A quantitative FastFUCCI assay defines cell cycle dynamics at a single-cell level

The fluorescence ubiquitination-based cell cycle indicator (FUCCI) system represents a quantum leap for real-time examination of cell cycle progression (Sakaue-Sawano et al., 2008). This technology relies on the choreography of two players of the DNA replication control apparatus, the licensing factor CDT1 and its inhibitor geminin (Zielke and Edgar, 2015). In eukaryotic cells, CDT1 level peaks during G1 and plummets upon S entry. Conversely, the geminin level is high during S and G2 but is low during late mitosis and G1. The alternating expression of these two proteins arises from the sequential activation of the E3 ubiquitin ligases SCF^Skp2 (a Skp1–cullin-1–F-box complex that contains Skp2 as the F-box protein), and the anaphase-promoting complex/cyclosome (APC/C), which target the degradation of CDT1 and geminin, respectively. By fusing fluorescent proteins to degrons derived from CDT1 and geminin, the FUCCI system thus enables continuous visualisation of live cells in either G1 or the S- and/or G2-to-M phases (hereafter denoted as S/G2-M).

To date, various FUCCI-based laboratory models have been employed together with live-cell imaging to elucidate the effects of therapeutic modulation (Sakaue-Sawano et al., 2011; Yano et al., 2013; Dan et al., 2012; Koh et al., 2015). However, despite the detailed molecular insights one can gain from such imaging, the approach remains uncommon owing to its technical complexity. In particular, advances in automated imaging hardware have not been matched by the ease and accessibility of multiparametric image-based processing workflows. As a result, many live-cell assays are not designed with good throughput capability and often depend on manual handling of a limited number of datasets (Chirieleison et al., 2011; Boutros et al., 2015; Singh et al., 2014).

Here, we developed a rapid selectable lentiviral-based system called FastFUCCI that allowed simultaneous introduction of both FUCCI transgenes, that is the CDT1- and geminin-tagged reporters, into cells. The method alleviated the conventional requirement for sequential cell line transduction and characterisation, thereby improving labelling efficiency. We coupled the FastFUCCI system to a cascade of image acquisition, processing and analysis streamlined for the single-cell level. This was achieved by developing a versatile live-cell assay implemented in a modular design, where each handling step was independently scalable and automated by open-source scripts. A key feature of this assay was its ability to perform batch management of high-volume data; the quantitative output of these data was image-based and therefore amenable to downstream validation.

To demonstrate the utility of the FastFUCCI imaging assay, we applied it to define the cellular and molecular events in cancer cells following treatment with antimitotic agents. Antimitotic agents such as taxanes are frequently used for cancer treatment. These agents disrupt mitotic progression, resulting in structural aberrations that in turn elicit mitotic arrest. Although the mechanism of these events is well described, comparatively little is known about how cells respond to such insults spatiotemporally (Gascoigne and Taylor, 2009; Rieder and Maiato, 2004). To this end, we first monitored the cell cycle dynamics at the single-cell level in response to treatment with varying concentrations of paclitaxel and nocodazole. We then correlated the cell cycle distributions with the temporal dimension, as quantified by single-cell tracking of the different cell cycle phases. Using the FastFUCCI assay revealed actionable spatiotemporal effects induced by antimitotic agents, leading to evidence that co-targeting mitosis and the DNA replication origin activity was a potential therapeutic strategy.

Traditional generation of a FUCCI-expressing cell line entails sequential expression of the two FUCCI reporters, CDT1 and geminin. To simplify this procedure, we directed the simultaneous expression of both reporters in equimolar amounts by cloning both transgenes [monomeric Kusabira Orange 2 (mKO2)–hCDT1(30–120) and monomeric Azami Green (mAG)–hGEM(1–110), where m denotes mouse and h denotes human], separated by a T2A sequence, downstream of a single EF1α promoter. This expression cassette was subsequently cloned into the lentiviral backbone pBOB to permit fast generation of a cell population that stably expressed both FUCCI reporters (Fig. 1A; Fig. S1, Table S1).

We introduced the expression of both FUCCI reporters in human pancreatic cancer MIA PaCa-2 cells through lentiviral transduction. In stably transduced cells, fluorescence was observed in the nuclei, and both red (mKO2) and green (mAG) signals alternated predictably according to the phases of the cell cycle (Fig. 1B). Flow cytometric analysis shows comparable cell cycle distributions in both parental and FastFUCCI-expressing cells (Fig. 1C). The red-, yellow- and green-emitting populations in the latter corresponded to cells in G1, the G1/S transition and S/G2-M, respectively.

The increasing use of multiparametric microscopy assays has demanded workflows that can manage practical issues related to processing and analysis of large datasets (Singh et al., 2014). To acquire and analyse time-lapse data in an unbiased and automated fashion, we proceeded to develop the FastFUCCI imaging assay (Fig. 2A; Fig. S2A,B). We streamlined the assay with automation at each principal step to afford user-independent handling of a high number of images. The output was time-lapse data with consistent image resolution and contrast, allowing downstream analyses to be performed in batch mode by commercial and open-source image analysis platforms, as tested with the Columbus™ Image Data Storage and Analysis and Fiji, respectively, in this work (Fig. 2A).

To demonstrate these functional features, we performed a 12-h study on the FastFUCCI-expressing MIA PaCa-2 cells treated with the previously characterised CDK1-selective inhibitor RO3306, which was expected to arrest cells at the G2/M boundary (Vassilev et al., 2006). CDK1 inhibition induced a dose-dependent increase in the green population and a corresponding decrease in the red population (Fig. 2B). Consistent with previous studies, this effect on single-cell spatiotemporal dynamics was rapid, as the apparent difference was noticeable as early as 1 h following treatment, reflecting an accumulation of S/G2-M cells (Vassilev et al., 2006; Marcus et al., 2015). Automated single-cell duration measurement showed that the increase in the S/G2-M population was correlated with the residence time in the green phase (representing this population), the latter also exhibiting a dose-dependent trend (Fig. 2C).

We exploited the assay to define the single-cell spatiotemporal dynamics induced by paclitaxel at high definition. Based on the IC₅₀ of paclitaxel (4.7±0.3 nM; mean±s.e.m.) (Fig. 3A), we subjected FastFUCCI-expressing MIA PaCa-2 cells to a low (3 nM) and a high (30 nM) concentration of paclitaxel. Cell cycle progression was monitored, with total number of cells being taken as a parameter for proliferation (Fig. 3B–E). We found that, in the vehicle control, the fractions of red and green cells each displayed a reciprocally oscillating profile as total cell count increased over time, in line with the predicted behaviour of cycling single cells (Fig. 3B,C). In 3 nM paclitaxel, the alternating oscillation remained evident, with minimal changes in the cell cycle distributions consistent with the minimal inhibition of proliferation (Fig. 3B,D). In contrast, in 30 nM paclitaxel, there was an initial accumulation of the green fraction up to 6 h (Fig. 3E). From 6 to 24 h, the green fraction decreased with a concomitant rise in the red fraction, implying that the initial mitotic arrest was followed by either mitotic death or mitotic slippage (the process by which cells proceed through mitosis to interphase without chromosome segregation). Given that cell proliferation did not appear to decline until after 24 h, this suggests that there was mitotic slippage, rather than extensive death in mitosis (Fig. 3B). Similar observations were made in paclitaxel-treated FastFUCCI-expressing Panc-1 cells, where accumulation of green cells was followed by accumulation of red cells (Fig. S3A–C).

To correlate the spatiotemporal dynamics with cell cycle residence time, we quantified the cell cycle durations of individual FastFUCCI-expressing MIA PaCa-2 cells under the same conditions (Fig. 3F,G). In control, the average S/G2-M duration was 8.9±0.2 h (mean±s.e.m.), comparable with previous results (Koh et al., 2015). Cells treated with 3 nM paclitaxel experienced similar S/G2-M duration (9.7±0.3 h), whereas those treated with 30 nM paclitaxel had almost double the duration (16.1±0.6 h) compared to control. The dose-dependent delay in S/G2-M was correlated with G1 prolongation (Fig. 3G). All these effects on cell cycle dynamics were similarly observed with nocodazole, another mitotic inhibitor, indicating that the findings were not specific to taxanes (Fig. S3D–H).

The temporal dichotomy of the red and green fractions induced by the antimitotic agents prompted us to use the FastFUCCI system to determine the basis of their cytotoxicity (Fig. 3E; Fig. S3C,F). First, we performed manual cell fate analysis of 30 nM paclitaxel-treated FastFUCCI-expressing MIA PaCa-2 cells (Fig. 4A). Of the 190 randomly selected cells tracked for 72 h, all of them attempted mitosis, with 2% eliminated during mitosis. In the remaining 98%, three distinct mitotic phenotypes were identified: abortive mitosis (no daughter, 57%), bipolar mitosis (two daughters, 32%), multipolar division (three or more daughters, 9%).

Next, we determined the ensuing fate of cells that had progressed through mitosis into the second cell cycle (Fig. 4B). Independent of the nature of the preceding mitotic phenotype, the majority of these cells (73–75%) died in the red phase (i.e. G1). An important observation was that some of these cells arrived at the red phase without overt reflattening before death and appeared morphologically similar to those that underwent mitotic death in green phase (Fig. 4C). The remaining fraction either died in the next green phase (10–15%) or appeared to remain alive (12–16%) (Fig. 4B).

To substantiate these live-cell FastFUCCI analyses, we performed quantitative immunofluorescence of MIA PaCa-2 cells to enumerate the proportion of cells that were dead or dying in mitosis. We employed cleaved caspase 3 (c-caspase 3) and cyclin B1 as markers of apoptosis and mitosis, respectively (Fig. 5A). Irrespective of treatment duration (0–48 h), we found that less than 30% of c-caspase-3-positive cells were positive for cyclin B1 (Fig. 5B). Validation using another apoptotic (γH2AX) and mitotic (pHH3) markers yielded the same conclusion – that mitosis was not the main phase in which cell death occurred (Fig. 5C,D; Fig. S4A). All these results were verified in Panc-1 cells (Fig. S4B–D).

Our findings suggest that post-mitotic entry into subsequent interphase is important for maximal paclitaxel cytotoxicity. To test this hypothesis, we performed a series of scheduled washout experiments on MIA PaCa-2 cells (Fig. 6A). We found that there was a duration-dependent decrease in proliferation index compared to time-matched controls (Fig. 6B). When the incubation time for 30 nM paclitaxel was limited up to 12 h (i.e. before the red fraction overtook the green fraction in Fig. 3E), the proliferation index remained above unity, indicating that cells had proliferated. A 24-h incubation largely prevented subsequent proliferation whereas a 48-h incubation had a killing effect, consistent with the total cell count dynamics in Fig. 3B.

Whereas drug incubation of at least 24 h appeared necessary for short-term growth suppression, from the fraction of cells that remained viable even following a 3-day treatment (Fig. 4B), 26% subsequently re-entered S phase (Fig. 6C). Based on the paradigm that re-entry into active cell cycle could underlie chemoresistance, we investigated whether inhibiting the DNA replication origin firing following paclitaxel treatment could improve chemosensitisation. We performed long-term cell proliferation studies, where we pretreated MIA PaCa-2 cells with 30 nM paclitaxel for 24 h and released viable cells into roscovitine, a pan-CDK inhibitor commonly used to suppress origin activity. As expected, we found that 24-h paclitaxel pretreatment impaired but did not obliterate proliferation compared to DMSO control (Fig. 6D,E). Additional treatment with roscovitine further reduced the proliferative capacity compared to the Bliss-predicted additivity value, suggesting an apparent synergistic interaction. Replacement of roscovitine with PHA767491 to inhibit CDC7, another replication origin regulator, reproduced this potential synergistic relationship with paclitaxel pretreatment. Similar results were obtained in Panc-1 cells (Fig. 6D,E).

Devising methods to visualise single-cell states is an area of active development (Walling and Shepard, 2011). In particular, as interest in multiparametric imaging assays grows, the need for efficient workflows that encompass data acquisition, processing and analysis has become more evident (Chirieleison et al., 2011; Singh et al., 2014). Here, we developed a selectable lentiviral reagent that enables rapid expression of the two FUCCI reporter transgenes in cells. We applied this FastFUCCI system to a streamlined imaging workflow that performs automated image processing and analysis. Practically, the FastFUCCI imaging workflow is supported by open-source scripts and is modular in its construction, with the output being digital images of accessible format (TIFF). Hence, it is amenable to current commercial and open-source image analysis programmes, as demonstrated here by its coupling with the Columbus™ and Fiji software, respectively. Analytically, the workflow enables quantification of large numbers of images, objects or events, conferring the statistical power necessary to detect subtle, transient or rapid real-time dynamics (Hamilton, 2009). Extraction of multidimensional data at such a scale would be a resource for studies attempting to integrate the FUCCI technology with mathematical modelling (Saitou and Imamura, 2016). Equally, addition of other imaging features, such as shape recognition, to the workflow would further diversify its experimental applications.

To test the utility of the FastFUCCI imaging assay, we applied it to dissect the spatiotemporal effects induced by antimitotic agents. Earlier studies had attempted to establish a link between paclitaxel-induced cell death and mitotic exit, but it remained unclear if mitotic exit represented a major pro-death event (Bekier et al., 2009; Tao et al., 2005; Gascoigne and Taylor, 2008; Shi et al., 2008; Huang et al., 2009; Orth et al., 2008). Although the ambiguity might have stemmed from interline variation, the understanding so far has been drawn from cell-population-based experiments, phase-contrast images or low-content microscopy, precluding the precise discrimination between cells that die in mitosis and cells that die following a mitotic attempt. More recently, several qualitative studies using FUCCI-expressing cells have reported the emergence of CDT1 expression following prolonged mitosis (Honda-Uezono et al., 2012; Kaida et al., 2011; Marcus et al., 2015). Through in-depth single-cell analyses, we found this phenomenon to be far more common than previously implied. Importantly, we forged a quantitative association between post-mitotic entry and cell death, the latter being independent of the mitotic phenotype. Our results are supported by previous studies that investigated the role of post-mitotic entry in finalising apoptosis (Tao et al., 2005; Orth et al., 2012). Further mechanistic studies are warranted to elucidate how mitotic arrest and subsequent interphase entry differentially engage the resolution of death.

Three therapeutic considerations arise from this work. First, conventional regimens are often based on the principle of maximum tolerable doses (Marshall, 2012). We have shown that, even at a cytotoxic concentration that elicits growth inhibition, cancer cells do regain some extent of proliferative capacity when drug residence falls below a critical threshold, that is when cells have not yet progressed beyond mitosis and become committed to apoptosis. This suggests that paclitaxel sensitivity depends not only on the strength of the mitotic checkpoint but also on the extent of checkpoint adaptation (Weaver and Cleveland, 2005). An understanding of how these mechanisms enact in vivo in relation to drug pharmacokinetics would inform future development of the recently proposed adaptive strategy for paclitaxel (Enriquez-Navas et al., 2016).

Second, our in vitro data using pancreatic cancer cell lines recapitulate the phenomenon seen in a fibrosarcoma xenograft model, whereby cancer cells accumulate in G1 following paclitaxel-induced mitotic arrest (Chittajallu et al., 2015). Our high-definition spatiotemporal analyses indicate that a large fraction of these cells are the direct consequence of treatment, having progressed through aberrant mitosis into G1 (Vakifahmetoglu et al., 2008; Shi et al., 2008). Some of these cells appear viable, and most of them stay in G1, presumably in an arrested or senescent mode. However, a few of the survivors subsequently re-enter S phase. It is conceivable that acquired resistance might arise from these cells that have been therapeutically challenged but do not succumb (Ogden et al., 2015).

Third, a second agent that targets the surviving post-mitotic population might confer further chemosensitisation. We demonstrated this possibility with the synergistic interaction between paclitaxel and roscovitine as well as PHA767491, two compounds known to affect DNA replication origins. Other small-molecule inhibitors with overlapping mechanisms, namely CINK4, flavopiridol, PD-0332991 and purvalanol, have previously been shown to enhance taxane cytotoxicity or reverse taxane resistance (Zhang et al., 2013; Ingemarsdotter et al., 2015; Bible and Kaufmann, 1997; O'Connor et al., 2002). As such a co-targeting strategy is schedule dependent in most cases, the challenge would be to identify optimum dosing regimens that could yield maximal efficacy and minimal toxicity (Jackson et al., 2007; Shah and Schwartz, 2001). We anticipate the FastFUCCI imaging assay to be a useful preclinical tool in these endeavours.

A synthetic linker was designed and cloned into pBluescript (pBS) to enable the cloning of the FUCCI reporters [mKO2–hCDT1(30–120) and mAG–hGEM(1–110); GeneArt Gene Synthesis] on either side of and in frame with a T2A sequence. This whole cassette was then dropped from pBS and cloned into a third generation lentiviral plasmid (pBOB), as a BamHI and NotI fragment. An EF1α promoter, to drive robust and constitutive transgene expression, was then cloned into pBOB upstream of the FUCCI cassette as a PacI fragment. Finally, a PGK-PURO cassette was cloned as an NheI fragment to enable in vitro selection of transduced cells with puromycin. The essential components of the construct are illustrated in Fig. 1A, full map is in Fig. S1, and all single and two-cutter 6+ nucleotide restriction enzymes with respective positions and actual nucleotide numbers are in Table S1.

Generation of FastFUCCI-expressing cells was performed using VSV-G pseudotyped lentiviral particles as previously described (Rodriguez et al., 2014). Briefly, HEK 293T cells were transfected with pBOB-EF1-FastFUCCI-Puro and packaging plasmids. Supernatant from the culture medium containing virus was then collected and used to transduce human pancreatic cancer MIA PaCa-2 and Panc-1 cells (European Collection of Cell Cultures). Following puromycin selection, polyclonal cell populations were expanded. The identity of both parental and transduced cell lines was authenticated by STR genotyping. In all experiments, cells were grown in Dulbecco's modified Eagle's medium (DMEM) with 10% foetal bovine serum (GIBCO) and were used up to a maximum of 20 passages. All cell lines were verified to be mycoplasma-free using the Mycoprobe Mycoplasma Detection Kit (R&D Systems).

Nocodazole (Sigma), paclitaxel (Tocris or Cayman Chemical), PHA767491 (Tocris), RO3306 (Calbiochem) and roscovitine (Sigma) were dissolved in dimethylsulphoxide (DMSO, Sigma), kept at −20°C and used within 3 months. The final working DMSO concentration was kept at 0.1%.

Cells were fixed with 4% paraformaldehyde and stained with DAPI. Red and green fluorescence were captured by the BD LSR II flow cytometer using a blue octagon detector array (488-nm laser) with 555 long-pass and 575/26 band-pass filters, and 505 long-pass and 530/30 band-pass filters, respectively. DAPI was captured using a UV trigon detector array (355-nm laser) with a 450/50 band-pass filter. 10,000 events were collected, and data were processed using the FlowJo software (Tree Star).

Cells were seeded in glass bottom wells (ibidi GmbH) and were kept in a humidified chamber under cell culture conditions. For each well, images were taken from three or more fields of view, where there were no excessive cell clumps so as to capture cells unaffected by potential cell–cell contact inhibition and to facilitate downstream analyses. Images were acquired with either a Nikon Eclipse TE2000-E microscope with a 20×0.45 NA long-working distance dry objective equipped with a sCMOS Andor Neo camera, or a Carl Zeiss Axio Observer Z1 with a 20×0.8 NA dry objective equipped with an Axiocam 506 CCD camera. With the Nikon microscope, red and green fluorescence were captured using a pE-300^white CoolLED filtered by Nikon FITC B-2E/C and TRITC G-2E/C filter cubes, respectively. With the Zeiss microscope, they were captured using a pE-4000 CoolLED filtered by Zeiss 38HE and 43HE filter cubes, respectively. To achieve optimal image resolution without excessive illumination, a binning factor of 2 by 2 for the sCMOS camera and 5 by 5 for the CCD camera was applied prior to imaging. Throughout each experiment, the ambient conditions of the imaging platform (e.g. external lighting) were maintained to minimise variations in optical resolution and illumination.

Processing of acquired videos was achieved automatically with a script written for Fiji (Schindelin et al., 2012). First, splitting of fluorescence images as single-channel time-lapse videos was performed using the Bio-Formats plug-in in order to convert files into the universal TIFF format. This processing step would also reduce video size, especially if binning was not applied during acquisition. Second, background subtraction and shading correction were applied to individual fluorescence channels. Depending on microscopy conditions (e.g. illumination source and objective shading effect), either the ‘Subtract background’ feature in Fiji was applied directly, or the Gaussian filter was used to evaluate and subtract the background. Optionally, the ‘Bleach correction’ plug-in could be used to further reduce auto-fluorescence. Together, these steps minimised intensity variation due to wide-field illumination and improved the signal-to-noise ratio, thus facilitating downstream analyses. A summary of the image processing pipeline is illustrated in Fig. S2A. Links to relevant scripts are provided below.

To monitor real-time cell cycle distribution, time-lapse images were processed as described above and were analysed in batch mode using the Columbus™ software (PerkinElmer). Image formatting was performed automatically with a Fiji script, which was designed to create an additional red–green merged channel and include time splicing. The resulting TIFF time-lapse images were then imported to the Columbus™ database using PerkinElmer Insight extension. Nuclear segmentation was performed on the merged red–green channel, optionally de-noised with Gaussian filtering, to detect all relevant nuclei. Sorting was performed to discard objects at the edges of field of view, as well as abnormally small and large objects by minimum and maximum size filter, respectively. Classification of the red-, yellow- and green-emitting populations was undertaken according to the mean and maximum intensities of the fluorescence within the nuclear area. Thresholds were defined based on the fluorescence charts generated by the Columbus™ interface. Fig. S2B depicts the procedure taken to analyse the data using the Columbus™ software. Scripts for image processing and analysis in this section are available at https://github.com/pmascalchi/FastFUCCI-cellcycle.

For single-cell tracking, images were processed as described above and were analysed in batch mode using the TrackMate plug-in available in the Fiji package (Tinevez et al., 2016). This plug-in operated on a user-specified framework that enabled automation of spot segmentation and frame-to-frame spot-tracking. The Laplacian of Gaussian filter was applied to detect spots, and the Linear Assignment Problem tracker was used to track the movement of detected spots. Elimination of false positives (i.e. tracks with short durations) was performed in the final step of the automated protocol. The output included the xy coordinates and the track durations of detected spots, as well as videos with tracking paths for quality control. Batch analysis of videos was achieved using the AutoClick Robot software (http://autoclick-robot.en.softonic.com), which allowed automatic mouse clicks and keyboard inputs, complemented with a customised Fiji script to manage sequential video opening, exported data handling and file formatting. Scripts for image processing and analysis in this section are available at https://github.com/pmascalchi/FastFUCCI-tracking.

Cells were fixed on glass bottom wells (ibidi GmbH) with 4% paraformaldehyde, immunostained and processed as previously described (Koh et al., 2015). Briefly, images were captured using the iCys laser scanning cytometer (CompuCyte). A digital image was generated for each photomultiplier on a pixel-to-pixel basis. Objects were scored based on preset signal thresholds. The output was processed by the iCys software.

Primary antibodies against c-caspase 3 (#9664, 1:400), cyclin B1 (#4135, 1:1000) and pHH3 (#3377, 1:1600) were from Cell Signaling Technology. γH2AX (#05-636, 1:2500) was from Millipore. Secondary antibodies conjugated to Alexa Fluor 488 (#4408, #4412) and Alexa Fluor 647 (#4410, #4414) were from Cell Signaling Technology and were used at 1:1000.

For sulforhodamine B cytotoxicity assays, cells were fixed with trichloroacetic acid following treatment and were stained with sulforhodamine B dye. Fluorescent readout was obtained using the Infinite 200 PRO microplate reader (Tecan) at excitation and emission wavelengths of 488 nm and 585 nm, respectively. For long-term cell proliferation assays, cells were first subjected to DMSO or paclitaxel treatment for 24 h. Cell viability was then determined using Trypan Blue exclusion, and equal numbers of viable cells were reseeded in fresh medium containing DMSO, roscovitine or PHA767491. After 2 weeks, cells were fixed and stained with 0.2% Crystal Violet (Sigma). Crystal Violet was solubilised using 10% acetic acid and the absorbance was measured at 590 nm using the PHERAStar plate reader (BMG LABTECH). The Bliss additivity prediction was derived from R_{AB Bliss} (a,b)=E_A(a)•E_B(b), where R denotes the reference effect, E denotes the observed effect, A and B denote the agents of interest, a and b denote their respective concentration.

We thank Masashi Narita and Fanni Gergely for the FUCCI transgene sequences, Inder Verma Laboratory (Salk Institute) for the pBOB lentiviral system, Cancer Research UK Cambridge Institute Light Microscopy (Heather Zecchini, Jeremy Pike, Lorraine Berry, Stefanie Reichelt) as well as Biorepository Core Facilities for technical provision, and Kevin Brindle for guidance and provision of laboratory space to E.R. and S.K.L. We acknowledge the support from the University of Cambridge, Cancer Research UK and Hutchison Whampoa Limited.