Abstract
The transcription factor Runx1 is essential for definitive hematopoiesis, and the RUNX1 gene is frequently translocated or mutated in leukemia. Runx1 is transcribed from two promoters, P1 and P2, to give rise to different protein isoforms. Although the expression of Runx1 must be tightly regulated for normal blood development, the mechanisms that regulate Runx1 isoform expression during hematopoiesis remain poorly understood. Gene regulatory elements located in non-coding DNA are likely to be important for Runx1 transcription. Here we use circular chromosome conformation capture sequencing to identify DNA interactions with the P1 and P2 promoters of Runx1, and the previously identified +24 enhancer, in the mouse multipotent hematopoietic progenitor cell line HPC-7. The active promoter, P1, interacts with nine non-coding regions that are occupied by transcription factors within a 1 Mb topologically associated domain. Eight of nine regions function as blood-specific enhancers in zebrafish. Interestingly, the +24 enhancer interacted with multiple distant regions on chromosome 16, indicating it may regulate the expression of additional genes. The Runx1 DNA contact map identifies connections with multiple novel hematopoietic enhancers that are likely to be involved in regulating Runx1 expression in hematopoietic progenitor cells.
Background
Runx1 is a key regulator of hematopoietic development. Deletion of Runx1 in mouse embryos is lethal in embryonic stage (E) 12.5 due to the complete absence of definitive blood cell progenitors accompanied by extensive hemorrhaging1,2. Runx1 is crucial for hematopoietic stem cell (HSC) emergence and maintenance during development3, since conditional ablation of Runx1 in adult mice results in HSC exhaustion4. In acute myeloid leukemia (AML) and myelodysplastic syndrome, RUNX1 function is frequently altered through mutations or translocations5, resulting in dysregulation of its target genes. While mutations directly affecting the RUNX1 protein are common in leukemia, mutations in regulatory elements that affect RUNX1 expression remain enigmatic. As yet unidentified mutations in regulatory elements, such as enhancers, could alter Runx1 expression, resulting in abnormal hematopoiesis.
Runx1 is transcribed from two promoters, P1 and P2, to give rise to different protein isoforms6. Expression of these isoforms is tightly controlled during hematopoiesis. At the onset of mouse hematopoiesis (E7.5), preceding the generation of HSCs, expression of the P2 isoform(s) is predominant7,8. P1 is expressed soon after P2, and its expression is synchronized with the generation of HSCs7,9. P1 expression is predominant in the mouse fetal liver, the main site of definitive hematopoietic stem/progenitor cell (HSPC) development from E12.5 onward9.
Regulatory elements, such as enhancers, can control the expression of genes via long-range chromatin interactions10. One previously identified Runx1 enhancer is located 24 kb downstream of the P1 transcriptional start site11,12. The +24 enhancer (also known as the +2313, or the +23.512 enhancer) is active in HSCs that express Runx1 during mouse embryogenesis11-13. The human equivalent to the +24 enhancer (+32 kb downstream of P1) directly contacts the promoters of RUNX1 in leukemia cell lines14. In addition to the +24 enhancer, putative regulatory elements for RUNX1 have been identified upstream of RUNX-P1 and between P1 and P2; however, whether they directly contact the RUNX1 promoters has not been investigated15,16.
Here we used circular chromosome conformation capture sequencing (4C-seq) to identify regulatory elements that interact with an active Runx1 P1 promoter, versus an inactive P2 promoter. While Hi-C provides genome-wide contact profiles, and Capture Hi-C enriches for interactions with preselected genomic features (usually promoters), 4C-seq can generate very high resolution contact profiles from ‘baits’ of particular interest17. Therefore, 4C-seq can yield richer information about a selected genomic region than Hi-C or Capture Hi-C. HPC-7 is a well characterized mouse HSPC line with genomic annotations, including transcription factor (TF) binding, histone modifications and chromatin accessibility18-20. 4C-seq in HPC-7 cells identified nine new hematopoietic enhancers that interact with the P1 promoter and +24 enhancer, and that are occupied by hematopoietic TFs. Eight of these were active in zebrafish hematopoiesis. Further, the +24 enhancer was highly interactive both within a topologically associated domain (TAD) harboring Runx1, as well as with loci outside the TAD. Collectively, our results point to the formation of a local ‘active chromatin hub’ controlling Runx1 expression in hematopoietic cells.
Results
We first confirmed that P1 is actively expressed in HPC-7 cells, while P2 is silent (Supplementary Fig. S1). 4C-seq in HPC-7 cells using the P1 and P2 promoters and the +24 enhancer as ‘baits’ identified genomic interactions at Runx1 (Fig. 1A). Bait locations were designed taking into account cohesin and CTCF binding sites near both promoters and the +24 enhancer. 4C baits were designed to regions of interest (P1, +24, P2), allowing for comparison of interactions between the active P1 promoter and inactive P2 promoter, with secondary baits located at nearby cohesin/CTCF (cc) binding sites (P1cc, +24cc, P2cc) (Fig. 1A).
4C-seq in HPC-7 cells identifies a 1.1 Mb domain harboring Runx1
For each bait, two replicate 4C-seq libraries and one control library were sequenced. Reads were predominantly located within a 1.1 Mb region surrounding the Runx1 gene (Fig. 1B-C), representing a TAD harboring Runx1. Domain boundaries are present at the Cbr1/Setd4 genes upstream of Runx1, and Clic6 downstream (Fig. 1B). A comparison with existing Hi-C data from mouse embryonic stem cells, mouse CH12 (erythroleukemia) cells, human GM12878 (lymphoblastoid), human K562 (myeloid leukemia) and human IMR90 (foetal lung) cells revealed that TAD boundaries are conserved, and are consistent with our 4C data (Supplementary Fig. S2). Most P1 and +24 interactions take place upstream of the Runx1 gene (Fig. 1B,C). In contrast, there are fewer upstream interactions from P2, while downstream interactions are retained (Fig. 1B,C). There are no other coding genes within the Runx1 domain; however, three inactive non-coding genes19, Mir802 and the long non-coding RNAs 1810053B23Rik and 1700029J03Rik, are located near the upstream border of the TAD.
Interactions from ‘cc’ baits (cohesin/CTCF binding sites) were similar to their nearby corresponding baits at P1, +24 and P2 (Fig. 1B). We investigated whether significant interactions for the ‘cc’ baits have more overlap with other cohesin or CTCF binding sites than for the non ‘cc’ baits, but did not find any difference (data not shown). The ‘cc’ baits may be too close to the corresponding P1, P2 and +24 baits to resolve unique interactions, although they are separated by at least one restriction fragment.
Chromatin interactions anchored by the Runx1 P1 and P2 promoters and +24 enhancer
Most significant interactions for P1, P2 and the +24 enhancer occur within the 1.1 Mb domain (Fig. 1B). There are ~40-50 significant interactions for P1cc/P1 baits, ~60-75 for +24/+24cc baits, and ~30 for P2/P2cc baits (Fig. 1D). Strikingly, the +24 enhancer in particular forms many significant interactions outside the Runx1 domain (Fig. 1B,D). Many of these long-range interactions are with other genes or gene promoters on mouse chromosome 16 (Fig. 2). Among the genes contacted by +24 are Erg, a hematopoietic TF, and Tiam1 (T-cell lymphoma invasion and metastasis 1), involved in cell adhesion and cell migration. These distant connections indicate that +24 may also regulate other hematopoietic genes in adjacent domains.
Identification of hematopoietic enhancers
We hypothesized that DNA connections formed from the active P1 promoter and the +24 enhancer may correspond to hematopoietic enhancers that regulate Runx1 expression. To identify putative enhancers, we aligned significantly interacting sites with the occupancy of thirteen TFs involved in hematopoietic progenitor cell production; enhancer histone modifications and DNase I hypersensitivity sites18-20; and conserved non-coding elements (CNE)11. We note that the +24 enhancer binds all thirteen hematopoietic progenitor TFs in HPC-7 cells (Fig. 3).
We found nine other potential enhancers within the Runx1 domain that form connections to the baits, each occupied by at least six hematopoietic TFs. These enhancers were named according to their distance from the P1 transcriptional start site: -371, -354, -327, -321, -303, -58, -48, and +110 (Fig. 3B). An additional interacting region at -368, a CNE, was located adjacent to a cluster of hematopoietic TFs (Fig. 3B). The putative enhancers either form distinct blocks upstream of Runx1-P1 (from -371 to -303 and -48 to -58), or fall between the P1 and P2 promoters (+24 and +110).
Eight out of nine putative enhancers (-371, -368, -354, -327, -321, -58, -48, and +110) form long-range interactions with the P1 promoter, the +24 enhancer, or both (Fig. 3B and Supplementary Table S1). The exception was -303, which does not interact with P1, but instead with the +24 and the P2 promoter. The +24 enhancer interacts promiscuously within the whole domain. Filtering at maximum stringency (see Methods) showed that the +24 enhancer connects to all putative enhancers and both Runx1 promoters (Fig. 3B). In contrast, the inactive P2 promoter connects only to -303 and +24 (Fig. 3B and Supplementary Table S1). Due to its proximity to P2, interactions between +110 and P2 cannot be resolved. A model of Runx1 interactions based on the 4C-seq results can be found in Supplementary Figure S3.
Long range chromatin interactions at hematopoietic enhancers
Long-range chromatin interactions can be mediated by cohesin and CTCF21, and cohesin is involved in transcription regulation at active genes22. We found that four of the nine enhancer loci (in addition to the +24 enhancer) coincide with Rad21 (cohesin) binding in the absence of CTCF (Fig. 3B and Supplementary Table S1). This is consistent with the idea that cohesin (but not necessarily CTCF) mediates local DNA-DNA interactions within TADs21. All Rad21 binding sites interacted with at least one ‘cc’ bait, therefore cohesin could mediate at least a subset of enhancer-promoter communication events in HPC-7 cells.
We compared the 4C interactions identified with recently published Capture Hi-C data in HPC-7 cells18 (Supplementary Fig. S4). Capture Hi-C data was only available for interactions anchored at P1, and has a lower coverage and resolution than our 4C-seq study (an average of ~18,000 reads per promoter for Capture Hi-C with a 6-cutter, and over 1 million reads per bait for 4C-seq with a 4-cutter). The Capture Hi-C study in HPC-7 cells identified 15 P1-interacting regions that were reproduced in our study (Supplementary Fig. S4). All of these are upstream of Runx1-P1, and most are within the -371 to -303 enhancer cluster (Supplementary Fig. S4). Therefore, our study provides additional Runx1-anchored interactions that were not previously described as connected to Runx1 promoters (enhancers -58, -48 and +110).
In vivo characterization of hematopoietic enhancers
Enhancer regions interacting with Runx1 recruit hematopoietic TFs in HPC-7 cells, therefore we determined if these regions act as enhancers in vivo. Each of the putative enhancers was tested for the ability to drive tissue-specific GFP expression in zebrafish embryos. Eight out of nine drove GFP expression specifically in the intermediate cell mass and posterior blood island, which are sites of hematopoietic progenitor cell production, at 20-24 hours post-fertilization (hpf) (Fig. 4A,B). The -303 enhancer also expressed GFP in keratinocytes, particularly after 24 hpf (Fig. 4C).
Interestingly, -303 was the only enhancer identified that interacted with the P2 promoter of Runx1, rather than P1. Despite the occupancy of multiple hematopoietic TFs in HPC-7 cells, and the presence of similar TF binding motifs compared to the other enhancers (Fig. 3 and Supplementary Table S1-2), we did not observe enhancer activity for -327.
In summary, we have assigned in vivo function to multiple putative enhancers upstream of Runx1 that were previously identified in silico18. These regions not only drive hematopoietic expression, but also physically connect with the Runx1-P1 promoter, lending confidence to the concept that they are bona fide regulators of Runx1 transcription.
Discussion
4C analysis in HPC-7 cells generated a high-resolution connectivity map of the genomic region harboring Runx1, and confirmed previously identified upstream connections from P1 to putative regulatory elements18. Runx1 appears to be contained within a ~1 Mb chromatin domain, consistent with Hi-C analyses in other cell types23,24. We observed multiple connections both up- and downstream from all 4C-seq baits (Runx1-P1, Runx1-P2 and +24 enhancer). Significantly, there were many more upstream connections anchored by the active elements (Runx1-P1 and +24).
Surprisingly, the +24 enhancer forms many up- and downstream connections outside of the Runx1 TAD. These comprise up to one-third of all connections formed and include other hematopoietic genes, such as Tiam1 downstream, and Erg upstream. Erg and Tiam1 dysregulation is associated with several tumor types, including AML and B- and T-cell lymphomas25,26. Strikingly, these genes are located over 2 Mb away from Runx1. These findings raise the interesting possibility that the +24 enhancer acts as a scaffold to recruit multiple promoters, enhancers and TFs over long distances in cis. Although the +24 enhancer is a specific marker of HSPCs, not all HSPCs that are marked by expression from this enhancer also express Runx127. Therefore, our results are consistent with the idea that the +24 enhancer regulates genes in addition to Runx1. Connections to +24 identified by our 4C-seq may point to the identity of some of these genes.
Our 4C-seq data identified multiple chromatin connections that intersect with some of the strongest indicators of upstream enhancers characterized by the binding of clusters of TFs and epigenetic modifications18,20. Eight out of nine of these DNA regions are able to drive hematopoietic expression in zebrafish, strongly suggesting an in vivo function. Two enhancers termed -59 and +110 were previously shown to drive LacZ expression in mice15; these are likely to be equivalent to the -58 and +110 enhancers identified here. Our data show in addition that these regions are in contact with P1 and the +24 enhancer.
The -303 enhancer interacts with P2 rather than P1 and, in addition to being active in hematopoietic sites, it drives expression in keratinocytes, indicating that it could act in a complex that keeps P2 silent; and/or that it regulates expression of Runx1 in other tissues. Interestingly, Runx1 is actively expressed in mouse keratinocytes where it is important for hair follicle development28. Neuronal cells also express Runx129, and there may be additional regulatory elements that control Runx1 expression in a manner distinct from hematopoietic expression. In support of this idea, we previously determined that cohesin and CTCF influence runx1 expression in hematopoietic, but not neuronal cells, in zebrafish30,31.
Cohesin and CTCF organize chromatin structure and when present in combination, they appear to negatively correlate with the HSPC TFs in HPC-7 cells18. However, we observed a coincidence of cohesin subunit Rad21 binding in the absence of CTCF with four out of nine identified enhancers, as well as the +24 enhancer. This suggests that CTCF-independent cohesin mediates a subset of enhancer-promoter looping in combination with TFs. This interpretation is consistent with previously identified CTCF-independent functions for cohesin in genome organization and transcription32. Importantly, cohesin mutations are prevalent in AML and other myeloid malignancies33, and are categorized together with RUNX1 and spliceosome mutations in a genetic category that confers poor prognosis in AML5. Cohesin mutations led to increased chromatin accessibility of Runx1 as measured by ATAC-seq34, raising the possibility that spatiotemporal regulation of Runx1 is cohesin-dependent in mouse and human, as was previously observed in zebrafish30.
4C-seq in HPC-7 cells has provided new high resolution connectivity data that sheds light on the genomic organization of Runx1, an important hematopoietic transcription factor. These data confirm and extend previous analyses, and furthermore, provide insight into the function of enhancers that have potential to regulate Runx1 expression. The data presented here set the scene for functional analyses to precisely determine how Runx1 is regulated, including CRISPR/Cas9-mediated interference with enhancer activity. They also provide a rationale for screening patients with myeloproliferative disorders for mutations in enhancer regions.
Methods
Cell culture
HPC-7 cells were maintained at a density of 1-10 x 105 cells/mL in Iscove’s modified Dulbecco’s media (Gibco®) supplemented with 3.024 g/L sodium bicarbonate, 10% fetal bovine serum (Moregate, New Zealand), 10% stem cell factor conditioned media and 0.15 mM monothiolglycerol (Sigma-Aldrich) as previously described35. SCF-conditioned media was obtained from culturing BHK-MKL cells maintained in Dulbecco’s modified Eagle media (Sigma-Aldrich) supplemented with 3.5 g/L glucose, 3.7 g/L Sodium Bicarbonate and 10% FBS.
4C-seq library preparation
4C library preparation was performed as previously described36 with modifications. Three libraries were generated, two replicates (passage 8 and passage 10 cells) and one control (a 1:1 mix of both replicates for which the first ligation step was omitted). Cells were cross-linked in 2% formaldehyde, 5% FBS and 1x PBS for 10 minutes at room temperature while rotating. Formaldehyde was quenched with a final concentration of 125 mM glycine for 5 minutes on ice while inverting several times. Cell pellets were washed twice with ice-cold 1x PBS.
Nuclei were harvested by lysing the cell pellets in ice-cold lysis buffer (10 mM Tris pH 8.0, 10 mM NaCl, 0.2% NP-40, protease inhibitors) for 10 minutes on ice. Nuclei were then resuspended in 1.2x DpnII restriction buffer (New England Biolabs) and 0.3% SDS and incubated for 1 hour at 37°C while shaking. Triton X-100 was then added to a final concentration of 1.8% and the reaction was left at 37°C while shaking for another hour. Chromatin was digested with 800 U of DpnII overnight at 37°C while shaking. DpnII was inactivated by adding SDS to a final concentration of 1.3% and incubating at 65°C for 20 minutes. Nuclei were diluted into a volume of 7 mL containing 1.01x T4 DNA ligase buffer (Life Technologies) and Triton-X100 at a final concentration of 1% and incubated at 37°C for 1 hour. Ligations were carried out with 100 U of T4 ligase (Life Technologies) for 4.5 hours at 16°C and 30 minutes at room temperature while shaking. For control libraries, ligase was omitted. Samples were proteinase K treated and reverse-crosslinked overnight at 65°C. Samples were then treated with RNase A at 37°C for 30 minutes. DNA was purified by phenol/chloroform extraction and ethanol precipitated.
A second digestion was performed with 25 U of BfaI (New England Biolabs) for P1 and P2 baits or MseI (New England Biolabs) for +24 baits in restriction buffer overnight at 37°C while shaking.
Restriction was inactivated by adding SDS to a final concentration of 1.3% and incubating at 65°C for 20 minutes. A second ligation was performed in the same way as the first ligation, except that ligations were incubated overnight. DNA was purified with two phenol/chloroform extractions and one chloroform extraction followed by ethanol precipitation. DNA concentrations were measured using a Qubit® 3.0 Fluorometer (Life Technologies) and Qubit® double-stranded DNA (dsDNA) High Sensitivity Assay kit (Life Technologies).
For each bait, a total of 1 μg of DNA was amplified by PCR using Q5® High-Fidelity DNA Polymerase (New England Biolabs). Bait primer sequences are listed in Supplementary Table S3. PCR products were purified using the QIAquick PCR Purification kit (QIAGEN). DNA concentrations were measured using a Qubit® 3.0 Fluorometer and average fragment size by 2100 Bioanalyser (Agilent Technologies) using a High Sensitivity DNA Kit (Agilent Technologies). Amplicons from the six different baits were mixed equally based on the concentration, average fragments size and ratio of demultiplexed 4C baits obtained from an initial MiSeq run. Libraries were prepared with Prep2SeqTM DNA Library Prep Kit from IlluminaTM (Affymetrix) and TruSeq® adaptors (Illumina). Libraries were mixed equimolarly and sequenced as 125 bp paired-end reads on two IlluminaTM HiSeq 2500 lanes by New Zealand Genomics Limited.
4C-seq data analysis
4C-seq data was analysed in command-line and the R statistical environment37 and visualized using the University of California, Santa Cruz (UCSC) genome browser (http://genome.ucsc.edu/) with mouse assemblies mm9 and mm1038,39 and using the R package ggplot240. Baits were demultiplexed based on bait primer sequences up to and including the digestion site using a custom awk script, allowing 0 mismatches. Only read pairs that had the forward and reverse bait sequences in the correct orientation were selected. Adapter sequences, bait sequences up to but excluding the digestion site, and bases with a Phred quality score under 20 were then trimmed from the reads, using the fastq-multx, fastq-mcf and cleanadaptors v1.24 tools41,42. Quality of reads was assessed using FastQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/)43.
Reads with a minimum length of 30 bp were mapped to the mm10 reference genome using Bowtie144, allowing 0 mismatches. Mapped reads were assigned to DpnII digestion fragments using fourSig45. The following reads were removed from the files: 1) self-ligated reads, 2) uncut reads (fragments adjacent to baits), and 3) reads at fragments that have at least 1 read in the control (non-ligated) library. The running mean was calculated from the sum of read counts from nine successive fragments, which was obtained using fourSig45, and was read per million normalized.
Significant interaction calling was performed using the R package fourSig with the following settings: window size of 3, 1000 iterations, fdr of 0.01, fdr.prob of 0.05 (which selects the top fifth percentile of interactions with a FDR of <0.01), and only included mappable fragments45. Significant interactions were called for two regions: 1) the whole of chr16, and 2) from chr16:92,250,000-93,635,000 (within the domain). Significant interactions in both replicates were overlapped using the bedIntersect tool from UCSC46.
In the fourSig package, significant interactions can be categorized into three categories: 1) interactions that are significant after the reads from the fragment with the highest read count is removed, 2) interactions that are significant when the fragment with the highest read count is averaged to the read counts of the neighbouring fragments, and 3) interactions that are significant only when all fragment read counts are included45. For this study, only category 1 and 2 interactions that overlap between both replicates were included, as they are more likely to represent true interactions (because they span multiple fragments), and were shown to be more reproducible between replicates than single-fragment interactions45. Furthermore, we distinguished category 1 interactions that overlap between both replicates from other category 1 and 2 interactions by coloring them red and orange, respectively, to visualise the most significant interactions. For conversion from assembly mm10 to mm9, the liftOver tool from UCSC was used (http://genome.ucsc.edu/)46. Gene annotations used in Figures are UCSC reference genes.
Zebrafish enhancer assay
Runx1 regulatory regions were amplified from HPC-7 gDNA or from I-SceI-zhsp70 plasmid containing the -368, +24 and +110 sequences11 (primer sequences are in Supplementary Table S3) and cloned into the zebrafish enhancer detection vector47. Primers amplified the TF binding peak +/-200 bp, except for -321. The -368, +24 and +110 fragments are 471, 529 and 579 bp, respectively11. 30 pg vector DNA and 120 pg Tol2 transposase mRNA48 was injected into 1-cell zebrafish embryos. Embryos were imaged at 20-24 or ~48 hpf using a Leica M205FA stereomicroscope with a DFC490 camera and LAS software (Leica Microsystems), images were processed using Adobe Photoshop. Zebrafish were maintained as described previously49 and zebrafish handling and procedures were carried out in accordance with the Otago Zebrafish Facility
Standard Operating Procedures. The University of Otago Animal Ethics Committee approved all zebrafish research under approval AEC 48/11.
Identification of conserved non-coding elements
Mouse conserved non-coding elements (mCNEs) were identified as described previously11.
ChIP-seq, Capture Hi-C and DNase I hypersensitivity data
Occupancy of the transcription factors Erg, Fli1, Scl, Runx1, Gata2, E2A, Ldb1, Lyl1, Lmo2, Gfi1b, Meis1, Myb, phospho-Stat1, Pu.1, Stat3, Eto2, Cebp-a, Cebp-β, Elf1, Nfe2, p53, cMyc, Egr1, E2f4, cFos, Mac and Jun; Rad21 and CTCF; H3K27ac and H3K4me3; DNase I hypersensitivity sites; and Capture Hi-C data in HPC-7 cells was obtained from previously published data18-20. Rad21, Smc3 and CTCF chromatin immunoprecipitation sequencing (ChIP-seq) data in MEL and CH12 cells were obtained from ENCODE50.
Authorship contributions
JM, AT, MO, JMO, JAH, designed experiments; JM, AT, performed experiments; JM, AT, MO, JMO, JAH analyzed data; JM and JAH wrote the paper with input from the other authors.
Additional information
Availability of data
The 4C-seq dataset is accessible through GEO Series accession number GSE86994 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE86994).
Disclosure of conflicts of interest
The authors have no conflicts of interest to disclose.
Acknowledgements
HPC-7 and BHK/MKL cells were kindly provided by Kathy Knezevic and John Pimanda. TruSeq® adaptors were kindly provided by Ian Morison. The authors would like to thank Anita Dunbier and Sofie Van Huffel for help with development of 4C protocols, and Noel Jhinku for expert management of the Otago Zebrafish Facility. This research was funded by Royal Society of NZ Marsden Fund [grant number 11-UOO-027 to JAH] and the Health Research Council of NZ [grant number 15/229 to JAH].
Footnotes
E-mail addresses of other authors: judith.marsman{at}otago.ac.nz (Judith Marsman), amarni.l.thomas{at}gmail.com (Amarni Thomas), csismo{at}nus.edu.sg (Motomi Osato) and justin.osullivan{at}auckland.ac.nz (Justin O’Sullivan).