DNA combing analysis of replication parameters in HeLa cells
We used DNA combing 
to measure replicon size and replication fork progression rate in HeLa cells at different stages of S phase (). Asynchronously growing cells were pulsed with the halogenated nucleotide IdU for 20 min followed by CldU for another 20 min, fixed and sorted into four temporal compartments of S phase (S1, S2, S3 and S4) according to total DNA content. DNA was stretched on coverslips by combing and total DNA was stained in red with an anti-DNA antibody. The replicative labels were revealed in blue (IdU) and green (CldU) using appropriate antibodies. The blue-to-green transitions show the position and orientation of mobile forks at the time CldU was added ().
DNA combing analysis of DNA replication in HeLa cells.
Replication fork velocities were determined by measuring the length of CldU or IdU tracts that could be unambiguously assigned to the progression of a single fork during an entire 20 min labeling interval. Fork velocities were narrowly distributed around a mean of 0.68 kb/min, with almost no values >2 kb/min, and did not change throughout S phase ().
The global density of replication forks (total number of forks divided by total length of DNA examined, corrected for contamination by non-replicating G1 or G2/M cells and for replicated genome fraction; see Material and Methods
) increased through S phase (from 2.64 to 3.88, 4.55 and 5.4 forks per Mb in S1, S2, S3 and S4, respectively; ). The measured replication fork densities and velocities were used to calculate the time required to duplicate the entire genome (see Material and Methods
). The result (6 h 18 min) was consistent with the length of S phase independently measured from the cell doubling time and the fraction of the sorted cells in S phase (22 h×1/3
7 h 20 min), corroborating the fork density and velocity measurements.
Statistical analysis of the parameters of DNA replication determined by DNA combing.
The inverse of the global fork density is the global fork-to-fork distance (FTFD). The global FTFD decreased from 379 kb to 258, 220 and 185 kb in S1, S2, S3 and S4, respectively. However, the local FTFDs measured on single DNA fibers containing forks were much smaller (mean ~19 kb) and did not decrease so much during S phase (from 22.0 kb in S1 to 17.0 kb in S4; ; ). Furthermore, the mean intra-fiber inter-origin distances and inter-termini distances were commensurate with the intra-fiber FTFDs, i.e. both were in the 35–42 kb range at all stages of S phase (; ). Thus, replicons were much shorter than global FTFDs would suggest. The discrepancy between local and global FTFDs might be attributed to the finite fiber size, which prevents measurement of large FTFDs, but actually results from the fact that origins are activated as clusters that fire at different times in S phase. Thus, only 10–20% of all fibers showed replication forks at any stage of S phase but among these many showed several forks (). To assess the clustering of replication forks, we compared the distribution of the number of forks per fiber with that generated in a simulation that assumed random initiation and a fiber size distribution and global fork density identical to the experimental samples. The observed distributions were significantly (P<10−4) different from the simulation, with a lack of fibers with one fork (whole S-phase average, 6.3% vs. 10.6%) and an excess of fibers with ≥2 forks (8.5% vs. 2.2%). This demonstrates a clustering of origin firing.
We next examined whether the global fork density increased because more origin clusters fired or because more origins per cluster fired during S phase. We found that the number of forks per fork-containing fiber (2.33, 2.54, 2.94 and 3.31 forks per fiber in S1, S2, S3, and S4, respectively; ; ), and the IdU-labeled fraction of fork-containing fibers (30.1% , 44.3%, 41.9% and 52.0%; ; ) increased throughout S phase. Thus, more origins per cluster fired as S phase progressed. The distances between origin clusters are generally too large to be measured, because they exceed the mean fiber size. Although such distances cannot be individually measured, their mean can be computed from the statistics of fibers with and without forks (by dividing the total length of DNA minus the sum of intra-fiber FTFDs by the number of fork-containing fibers, assuming at most one cluster per fiber). Note that intercluster segments mainly consist in unreplicated DNA in early S phase and already replicated DNA in late S phase and that the total DNA length used in our calculations is corrected for the extent of DNA replication (see Material and Methods
). The mean intercluster distance decreased from 772 kb in S1 to 484, 465 and 501 kb in S2, S3 and S4. Thus, inter-cluster distances were reduced as S phase progressed from S1 to S2 but did not change thereafter. This reduction was too large to be explained by the increase in cluster size. Therefore, the number of active clusters increased from S1 to S2.
To further evaluate the tightness of origin synchrony we reasoned that the consecutive IdU/CldU labeling scheme allows us to distinguish origins that fired before (type 1) or after (type 2) CldU addition. Type 1 origins are flanked by two divergent blue-to-green transitions whereas type 2 origins give rise to doubly-labeled, isolated tracks. For example, most origins shown on fired before CldU addition but the leftmost one fired after CldU addition. We first noticed that when inter-origin distances were plotted separately for type 1 and type 2 origins (not shown), their distributions were not markedly different from those shown on , where all origins were taken into account. This suggested that type 1 and type 2 origins were not frequently interspersed with each other. We then selected fibers containing more than one origin and found that adjacent origins were significantly more frequently of the same type than if randomly interspersed (254 type1/type1; 60 type2/type2; 99 type1/type2; P<10−4, chi-square test of homogeneity). Thus, adjacent origins tended to fire within 20 min of each other. Together these observations suggested that a wave of initiations propagates on the DNA molecule.
In conclusion, DNA combing showed that in HeLa cells i) replication origins are spaced at mean ~40 kb intervals; ii) adjacent origins fire within 20 min of each other, resulting in a spatial clustering of replication forks; iii) replication fork velocity (~0.68 kb/min) does not change during S phase; iv) the global fork density increases during S phase, because more replicon clusters and more origins within clusters become active as S phase progresses. Therefore, the global rate of DNA replication increases during S phase due to increasing origin synchrony.
Comparison with replication parameters found in other studies
Our conclusion that fork speed is constant through S phase contrasts with earlier reports of changes in fork speed during S phase 
. However, in these studies, chemicals or serum starvation were used to synchronize cells, which may affect nucleotide pools and replication fork progression, whereas the retroactive (FACS) synchronization we used does not perturb the cell cycle. Furthermore, these studies used less precise techniques than DNA combing to spread DNA fibers, and some of the track length changes interpreted as changes in fork progression may in fact have resulted from changes in the synchrony of adjacent origins and consequent merging of forks. We have minimized such potential artifacts thanks to the use of two short, consecutive labeling pulses and the better resolution of DNA combing, which allowed us to demonstrate an increase in adjacent origin synchrony during S phase.
The fork speed (~0.7 kb/min) and interorigin distance (~40 kb) we found are somewhat, though not much, lower than usually reported in other human cell lines (typically 1.0–2.0 kb/min and 100–200 kb) 
. Small interorigin distances (57 kb) and slow forks (0.37 kb/min) have also been found by DNA combing in K562 leukemic cells 
. Small interorigin distances were also reported using another DNA fiber technique both in U2OS osteocarcinoma cells (50 kb) and in nontransformed MRC5 cells (42.5 kb) 
. More intriguingly, our estimates also differ from those reported by other investigators in HeLa cells (fork rates of 0.59–1.37 kb/min 
and 1.7±0.3 kb/min 
and interorigin distances of 144±66 kb 
). In yet another HeLa clone (data not shown) we observed slightly larger replicons (50 kb) and faster forks (1.0 kb /min) than in this work. Thus, clonal variation as well as differences in labeling scheme, DNA fiber technique and track choice probably explain these differences.
Such clonal variation is consistent with the possibility that the cancerous nature and genetic or epigenetic instability of HeLa cells influence origin activity and fork progression and their response to a number of physiological and pathological stimuli 
. Indeed, recent work showed that forced expression of oncogenes in primary keratinocytes can slow down replication fork progression and trigger activation of dormant origins due to decreased nucleotide pools 
. However, in another study, no change in origin spacing and fork velocity could be observed between primary keratinocytes and a keratinocyte-derived tumour cell line 
. Thus, it remains possible that the fork speed and origin spacing observed in HeLa cells just reflect some physiological tissue variation range.
Whole-genome replication timing profile of HeLa cells
We generated a high-resolution, genome-wide replication timing profile in HeLa cells as described previously 
with minor modifications detailed in the Material and Methods
. Briefly, HeLa cells were pulsed with BrdU, sorted into four temporal compartments of the S phase and nascent DNA was immunoprecipitated with anti-BrdU antibodies and sequenced using the Illumina technology to yield a total of 50 million reads that mapped uniquely to the human genome sequence. The abundance of sequence reads along the genome was computed every 10 kb in a 100 kb sliding window in each S phase compartment allowing to cover 90% of the genome. The resulting profile was used to compute in each window the fraction of the S phase at which 50% of the DNA was replicated (S50
). Using the FACS DNA fluorescence histogram to extract the proportion of cells at different stages in S phase and the temporal profile of the rate of DNA synthesis, we calculated the profile of DNA content as a function of the time spent by a cell in S phase. The S50
values were then used to deduce the time (TR50
) at which a defined genome region had replicated in 50% of the cells (see Material and Methods
). A biological replicate showed excellent reproducibility of the TR50
). The average of the two TR50
determinations was used for subsequent analyses.
The genome-wide TR50
histogram () showed a continuum of replication times with no dearth of replicating regions in mid-S phase and an increasing number of replicating regions during S phase. This is consistent with the increase in global fork density observed by DNA combing () and the one of global rate DNA replication observed by flow cytometry analysis (). This is also consistent with the observed dip in the FACS histogram of DNA content from S1 to S3, due to cells moving faster through this DNA content (Figure S1A
). The expected dip in S4 was not observed but this was due to the spreading of the adjacent G2 peak. The continuous dip from S1 to S4 was indeed visible in the post sort control, where the DNA content of sorted S1 to S4 cells was reexamined in a second round of sorting (Figure S1B
The global rate of DNA replication increases during S phase in HeLa cells.
Multiscale analysis of apparent replication speeds in HeLa cells
profile along the genome showed a landscape of peaks and valleys interspersed with flat domains of uniform replication time ( shows an exemplary 15 Mb chromosomal segment; see Figure S2
for a whole-genome profile). The slope of replication timing profiles has often been taken as a measure of replication fork velocity. However, since replication timing profiles are population averages, this is only true for regions in which forks progress in the same direction in all cells. Here we demonstrate (see Material and Methods
), as first proposed by de Moura et al 
in a recent analysis of yeast replication timing profiles, that the derivative of the replication timing, dt/dx
, depends not only on the fork speed, v
, but also on the local proportion of rightward (R
) and leftward (L
) moving forks in the cell population, such that dt/dx
. The apparent replication speed is defined here as the inverse of this derivative, dx/dt
. Note that the equality dx/dt=v/(R−L)
implies that the sign of the apparent replication speed indicates the predominant direction of replication progression and that in flat domains of uniform replication time (infinite apparent replication speed), forks move equally in both directions.
Replication timing profiles segmented in CTRs/TTRs and multiscale analysis of apparent replication speeds in HeLa cells.
We performed a multiscale analysis of the apparent replication speed genome wide, using the continuous wavelet transform, a robust method to obtain a well defined and numerically stable measurement of the local slope of the timing profile at any scale of observation (; Figure S2
). The replication speed modulus, |dx/dt
|, critically depended on the measured segment scale dx
. At very large scales (>2 Mb), the entirety of the genome appeared to replicate at >10 kb/min. At smaller scales, a differentiation of the genome into smaller and slower replicating segments was observed, revealing finer details of the replication profile. The landscape of replication speeds stabilized below the 100 kb scale, as expected from the spatial resolution of the profile. At this scale, a broad distribution of replication speeds was observed in the HeLa cell genome (), with 1% of 100 kb segments replicating at an apparent speed ≤2 kb/min, 53% in the 2–10 kb/min range, and 46% at >10 kb/min. We noted that the speed distribution was shifted toward higher speeds for S1 and S4 compared to S2 and S3 fractions ().
The observed range of apparent replication speeds cannot be explained by the range of single fork velocities measured by DNA combing in the same cells. The mean and maximum fork velocities are 0.68 kb/min and 2.0 kb/min, whereas 99% of the genome replicates at an apparent speed >2 kb/min. The possibility that regions with the slowest apparent replication speed are specifically replicated by the fastest forks seems unlikely since fork velocities at single loci usually show the same degree of heterogeneity as in the bulk genome (e.g. 
, and see below our data on the IGH locus). These results imply that in HeLa cells, |R−L
|<1, i.e. replication forks move in both directions, in most of the genome and that the proportion of right and left forks varies widely along the genome. There is a complete gradation between regions where forks progress predominantly (if not exclusively) in one direction (steep timing gradient, apparent speed ≤vmax
), and regions where they progress equally in both directions (flat timing gradient, high apparent speed). The apparent speeds must therefore reflect the statistics of origin activation around and within the timing gradients. Essentially similar results were obtained for several other cell lines (see below).
Segmentation of the genome into CTRs and TTRs
To address the mechanism by which different proportions of rightward and leftward moving forks are established in different parts of the genome in HeLa cells, we segmented the whole genome into constant timing regions (CTRs) replicating at >10 kb/min and timing transition regions (TTRs) replicating at ≤10 kb/min and analyzed them separately. shows the size distribution, genome coverage, TR50
and apparent replication speed of CTRs and TTRs defined at 100 kb (blue), 200 kb (green) and 500 kb (red) scales. At the 100 kb scale, the whole genome was segmented into 7548 CTRs and 7504 TTRs (; Figure S2
). All CTRs were ≤2 Mb and 53.8%<100 kb (), with CTRs>100 kb covering 34.2% of the genome (). All TTRs were ≤900 kb and 64.4%<200 kb (), with TTRs>200 kb covering 32.4% of the genome (). At larger scales, as expected, the mean size of both CTRs and TTRs increased and the genome fraction covered by CTRs increased at the expense of TTRs. The TR50
distribution of CTRs was relatively insensitive to scale () and was similar to that of the whole genome, but the apparent replication speed of TTRs increased with scale (). The small oscillations in the TR50
distribution of CTRs are an artifact of the finite number of S phase fractions, which we have not attempted to correct. The proportion of CTRs was higher in S1 (48%) and S4 (56%) than in S2 (28%) and S3 (32%), consistent with the fastest distribution of speeds in these two S phase compartments ().
Characteristics of CTRs and TTRs.
Replication mode of Constant Timing Regions
One possible mechanism for explaining why an equal proportion of rightward and leftward moving forks replicate a CTR is that it does not contain origins and is passively replicated from an outside origin that is activated equally often on its right or its left side (Figure S3A
). Given a mean fork velocity of 0.68 kb/min (40 kb/h) this mechanism could only apply to short enough CTRs (<300 kb) to replicate within a 7–8 h S phase in HeLa cells. At the 100 kb scale, CTRs<300 kb and >300 kb cover 19.7% and 21.2% of the genome, respectively (). This mechanism predicts that i) the edges of small CTRs would replicate asynchronously in non-adjacent S-phase compartments whereas their centers would replicate synchronously in mid-S phase; ii) that small CTRs lying between 150 and 300 kb would replicate rather in mid-S phase. A previous study of Hela cells replication timing determined that about 20% of the ENCODE regions present a pan-S replication profile 
. However, we reported that in HeLa cells only 7.4% of all genomic sequences replicate with such a pan-S profile 
. Although a significant correlation was observed between these two studies (Pearson, R
), the differences may result from the use of microarray hybridization and cell synchronisation by drug treatment in the first study vs. massive sequencing and no drug treatment in our study. Furthermore, we have found that the TR50
distribution of CTRs spans the entire S phase whatever their size (data not shown), inconsistent with the mechanism proposed above.
Alternatively, CTRs might consist of regions in which multiple origins are synchronously activated (Figure S3B
). This mechanism would result in an equal number of forks moving in both directions whatever the size and the replication time of the CTR. The fact that the TR50
distribution of CTRs spans the entire S phase whatever their size suggests that all long CTRs and most small CTRs replicate during defined intervals of S phase by synchronous firing of multiple replication origins. The small-scale changes in fork polarity around individual origins are not seen due to the small replicon size and/or to the use of different potential origins in different cells, which effectively smooth replication timing gradients across multiple replicons.
Replication mode of Timing Transition Regions
Our demonstration that the apparent replication speed is equal to v/(R−L)
(assuming that v
is locally constant), implies that in TTRs replication forks move predominantly but perhaps not exclusively in one direction. To further investigate this we analyzed TTRs individually. We found that the temporal transitions were directly proportional to the length of the TTRs (). Even at the smallest scale analyzed (100 kb), only 24 out of these 7504 transitions were compatible with the progression of a single fork even at maximum rate (vmax
2 kb/min) and together they only covered 0.13% of the genome. None of them was >250 kb, as expected from the maximum distance that a single fork can travel during S phase. Therefore, systematic unidirectional replication of large regions is not observed in HeLa cells. Replication forks instead appear to move in both a major and a minor direction in most TTRs. One potential explanation is that some TTRs support no internal initiation and are replicated from alternative origins located on either side of the TTR and used in unequal fractions of the cells (Figure S3C
). As discussed for CTRs, this mechanism could only apply to TTRs<300 kb and would predict asynchronous replication of their edges, for which we did not find convincing evidence.
Alternatively, multiple origins could fire in a progressive manner along the TTRs (Figure S3D
). The mean replication progression rate along TTRs was 3.63 kb/min, 5 times the mean progression rate of single forks (). This suggests that on average 2–3 adjacent replicons simultaneously operated along the gradient or, in other words, that on average adjacent origins spaced at ~36 kb intervals were consecutively activated at ~10 min intervals. Faster (slower) apparent speeds may result from shorter (larger) space and/or time intervals between adjacent initiations. This mechanism not only explains why replication progresses faster than single forks in TTRs but also why a higher proportion of forks move downstream than upstream the gradient, because when a new origin fires, the upstream moving fork will rapidly merge with the converging fork emanating from the upstream origin, whereas the downstream moving fork will progress for some distance before the next downstream origin fires. According to this mechanism, the faster distribution of speeds in late S phase is due to an increased synchrony of origin firings, consistent with the DNA combing results.
Replication accelerates along the TTRs
A visual inspection of the replication timing profile suggested that the slope of a large fraction of the TTRs tended to flatten with distance from their early edge. To asses this point, we selected TTRs>400 kb and measured the apparent replication speed at different positions along the slope. It was found that the apparent replication speed increased for about two thirds of the TTRs (). Furthermore, the distribution of apparent replication speeds along the TTRs was shifted to higher values at increasing distances from the early edge of the TTR (). These results suggest that forks move more and more in both directions along the TTRs as S phase progresses. These results are consistent with the DNA combing data showing that origins fire in an increasingly synchronous manner as S phase progresses.
Multiscale analysis of apparent replication speeds in other cell lines
The recent availability of high-resolution replication timing data in six other human cell lines (BG02, a human embryonic stem cell line; K562, a chronic myelogenous leukemia cell line; BJ, normal fibroblasts; GM06990, TL010, and H0287, lymphoblastoid cell lines) 
prompted us to carry out a similar multiscale analysis of their apparent replication speeds. As shown in , the distributions of replication speeds at the 100 kb scale were quite similar to HeLa cells with 3% (BJ) and <1% (other cells) of apparent speeds ≤2 kb/min, except for BG02 cells where a higher proportion of speeds ≤2 kb/min was observed (14.3%). Note that in the absence of published measurements of S phase length in these cell lines we have assumed a uniform S phase length of 8 h, typical of most mammalian cell lines. These distributions would be shifted toward proportionately faster (slower) speeds if S phase turned out to be shorter (longer). It is interesting to note that apparent speed distributions were much more similar among cell lines than single fork speeds and, by inference, origin activation patterns. This is consistent with a number of observations suggesting that replication timing is a more conserved feature among cell types than replication origin distribution 
. The difference between BG02 and the other cell lines presumably reflects the previously described smaller replication domain size and higher density of timing transition regions in embryonic stem cells than in differentiated cells 
Analysis of apparent replication speeds in multiple cell types.
As in HeLa cells, the observed ranges of apparent replication speeds in these cells cannot be explained by the range of single fork velocities measured by DNA combing in identical or comparable cells. In K562 cells, mean and max fork velocities are 0.37 kb/min and 1.0 kb/min 
whereas >99.9% of the genome replicates at apparent speed >1.0 kb/min. To our knowledge, replication fork velocities have not been measured in the five other cell lines. However, mean and max fork velocities have been estimated to 1.73 and 2.9 kb/min in MRC5 fibroblasts (M. Debatisse, pers. comm.) and to 2.06 and 4.4 kb/min in JEFF lymphoblastoid cells 
. Taking these values as reasonable estimates for BJ fibroblasts and for GM06990, TL010, and H0287 lymphoblastoid cells, respectively, it appears that 99.5–99.8% and 76–85% of the genome replicate faster than the mean and max fork velocity, respectively, in all those cell lines. Furthermore, mean fork velocities of 1.53–2.49 kb/min have been found in H9 and H14 embryonic stem cells 
. Assuming that mean and max velocity in BG02 embryonic stem cells are 2.0 kb/min and 4.0 kb/min, respectively, we find that 85.7% and 61.9% of the genome replicate faster than these respective speeds. Thus, a higher proportion of the genome replicates at an apparent speed compatible with unidirectional progression of a single fork in BG02 cells.
To further investigate this we analyzed the TTRs of these six cell lines individually (). The number of TTRs is about 2-fold higher in BG02 embryonic stem cells than in the differentiated cells (numbers in legend). Interestingly, a large fraction of the BG02 TTRs replicated at an apparent speed compatible with unidirectional progression of a single fork (; average apparent speed 2.34 kb/min, mean fork velocity 2.0 kb/min). In all the other cell lines (), however, the TTRs replicated faster than in BG02 (average apparent speed ranging from 3.24 kb/min to 4.21 kb/min, green lines), and faster than the mean fork velocity (compare dots with orange dashed lines). The discrepancy was most pronounced in K562 cells (), where no TTR replicated slower than the fastest single forks (vmax
1.0 kb/min, purple dashed line). In BJ fibroblasts () and in the three lymphoblastoïd cell lines (), however, many TTRs replicated at an intermediate speed between the mean and max fork velocity (orange and purple dashed lines, respectively). The possibility that the slowest TTRs are specifically replicated by the fastest forks cannot be formally discounted but seems unlikely, as explained above. Therefore, most of the TTRs in these five cell lines replicate faster than by a single unidirectional fork. In other words, internal initiation in TTRs is more frequent in differentiated cells than in BG02 stem cells. Furthermore, in cancerous cells K562 and HeLa, replication forks progress more slowly and this likely triggers additional origin activation in TTRs.
Size and replication time of individual TTRs in multiple cell types.
Importantly, if domains of constant replication time were separated by timing transition regions of uniform and slow replication speed 
, a biphasic distribution of apparent replication speeds should have been observed. This was not the case in any of the cell lines investigated. We found instead that the apparent replication speed, dx/dt
, has a continuous and wide-range distribution significantly faster than the known range of fork velocities, v
, in the vast majority of the genome. This implies that in all these cell lines, the statistics of origin activation creates throughout the genome a complete gradation in the predominance with which forks move in a preferred direction.
Comparison with previous genome-wide replication timing studies
Our findings appear to contradict earlier views of genome-wide replication timing in human and mouse cells, which proposed a strict dichotomy between large (0.2–2.0 Mb) CTRs containing multiple synchronous origins and smaller (0.1–0.6 Mb) TTRs with slopes consistent with unidirectional replication fork progression 
In the study by Desprat et al 
, the profiles were generated from the <2-fold copy number difference between S and G1 cells, which resulted in a low signal-to-noise ratio, and TTRs were defined as regions >250 kb in which the slope did not differ by more than 0.1 kb/min over their entire lengths. Such TTRs had slopes consistent with unidirectional fork progression (0.8–3.5 kb/min) but they only encompassed 5–8% of the genome. In three other studies 
, the profiles were generated from the abundance ratio of newly replicated DNA in different fractions of S phase and were segmented into CTRs and TTRs using a clustering algorithm. In all three cases, the resulting TTRs again only encompassed a small fraction of the genome (<10%). Hiratani et al 
and Ryba et al 
, who used only two fractions of S phase, found slopes consistent with unidirectional fork progression (0.8–3.5 kb/min), but Farkash-Amar et al 
, who used up to seven fractions of S phase, found faster slopes (1.5–6.5 kb/min). As can be seen in Figure S3
in Hiratani et al 
, having only two S phase fractions creates an essentially biphasic distribution of replication times, an artifact that is much attenuated by the use of four to six S phase fractions ().
The profiles we analysed in this work were generated from four 
or six 
fractions of S phase, allowing us to discern replication timing differences within regions that were merged as a single replication timing domain in previous studies. Furthermore, we determined the full distributions of apparent speeds before any segmentation of the genome. These distributions were continuous, not biphasic, which implies that any segmentation in CTRs and TTRs entails a degree of arbitrariness. When we delineated CTRs and TTRs as contiguous regions which replicate faster (resp. slower) than 10 kb/min at a 100 kb scale, the genome was partitioned in two nearly equal halves. However, to obtain a set of TTRs that encompass <10% of the genome, we would need to set the threshold at ~3 kb/min. Interestingly, the size range (0.1–0.5 Mb) and mean replication speed (2.3 kb/min) of such TTRs would be similar to those reported in the other studies, yet mostly incompatible with unidirectional fork progression given the fork speed measured by DNA combing in HeLa cells (, and data not shown). In none of the previous studies was the speed of replication forks directly measured on single DNA molecules in the same cells. We therefore believe that the rigid dichotomy reported in these studies overlooked the existence of a broad range of timing transition slopes, due to insufficient temporal resolution and/or to the use of a segmentation algorithm, and needs to be replaced with a more nuanced picture of DNA replication kinetics.
Although we do not exclude passive (but bidirectional) replication of TTRs<300 kb in HeLa cells, our data show that the mean replication progression rate along most of the genome is remarkably high, meaning that most TTRs are preferentially replicated by the progressive firing of multiple origins in most cells of a population. This is also the case for K562 cells. Nevertheless, we observed a higher proportion of apparent replication speeds consistent with unidirectional progression of a single fork in BG02 stem cells, and, to a smaller extent, in fibroblasts and lymphoblastoid cell lines in which replicons are longer and replication forks move faster than in HeLa cells.
DNA combing analysis of the IGH TTR
In the study by Desprat et al 
, the notion that TTRs are originless regions that replicate by unidirectional fork progression was strongly supported by a single molecule analysis of the human IGH locus. This experiment unambiguously demonstrated that most forks progress unidirectionally in this transition region in human ES cells, in agreement with ample evidence for unidirectional replication of the homologous locus in mouse ES cells and T lymphocytes 
. This behavior is cell-type dependent, however, since abundant initiation events were detected in the same region during early and late stages of mouse B cell development 
We found that in HeLa cells the IGH locus is included in a 440 kb TTR whose apparent replication speed is 3.77 kb/min, inconsistent with unidirectional replication and significantly faster than reported by Desprat et al 
in other cells (). We used DNA combing to determine the replication mode of this region ( and Figure S4
). We observed 43 initiation events on 25 DNA fibers evenly spread over a >700 kb region including the three restriction fragments studied by Desprat et al 
. Only two out of these 26 fibers were found to contain a single fork. We also found that replication fork velocities (1.48±0.21 kb/min, N
38) and inter-origin distances (46.0±5.1 kb, N
20) in this region were approximately similar to that of the bulk genome. These results unambiguously demonstrate that this TTR replicates by progressive activation of multiple replication origins in HeLa cells and confirm the validity of our multiscale analysis of apparent replication speeds in predicting regions that cannot replicate by unidirectional progression of a single fork. Together with the results of Desprat et al 
, they also confirm that the replication mode of the human IGH locus can change according to cell type, as previously reported for the mouse IgH locus 
. Since HeLa cells are derived from an adenocarcinoma, they show that replication origins in this region can be activated in non-B cells, although it is not clear if this results from a normal developmental program or from the tumoral nature of HeLa cells.
DNA combing analysis of the IGH TTR.
Mustiscale analysis of apparent replication speeds in the FRA3B fragile site
To further check our predictions of bidirectionally replicating CTRs and TTRs, we took advantage of the recent work of Letessier et al 
, who used DNA combing in fibroblasts and lymphoblastoid cells to reveal cell-type specific replication initiation programs at the FRA3B
chromosome fragile site.
Analysis of the replication timing data of Hansen et al 
shows that in BJ fibroblasts, the FRA3B
region is embedded into a late-replicating 0.9 Mb CTR that is predicted to replicate by synchronous initiations (). The DNA combing results of Letessier et al. 
entirely confirm this prediction, showing initiation and termination events evenly distributed all along this locus in MRC5 fibroblasts ().
Replication mode of the FHIT locus in fibroblasts and lymphoid cells.
In GM06990 lymphoblastoid cells, the FRA3B
region lies within a V-shaped replication timing trough formed by two converging TTRs, each about 1 Mb in length (). Both TTRs replicate at an apparent speed of 6–8 kb/min, inconsistent with the mean (1.87 kb/min) and max (3.2 kb/min) velocity of single forks measured within this locus by DNA combing in JEFF lymphoblastoid cells 
. Each of these two TTRs is therefore predicted to contain forks moving in both directions. A single fork moving at 2 kb/min could replicate up to 1 Mb of DNA within an 8 h S phase. Therefore, in principle, each TTR could be replicated without internal initiation if it is traversed by a single fork that is initiated two-thirds of the time on its early edge and one-third of the time on its late edge, since the resulting apparent speed would be v
6 kb/min (Figure S3C
). However, this scenario would predict that the edges of these TTRs would replicate either very early or very late in S phase, which is not supported by the timing data of Hansen et al 
(see Figure 3
in Letessier et al 
). The alternative hypothesis is that these TTRs replicate by internal initiations (Figure S3D
). The data of Letessier et al 
in JEFF cells indeed show initiations over the early and middle parts of each TTR, although initiations are excluded from a 700 kb region that corresponds to the late edges of both TTRs. Forks nevertheless are found to move in both directions in this 700 kb originless region 
, consistent with our predictions (). Another interpretation of these data would be that the edge of the early CTR is different in individual cells but the TTR is unidirectional in all cells, thus termination occurs at different points in the TTR. However, in order to quantitatively explain the discrepancy between the TTR slope and fork velocity, the position of this edge should differ by up to 1–2 Mb in different cells, which is again not supported by the timing data  
. Therefore, these results again confirm the validity of our analysis of apparent replication speeds in predicting regions that cannot replicate by unidirectional progression of a single fork.
Comparison with replication origin maps in ENCODE regions
Mesner et al 
have recently provided a reliable map of replication origins in 1% of the human genome in HeLa and GM06990 cells, using a novel replication-bubble trapping procedure to prepare nearly pure origin libraries that were hybridized to encyclopedia of DNA elements [ENCODE] microarrays 
. We compared the coverage of CTRs and TTRs by replication bubbles within these regions in HeLa cells (, and Figure S5
). Most CTRs and TTRs contained replication bubbles, consistent with our proposal that most of the genome replicates by internal initiations (see e.g. region ENm001, ), and we found no major difference in replication bubble coverage in TTRs (22%) vs. CTRs (29%). We noted however a higher bubble coverage in early replicating regions (). This was surprising because we found by DNA combing that interorigin distances did not change during S phase. Similar results were obtained with bubbles mapped in GM06990 cells (data not shown). A potential explanation for this discrepancy is that early bubbles are more efficiently trapped, perhaps because they are more efficient and less delocalised than late ones. Early bubbles may also have a longer dwell time than late ones because they are less synchronous and slower to merge with neighboring bubbles. This interpretation would imply that an even larger fraction of the genome than found by Mesner et al 
can support a significant level of delocalised, replication initiation of low efficiency.
Analysis of replication bubble and λ-SNS coverage in ENCODE CTRs and TTRs.
We also compared the coverage of CTRs and TTRs by short RNA-primed, nascent DNA strands purified by λ exonuclease digestion (λ-SNS) by Cadoret et al 
. Although there is only modest concordance between bubble and λ-SNS maps 
, we again found no major difference in λ-SNS coverage ( and Figure S5
) in TTRs (1.05%) vs. CTRs (1.71%) and a higher λ-SNS coverage in early replicating regions (). It is expected that λ-SNS peaks are less efficiently detected if initiation is more random in late replicating regions.
A gradient of chromatin openness along TTRs
A general correlation between replication timing, chromatin openness and transcriptional activity has been reported 
. To examine this in further detail, we analyzed the distribution along TTRs of an experimental marker (DNase I hypersensitive sites determined in HeLa S3 cells) and a DNA sequence marker (CpG islands) of open and transcriptionally active chromatin available genome-wide. We observed that the average coverages are maximum at TTRs early border and steadily decrease when going towards the late border (). These gradients of open chromatin marker distribution are unchanged when considering small, intermediate and large TTR size classes, suggesting there exists a characteristic scale for the change of chromatin state along TTRs. These results raise the possibility that there is a direct link between the gradients of origin firing time and a gradient in chromatin openness along the TTRs.
Representation of open chromatin markers along all TTRs relative to the corresponding genome-wide average value.
Temporal control of origin firing
The mechanisms that regulate the timing of replication are unknown. A simple model to account for our data is that origins have different relative firing probabilities and fire stochastically, and that the firing probability of all origins increases during S phase. Thus, efficient origins are likely to fire during early S phase and weak origins are unlikely to fire early but become more likely to fire during late S phase 
. The firing probability of origins may be specified by chromatin structure, since there is a general correlation between replication timing and chromatin openness 
(). Consistent with this model, we show here that markers of open chromatin are correlated with early replication throughout TTRs (), and we have previously reported that origin firing probability increases during S phase in a wide range of eukaryotes including human 
. Furthermore, both the combing data and the distributions of apparent replication velocities at different stages of S phase provide evidences for increasing origin firing during S phase. One observation, however, argues against a purely uniform and uncorrelated stochastic model: origin firings are temporally and/or spatially correlated. It is possible that neighbor origins fire independently of each other but are nevertheless temporally correlated because their timing is set by some underlying chromatin features that change over a characteristic distance longer than individual replicons.
Two alternative models for origins activation along TTRs.
An attractive alternative mechanism to explain the progressive activation of neighboring origins along the TTRs is that active forks stimulate nearby initiation in unreplicated DNA. As discussed elsewhere 
, forks may stimulate initiation due to changes in DNA supercoiling in front of the fork or to association of chromatin remodellers or origin triggering factors with replication fork proteins. Early studies of replication foci labelled by two consecutive pulses showed that the intranuclear distance between consecutively replicated domains increased linearly with the time interval between the labels 
. Studies on the dynamics of PCNA assembly at replication foci indicated that once replication is completed at a given site, a new replication focus assembles de novo at a neighboring site, consistent with a domino effect in activation of neighboring origins 
. A more recent study of S phase progression in HeLa cells suggested that replication foci that lie side-by-side in the nuclei are replicated in consecutive intervals of S phase because of their genetic continuity along the chromosomal fiber and that a “next-in-line” principle defines the efficiency with which origins are activated once S phase has begun 
. In this work, we have quantitatively analyzed the speed of the replication wave progression and shown that it is consistent with a cascade of origin activation along TTRs as predicted by a domino model for origin activation. Thus, replication would first initiate in efficient zones of variable size specified by an open chromatin structure 
, followed by progressive activation of flanking origins in less open chromatin due to the approach of an incoming fork (). This model explains why adjacent origins tend to fire synchronously, why replication progresses faster than a single fork and why origins embedded in closed chromatin do not fire in early S phase but fire efficiently when the replication wave reaches them. With an increasing rate of origin firing during S phase 
, this domino model can further explain why the apparent speed of replication increases along replication timing gradients, and predicts a progressive change in replication fork polarity along these gradients.
Works from several groups suggest that activation of one origin within a potential initiation zone suppresses rather than activates the activation of immediately surrounding origins 
. However, the range of this negative origin interference is limited to distances smaller than the typical interorigin distance and is not incompatible with positive origin interference acting over larger distances 
. Data on origin spacing and synchrony in Xenopus egg extracts are indeed consistent with a mechanism whereby loop formation between a potential origin and an approaching fork suppresses initiation at very close spacing and enhances initiation at a larger, characteristic distance 
In favor of a role of fork progression in controlling sequential origin activation, a recent study in yeast has shown that mutants deficient in chromatin remodeling activities located at replication forks specifically delay the replication of late replicating domains 
. On the other hand, a study with mammalian cells has shown that exposure of aphidicolin-arrested cells to checkpoint inhibitors results in initiation of replication at successively later-replicating domains in the absence of detectable elongation of replication forks 
. This suggests that fork elongation is not strictly required for at least the global aspect of temporal origin activation, but does not prove that it has no role in this process. Furthermore, it is possible that only the earliest origins are activated in successive large-scale replication domains, and that secondary origins within a domain require activation by replication forks.
In this work, we have performed a quantitative analysis of human genome replication in cells sorted into four or six stages of S phase, using DNA combing, mathematical analysis of replication timing profiles generated by massive sequencing of newly replicated DNA, and bioinformatic analysis of replication origin maps and chromatin structure data. The results show that i) replication origins fire in a correlated manner and at an increasing rate during S phase, ii) the apparent speed of replication progression throughout the genome depends on both the velocity of single forks and the proportion of rightward and leftward moving forks in the cell population, and ultimately reflects the pattern of origin firings along replication timing gradients rather than the unidirectional progression of a single fork. The correlation between adjacent origin firings may be due to their common chromatin environment or to a stimulation of origin firing by approaching forks. Further analyses and mathematical modelling of replication timing profiles are underway to explore these issues.