is capable of identifying the enrichment of motifs that are longer than six residues. As for the SVM analysis, we used 1 kb from either side of the replication peaks as the positive loci. However, because fission yeast origins are generally intergenic 16 17 , our positive loci are enriched for promoter sequences. To avoid training on promoter sequences and thus inadvertently identifying promoter motifs, we masked coding sequence from both our positive and negative loci. We performed the analysis using two background models. The first approach uses only the mono-nucleotide composition of the input loci as background, and finds motifs that are enriched relative to that bias. This approach finds motifs that are unexpectedly frequent in the input loci, but does not compare the frequency between positive and negative loci. The second approach uses a fifth-order hidden Markov model of the background in the negative loci, which reduces the background noise and allows for more rigorous identification of enriched motifs. We initially limited our search to motifs of less than 17 bases, to increase sensitivity. No additional motifs were found upon extending the search space to 50 bases.

In both S. pombe and S. octosporus, the motifs identified by MEME using both background models reflected the polyA-rich nature of origins seen in the SVM analysis (Figure 3; Table S6 in Additional file 2). In addition, a polyAG motif was frequently found in S. pombe origins (Table S6 in Additional file 2).

In S. japonicus, using the mono-nucleotide composition background, we find polyA motifs, suggesting that origins have significant enrichment of polyA tracts above that expected from their AT content, consistent with the SVM results (Figure 3; Table S4 in Additional file 2). However, using the hidden Markov model background model, we find no polyA motifs, suggesting that although intergenes in general tend to have high numbers of polyA tracts, those tracts do not directly determine origin function. Instead, we find three highly enriched motifs (Figure 3): polyG, a head to tail repeat of CTCGCT and the binding site for the Sap1 protein (Figure S3 in Additional file 1), with 87% of all replication peaks containing at least one of the motifs. polyG motifs are known to exclude nucleosomes - like polyA, due to the intrinsic stiffness of the sequence 19 20 33 - and Sap1 has been shown to exclude nucleosomes in S. pombe 33 , suggesting that these motifs may function to deplete origins of nucleosomes in S. japonicus.

Previous work has demonstrated that origins tend to be nucleosome depleted 3 4 9 21 34 . To investigate the nucleosome occupancy of origins in the three fission yeast, we mapped mono-nucleosomes by deep sequencing of all three species. Average nucleosome occupancy profiles centered on replication peaks from S. pombe and S. japonicus show them to be nucleosome depleted, as expected (Figure 4a). The replication peaks for S. octosporus show no significant nucleosome depletion (data not shown), possibly due to the lower resolution of that replication data set. To compensate for the fact that our replication peaks are not perfectly aligned with their underlying origins, we identified the nearest NFR to each replication peak and to an equal number of control sites (Figure 4b). Average nucleosome occupancy profiles centered on origin-proximal NFRs tend to be deeper and broader than nucleosome profiles centered on origin-distal NFRs (P < 5 × 10^-3). Note that both the origin NFRs and non-origin NFRs in S. pombe tend to be more nucleosome depleted than in S. japonicus. In S. octosporus, the non-origin NFRs are similar to those in S. pombe, but the S. octosporus origin NFRs appear to be less depleted than the origin NFRs in S. pombe, possibly because the lower resolution of the data set caused some of the replication peaks to be assigned to the wrong NFR.

Several of the identified sequence motifs associated with origins, in particular polyA and Sap1, are known to exclude histone binding 20 33 . To investigate the roles of other origin-associated motifs, we examined their in vivo nucleosome occupancy. The motifs tend to be nucleosome depleted in all three species, with three exceptions (Figure 4c). The first exception is the CTCGCT motif in S. pombe and S. octosporus, which is largely absent from these genomes. The second exception is polyG, which is not nucleosome depleted in S. pombe and S. octosporus. Although it seems counter-intuitive that a sequence believed to exclude nucleosomes because of its intrinsic stiffness should show species-specific effects, such effects have been seen before, due to the fact that polyG motifs in relatively G-rich genomes (like S. japonicus) tend to be embedded in longer G-rich sequences 33 . The third exception is the lack of nucleosome exclusion for polyAG in S. octosporus. This result is consistent with the fact the polyAG is not enriched at S. octosporus origins and may reflect a different dinucleotide bias between the S. pombe and S. octosporus polyAG sequences.

To further investigate the roles of the identified origin motifs in nucleosome exclusion, we compared the nucleosome occupancy of motifs at origins and non-origin loci. The motifs all tend to be as nucleosome depleted at non-origin loci as they are at origins (Figure S5 in Additional file 1). This result suggests that the motifs exclude nucleosomes, as opposed to being nucleosome depleted because of their enrichment in origins. Therefore, the sequences that characterize origins in fission yeast appear to fulfill two roles: excluding nucleosomes and binding ORC. This conclusion is consistent with a general model for eukaryotic origins in which origins are simply the NFRs with the highest affinity ORC binding sites.

All three species were grown in rich medium (YES) and manipulated using standard fission yeast protocols 45 . The strains used were yFS101 (S. pombe h-), HM1826 (S. pombe h+ nmt1-TK dfp1-3A::kanMX6) 31 , yFS128 (S. pombe h- leu1-32 ura4-D18 cdc25-22), yFS286 (S. octosporus), and yFS275 (S. japonicus).

DNA sequencing and sequence data analysis were performed as previously described 26 . We used the January 2007 assembly of the S. pombe genome 46 and SO3 and SJ1 assemblies of the S. octosporus and S. japonicus genomes 35 . Reads that mapped to multiple locations of the genome were discarded. Post analysis, S. japonicus coordinates were mapped to genome assembly SJ4, the current release.

To construct replication profiles, we extended the reads by 100 bp in the 3' direction, because the average length of DNA fragments was 200 bp. Then we smoothed the data using a 200-bp sliding window with a 20-bp step. We excluded regions of S. octosporus (contig6:1-400000, contig2:1-5600, contig7:766887-886887) and S. japonicus (contig1:1-62000, contig3:1-74000, contig3:1776000-1809320, contig4:1-16000) that had anomalously high read counts, presumably due to under-representation of repetitive sequences in the genome assemblies. To account for different sequencing depths in the various datasets, we normalized each dataset so that the mean across the genome was 1. Using flow cytometry analysis of percentage of the genome replicated in the S phase samples (Table S1 in Additional file 2), we applied the following formula to combine the S and G2 datasets into replication profiles: R = S × (1 + n%) - G2 + 1. Finally, we performed LOESS smoothing (Igor PRO v6.12, WaveMetrics (Lake Oswego, OR USA) on the replication profile R using empirically determined smoothing factors (Table S1 in Additional file 2).

We used the MATLAB ipeak program (MathWorks (Natick, MA USA) to detect peaks in each replication profile R. The ipeak algorithm first looks for downward zero crossings in smoothed first derivative, and then detects peaks by least square curve fitting. The ipeak parameters used were AmpT = 1, SlopeT = 0, SmoothW = 58, FitW = 58 for S. pombe, AmpT = 1, SlopeT = 0, SmoothW = 50, FitW = 50 for S. octosporus, and AmpT = 1, SlopeT = 0, SmoothW = 10, FitW = 10 for S. japonicus. We set the peaks to be the highest positions in R within a ± 500-bp window around the peak locations detected by ipeak. Peaks with anomalous heights (greater than 3 in S. pombe and S. octosporus or 10 in S. japonicus) were discarded; these regions are presumed to correspond to unannotated repetitive regions. We used the Genomic Signal Aggregator 47 to generate an average peak profile of all peaks detected by ipeak in the entire genome. The center of the average peak profile was then fitted to a Gaussian distribution, which was then used serve as a peak template for refining peak calling.

In order to remove split peaks and sub-peaks, we used the following strategy. First we sorted all peaks detected by the ipeak algorithm by their amplitudes. Then we went down the list and performed the following steps: 1) we scaled the peak template to the same height as the data peak under investigation and positioned the template such that its highest point coincides with the highest point of the data peak; 2) we slid the peak template in a ± 250-bp window to search for the position that would lead to maximal correlation between the peak template and the data peak; 3) we searched in the height dimension with a step of 0.01 for minimal root mean square deviation (RMSD) between the peak template and the data peak. To avoid noise in the data, the correlation and RMSD were computed only for the region that was within 60% of the highest point of the data peak. We assigned the peak location and height to those of the resulting template, saved the information, then subtracted the newly detected peak from the replication profile R so that minor side peaks would be ignored. We iterated this procedure until all peaks above the peak threshold had been detected.

We compared our data to published S. pombe origin lists compiled by three different approaches: MCM ChIP-chip mapping, a microarray-based, HU-arrest, copy-number strategy similar to ours, and a microarray-based mapping of single-stranded DNA in HU arrested cells 11 12 13 . We calculated the statistical significance of the overlap between these origin lists (the test sets) and two well-defined origin lists (the validation sets): a list of origins confirmed by two-dimensional gels (collated in 15 ) and a list of AT-rich intergenes that are strongly correlated with origins 17 . We varied the resolution for the test sets, from 500 to 5,000 bp, by varying the distance from a validation set origin that a test set origin could be and still be considered to overlap it. We reasoned that the statistical significance of the overlap between the test and validation sets would increase as the size of the test sets increased up to the resolution of the test set. After that point the significance would decrease because the test set would get bigger without improving the overlap. By this analysis, our data are of similar resolution, between about 1 and 3 kb, to the other HU-arrest approaches (see Materials and methods; Table S7 in Additional file 2), which presumably reflects the resolution of the HU-arrested replication bubbles. The overlap between our replication peaks is comparable to previously published origin prediction, with approximately 50% overlap between any pairwise comparisons of datasets (Figure S2 in Additional file 1).

We used MEME v4.4 to find sequence motifs de novo in the replication origin peaks with parameters nmotifs = 30, minsites= 20, minw = 11, maxw = 17 48 . MEME is an expectation-maximization algorithm that identifies motifs that are enriched in a positive set of sequences, in our case the sequences from 1 kb on either side of the replication peaks, compared with a background set of sequences, which were a set of sequences from 1 kb on either side of loci chosen randomly from the genome excluding 8 kb around the peaks taller than 1.1, 1.14, 1.18, 1.2, 1.2, and 1.25 for Sp1, Sp2, Sp3, So1, Sj1, and Sj2, respectively. We ran the algorithm with two background models: fifth-order Markov model (defined as frequencies of all 6-mers) or simply mononucleotide frequencies. If two neighboring peaks were closer than 2 kb, the overlapping sequence was counted only once. We only used the intergenic portions for both positive and background sequences.

In order to predict replication origins from sequence features, we constructed SVMs with a linear kernel using the libSVM algorithm in the R package (e1071). We used k-mer (k = 1 to 6; 2,772 k-mers in total) and significant motifs identified by MEME as sequence features. The frequency of each k-mer was calculated for each sequence in the peak set and the background set as described in the Motif Finding section, and was 2-norm normalized. For each MEME motif, we ran the CLOVER algorithm to obtain a cumulative score for each sequence 49 . Then we normalized the CLOVER scores for each motif to have the same mean and standard deviation as the k-mer frequencies across the training sequences, so that k-mers and motifs could be combined with the SVM.

We used the positive and background sequences described in the Motif Finding section for training and testing the SVMs. We performed five-fold cross-validation, in which the sequences were divided into three equal sized portions and in each of five runs one portion is held out for testing while the other two portions were used for training. We computed the auROC score, which is a performance measure that combines sensitivity and specificity, using libSVM and colAUC (R package caTools).

Nucleosomal DNA libraries were constructed and sequenced as previously described 50 . To assess the quality of the nucleosome maps, we aligned our nucleosome signal on the annotated transcriptional start site for each genome 35 using the Genomic Signal Aggregator 47 (Figure S6 in Additional file 1). The profiles show a deep NFR at the transcriptional start site, with phased nucleosomes downstream and heterogeneous nucleosomes upstream, as previously reported for S. pombe 21 33 . Nucleosome mapping and motif occupancy analysis were performed as previously described, using default settings 51 . We identified NFRs genome-wide as described 50 but without the constraint of searching relative to the translational start site of each gene. We defined the NFRs nearest each species' replication peaks as the origin-proximal NFRs and computed their average nucleosome occupancy, aligning the data on the NFR centers. As controls, we chose an equal number of randomly distributed genomic coordinates and defined the nearest NFRs as non-origin-proximal and then measured the non-origin-proximal average nucleosome occupancy. We repeated this analysis for 10,000 randomly chosen genomic regions for each species. To test the significance of the hypothesis that origin-proximal NFRs are more nucleosome depleted than non-origin-proximal NFRs, we calculated a P-value of n/10,000, where n is the number of times that the non-origin-proximal NFR had an average nucleosome occupancy that was more depleted than the origin-proximal NFRs.

Filtered reads for all datasets have been submitted to the Sequence Read Archive (Table S8 in Additional file 2). S. pombe replication profiles have been submitted to OriDB 52 and ReplicationDomain 53 .