Introduction

Tuberculosis (TB) continues to represent a major threat to global public health, with the World Health Organization (WHO) estimating 10 million cases and 1.4 million deaths in 2019 alone [1]. Multidrug resistance (MDR) poses a major challenge to tackling TB; it is estimated that there were 465,000 cases of rifampicin-resistant TB in 2019, of which 78% were resistant to the first-line drugs rifampicin and isoniazid—called MDR-TB [1]. While treatment is 85% successful overall, that drops to 57% for rifampicin-resistant and MDR-TB [1]; underdiagnosis and treatment failures then amplify the problem by encouraging onward transmission of MDR-TB [2]. New treatment regimens for MDR-TB are therefore an important focus, introducing new and repurposed drugs such as bedaquiline, clofazimine, delamanid, and linezolid [3,4]; however, resistance is already emerging [5–7].

Understanding mechanisms of resistance in TB is important for developing rapid susceptibility tests that improve individual patient treatment, recommending drug regimens that reduce the development of MDR and developing new and improved drugs that expand treatment options [8,9]. Genomics can accelerate drug susceptibility testing, replacing slower culture-based methods by predicting resistance from the sequenced genome rather than directly phenotyping the bacteria [10]. Genome sequencing-based susceptibility testing for first-line drugs has achieved sensitivities of 91.3% to 97.5% and specificities of 93.6% to 99.0% [11], surpassing the thresholds for clinical accreditation, motivating its adoption by multiple public health authorities [12]. In low-resource settings, molecular tests such as Cepheid GeneXpert and other line probe assays offer rapid and more economical susceptibility testing by genotyping a panel of known resistance-conferring genetic variants [13], with performance close to that achieved by whole-genome sequencing [14,15]. However, the limited number of resistance-conferring mutations that can be included in such tests can lead to missed MDR diagnoses and incorrect treatment [11,16]. Both approaches rely on the development and maintenance of resistance catalogues of genetic variants [11,17].

In the discovery of resistance-conferring variants, traditional molecular approaches have been replaced by high-throughput, large-scale whole-genome sequencing studies of hundreds to thousands of resistant and susceptible clinical isolates [18–23]. Despite the strong performance of genome-based resistance prediction for first-line drugs, knowledge gaps remain, especially for second-line drugs [17,24,25]. There are numerous challenges in the pursuit of previously uncatalogued resistance mechanisms. Very large sample sizes are needed to identify rarer resistance mechanisms with confidence. The lack of recombination in Mycobacterium tuberculosis makes it difficult to pinpoint resistance variants unless they arise on multiple genetic backgrounds, reiterating the need for large sample sizes. Sophisticated analyses are required that attempt to disentangle genetic causation from correlation [26]. A reliance on a binary resistance/sensitivity classification paradigm has hindered reproducibility for some drugs by failing to mirror the continuous nature of resistance [27–29].

The aim of Comprehensive Resistance Prediction for Tuberculosis: An International Consortium (CRyPTIC) was to address these challenges by assembling a global collection of over 10,000 M. tuberculosis isolates from 27 countries followed by whole-genome sequencing and semiquantitative determination of minimum inhibitory concentration (MIC) to 13 first- and second-line drugs using a bespoke 96-well broth micodilution plate assay. The development of novel, inexpensive, high-throughput drug susceptibility testing assays allowed us to conduct the project at scale, while investigating MIC on a grid of 2-fold concentration dilutions [30,31]. Here, we report the identification of previously uncatalogued resistance-conferring variants through 13 genome-wide association studies (GWAS) investigating MIC values in 10,228 M. tuberculosis isolates. We identify these discoveries relative to specific catalogues for the sake of concreteness and tractability, while acknowledging the catalogues do not include all credible resistance mechanisms known from the literature; we expand on this limitation in the Discussion. Our analyses employed a linear mixed model (LMM) to identify putative causal variants while controlling for confounding and genome-wide linkage disequilibrium (LD) [20,32]. We developed a novel approach to testing associations at both 10,510,261 oligopeptides (11-mers) and 5,530,210 oligonucleotides (31-mers) to detect relevant genetic variation in both coding and non-coding sequences and to avoid a reference-based mapping approach that can inadvertently miss significant variation. We report previously uncatalogued variants associated with MIC for all 13 drugs, focusing on variants in the 20 most significant genes per drug. We highlight notable discoveries for each drug and demonstrate the ability of large-scale studies to improve our knowledge of genetic variants associated with antimicrobial resistance in M. tuberculosis.

Results

CRyPTIC collected isolates from 27 countries worldwide, oversampling for drug resistance [31]. A total of 10,228 genomes were included in total across the GWAS analyses: 533 were lineage 1; 3,581 were lineage 2; 805 were lineage 3; and 5,309 were lineage 4. Due to rigorous quality control, we dropped samples for each drug as detailed in the methods, resulting in a range of 6,388 to 9,418 genomes used in each GWAS, for which we constructed a phylogeny (Fig 1A). MICs were determined on a grid of 2-fold concentration dilutions for 13 antimicrobials using quantitative microtiter plate assays: first-line drugs: ethambutol, isoniazid, and rifampicin; second-line drugs: amikacin, ethionamide, kanamycin, levofloxacin, moxifloxacin, and rifabutin; and the new and repurposed drugs: bedaquiline, clofazimine, delamanid, and linezolid. The phenotype distributions differed between the drugs, with low numbers of sampled resistant isolates for the new and repurposed drugs that have not yet been widely used in TB treatment (Figs 1B and S1). Applying epidemiological cutoffs (ECOFFs) to the MIC [31], the GWAS featured 66 isolates resistant to bedaquiline, 97 resistant to clofazimine, 77 resistant to delamanid, and 67 resistant to linezolid. We performed oligopeptide- and oligonucleotide-based GWAS analyses, controlling for population structure using LMMs. We focused initially on oligopeptides, interpreting oligonucleotides only where necessary for clarifying results.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

(A) Phylogeny of 10,228 isolates sampled globally by CRyPTIC used in the GWAS analyses. Lineages are coloured blue (lineage 1), green (2), orange (3), and yellow (4). Branch lengths have been square root transformed to visualise the detail at the tips. (B) Distributions of the log₂ MIC measurements for all 13 drugs in the GWAS analyses, AMI, BDQ, CFZ, DLM, EMB, ETH, INH, KAN, LEV, LZD, MXF, RFB, and RIF. The red line indicates the ECOFF breakpoint for binary resistance versus sensitivity calls [31]. AMI, amikacin; BDQ, bedaquiline; CFZ, clofazimine; DLM, delamanid; ECOFF, epidemiological cutoff; EMB, ethambutol; ETH, ethionamide; INH, isoniazid; KAN, kanamycin; LEV, levofloxacin; LZD, linezolid; MIC, minimum inhibitory concentration; MXF, moxifloxacin; RFB, rifabutin; RIF, rifampicin.

https://doi.org/10.1371/journal.pbio.3001755.g001

Estimates of sample heritability (variance in the phenotype explained by additive genetic effects) were higher for MIC compared to binary resistant versus sensitive phenotypes for the new and repurposed drugs bedaquiline, clofazimine, delamanid, and linezolid by at least 26%. Across drugs, binary heritability ranged from 0% to 94.7% and MIC heritability from 36.0% to 95.6%, focusing on oligopeptides (Figs 2 and S2 and S1 Table). For delamanid, binary heritability was not significantly different from zero (2.99 × 10⁻⁶; 95% confidence interval (CI) 0.0% to 0.5%), while MIC heritability was 36.0% (95% CI 28.9% to 43.1%). Estimates of sample heritability were more similar between binary and MIC phenotypes for the remaining drugs, differing by −3.6% to +5.2%.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Sample heritability for MIC (orange) versus binary resistance/sensitivity (blue) assuming additive genetic variation in oligopeptide presence/absence across 13 drugs, DLM, CFZ, LZD, BDQ, MXF, LEV, KAN, EMB, ETH, AMI, RIF, INH, and RFB.

Lines depict 95% CIs. MIC heritability was at least 26% higher than binary heritability for the new and repurposed drugs BDQ, CFZ, DLM, and LZD. AMI, amikacin; BDQ, bedaquiline; CFZ, clofazimine; CI, confidence interval; DLM, delamanid; EMB, ethambutol; ETH, ethionamide; INH, isoniazid; KAN, kanamycin; LEV, levofloxacin; LZD, linezolid; MIC, minimum inhibitory concentration; MXF, moxifloxacin; RFB, rifabutin; RIF, rifampicin.

https://doi.org/10.1371/journal.pbio.3001755.g002

GWAS identified oligopeptide variants associated with changes in MIC for all 13 drugs after controlling for population structure (Table 1 and Figs 3, S3 and S4). In total, across the drugs, we tested for associations at 10,510,261 variably present oligopeptides and 5,530,210 oligonucleotides; these captured substitutions, insertions, and deletions. The drugs differed in the number of genes or intergenic regions that were significant, the drugs with fewest significant genes being isoniazid (12), levofloxacin (13), and moxifloxacin (6). We defined the significance of a gene or intergenic region by the most significant oligopeptide within it and assessed all significant variants above a 0.1% minor allele frequency (MAF) threshold for the top 20 significant genes. The top 20 genes for each drug are detailed in Table 1. Some variants were identified in novel genes, some were novel variants in known genes, and some were known variants. We highlight examples of these (in reverse order) in the following sections. Highlighted examples have been chosen to exclude genes or variants in LD with other regions where possible; some are in LD with other less significant variants.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Manhattan plots of regions containing oligopeptide variants associated with MIC across 13 drugs.

Significant oligopeptides are coloured by the direction (orange = increase, blue = decrease) and magnitude of their effect size on MIC, estimated by LMM [32]. Bonferroni-corrected significance thresholds are shown by the black dashed lines. The top 20 genes ranked by their most significant oligopeptides are annotated alphabetically. Gene names separated by colons indicate intergenic regions. Gene names for those annotated with letters can be found in Table 1. Oligopeptides were aligned to the H37Rv reference; unaligned oligopeptides are plotted to the right in light grey. LMM, linear mixed model; MIC, minimum inhibitory concentration.

https://doi.org/10.1371/journal.pbio.3001755.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The top genes or intergenic regions ranked by their most significant oligopeptides per drug, up to a maximum of 20 (more only when the 20th was tied).

Genes are highlighted in bold if they were catalogued for that drug by [11,17]. Gene names separated by colons indicate intergenic regions. Genes or intergenic regions capturing repeat regions are highlighted with the superscript ^R. Alphabetic characters following gene names are used to cross-reference with the corresponding Manhattan plots in Fig 3.

https://doi.org/10.1371/journal.pbio.3001755.t001

We assessed whether the top genes for each drug were in either of 2 previously described resistance catalogues [11,17]; we describe variants in genes not in these catalogues as uncatalogued (Table 1). The interpretation of oligopeptides and oligonucleotides required manual curation to determine the underlying variants they tagged; the most significant oligopeptide or oligonucleotide for each allele captured by the significant signals are described in S1 Data. For 8/13 drugs with previously catalogued resistance determinants, the most significant GWAS signal in CRyPTIC was a previously catalogued variant, consistent with previous GWAS in M. tuberculosis [18–23]. The most significant catalogued variants for each drug were (lowercase for nucleotides, uppercase for amino acids): rrs a1401g (amikacin, kanamycin), embB M306V (ethambutol), fabG1 c−15t (ethionamide), katG S315T (isoniazid), gyrA D94G (levofloxacin, moxifloxacin), and rpoB S450L (rifampicin) [11,17]. For the remaining drugs, which had no resistance determinants in the catalogues to which we referred [11,17], the genes identified by the top signals were: Rv0678 (bedaquiline, clofazimine), ddn (delamanid), rplC (linezolid), and rpoB (rifabutin). However, for all these associations, there does exist credible evidence elsewhere in the literation (e.g., [85,86]). The top variants identified for each drug were all significant at p < 1.04 × 10⁻¹⁵.

For many drugs, the most significant oligopeptide was high frequency, and the direction of effect was to decrease MIC relative to alternative alleles (S5 Fig). This implies that oligopeptides and oligonucleotides associated with lower MIC are more likely to be genetically identical across strains than those associated with higher MIC. This would be consistent with the independent evolution of increased MIC from a shared, low-MIC TB ancestor. Often there were multiple low-frequency oligopeptides mapping to the same positions, supporting this idea.

Uncatalogued variants significantly associated with MIC are important because they could improve resistance prediction and shed light on underlying resistance mechanisms; they may be novel or previously implicated in resistance but not to a standard of evidence sufficient to be catalogued. We discuss the choice of catalogues in the Discussion [11,17].

We next looked at uncatalogued variants in known resistance-conferring genes. We identified uncatalogued variants in gyrB associated with levofloxacin and moxifloxacin MIC (minimum p-value levofloxacin: p < 10^−15.6, moxifloxacin: p < 10^−11.6). The primary mechanisms of resistance to the fluoroquinolones levofloxacin and moxifloxacin are mutations in gyrA or gyrB, the subunits of DNA gyrase. The gyrB Manhattan plots for levofloxacin and moxifloxacin both contained 2 adjacent peaks within the gene, but for each drug just 1 of the 2 peaks was significant, and these differed between the drugs (Fig 4). Interpretation of oligopeptides and oligonucleotides requires an understanding of the variants that they capture, which we visualised by aligning them to H37Rv and interpreting the variable sites (e.g., Fig 4C and 4D). For levofloxacin, the peak centred around amino acid 461. Significant oligopeptides captured amino acids 461 and 457, which are both uncatalogued [11,17] with 457 falling just outside of the gyrB quinolone resistance-determining region (QRDR-B) [33]. Oligopeptides capturing 461N were associated with increased MIC (e.g., NSAGGSAKSGR, −log₁₀p = 15.65, effect size β = 2.46, present in 15/7,300 genomes). Oligopeptides capturing the reference alleles at codons 461 and 457 were significantly associated with lower MIC (e.g., 461D: DSAGGSAKSGR, −log₁₀p = 13.47, β = −2.14, present in 7,278/7,300 genomes; 457V/461D: SELYVVEGDSA, −log₁₀p = 12.51, β = −1.96, present in 7,272/7,300 genomes). For moxifloxacin, the peak centred around amino acid 501. Significant oligopeptides captured amino acids 499 and 501. Oligopeptides capturing 501D were associated with increased MIC (e.g., NTDVQAIITAL, −log₁₀p = 10.64, β = 1.86, present in 23/6,388 genomes). Oligopeptides capturing the reference allele at codons 499 and 501 were associated with lower MIC (e.g., NTEVQAIITAL, −log₁₀p = 11.63, β = −1.33, present in 6,332/6,388 genomes). Amino acids 461 and 501 are at the interface between gyrB and the bound fluroquinolone [34]. gyrB is included in the reference catalogues for predicting levofloxacin (including D461N) but not moxifloxacin resistance; therefore, our results support inclusion of gyrB (in particular E501D) in future moxifloxacin catalogues [11,17].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Interpreting significant oligopeptide variants for levofloxacin and moxifloxacin MIC in gyrB.

Oligopeptide Manhattan plots are shown for (A) levofloxacin and (B) moxifloxacin. Oligopeptides are coloured by the reading frame that they align to, black for in frame and grey for out of frame in gyrB. Oligopeptides aligned to the region by nucmer but not realigned by BLAST are shown in grey on the right-hand side of the plots. The black dashed lines indicate the Bonferroni-corrected significance thresholds—all oligopeptides above the line are genome-wide significant. Alignment is shown of oligopeptides significantly associated with (C) levofloxacin and (D) moxifloxacin. The H37Rv reference codons are shown at the bottom of the figure, grey for an invariant site, coloured at variant site positions. The background colour of the oligopeptides represents the direction of the β estimate, light grey when β < 0 (associated with lower MIC), dark grey when β> 0 (associated with higher MIC). Oligopeptides are coloured by their amino acid residue at variant positions only. MIC, minimum inhibitory concentration.

https://doi.org/10.1371/journal.pbio.3001755.g004

Next, we looked at specific examples of significant associations identified by GWAS in genes not catalogued by [11,17] for each of the drugs. A well-recognized challenge in GWAS for antimicrobial resistance is the presence of artefactual cross-resistance. To mitigate this risk, we preferentially highlight variants significantly associated with a single drug. However, many catalogued resistance variants demonstrated artefactual cross-resistance. For example, variants in the rifampicin resistance-determining region were in the top 20 significant associations for all drugs except for delamanid (Table 1). Interestingly, we observed that the magnitude of effect sizes was often larger on MIC of the drug to which catalogued variants truly confer resistance (S6 Fig). For example, the effect sizes for significant oligopeptides in rpoB were greater for rifampicin and rifabutin than for all other drugs. This suggests that the β estimates could help to prioritise drugs for follow up when genes are significantly associated with multiple drugs.

Discussion

In this study, we tested oligopeptides and oligonucleotides for association with semiquantitative MIC measurements for 13 antimicrobials to identify novel resistance determinants. Analysing MIC rather than binary resistance phenotypes enabled identification of variants that cause subtle changes in MIC. This is important, on the one hand, because higher rifampicin and isoniazid MIC in sensitive isolates are associated with increased risk of relapse after treatment [68]. Conversely, low-level resistance among isolates resistant to rifampicin and isoniazid mediated by particular mutations may sometimes be overcome by increasing the drug dose, or replacing rifampicin with rifabutin, rather than changing to less desirable drugs with worse side effects [69–73]. The investigation of MIC was particularly effective at increasing sample heritability for the new and repurposed drugs.

The MICs were positively correlated between many drugs, particularly among first-line drugs (see Figure 4A of [78]). Consequently, many of the 10,228 isolates we studied were MDR and XDR. In GWAS, this generates artefactual cross-resistance, in which variants that cause resistance to one drug appear associated with other drugs to which they do not confer resistance. In practice, it is difficult to distinguish between associations that are causal versus artefactual without experimental evidence. Nevertheless, we found frequent evidence of artefactual cross-resistance: several genes and intergenic regions featured among the top 20 strongest signals of association to multiple drugs, including rpoB (12 drugs), embB (7), fabG1 (7), rrs (6), gyrA (6), katG (6), lprF:Rv1371 (6), pncA (5), ethA (4), Rv3183:Rv3188 (4) dxs2:Rv3382c (4), rpsL (3), and moaC3:Rv3327 (3). Among previously catalogued variants, we observed that the estimated effect sizes were usually larger in magnitude for significant true associations than significant artefactual associations (S4 Fig). In future GWAS, this relationship could help tease apart true versus artefactual associations when an uncatalogued variant is associated with multiple drugs.

We focused on variants in the top 20 most significant genes identified by GWAS for each of the 13 drugs, classifying significant oligopeptides and oligonucleotides according to whether the variants they tagged were previously catalogued among known resistance determinants or not. While the interpretation of oligopeptides and oligonucleotides required manual curation to determine the underlying variants they tagged, the approach had the advantage of avoiding reference-based variant calling that can miss important signals, particularly at difficult-to-map regions. Exhausitvely and concisely calling variants relative to a reference genome in a large dataset of over 10,000 genomes is not trivial [80]. An additional benefit of analysing oligopeptides (versus oligonucleotides) is to pool signals of association across nucleotides that encode the same amino acid. Since we expected most resistance-conferring mutations to affect coding sequences, we anticipated that this should improve power and interpretability for most associations. A combination of strong linkage disequilbrium and lower diversity at the protein versus nucleotide level meant that, despite testing all 6 reading frames, the Bonferroni thresholds for the oligopeptide analyses were 2-fold less stringent than for the oligonucleotide analyses.

For 8/13 drugs with previously catalogued resistance determinants, the most significant GWAS signal in CRyPTIC was a previously catalogued variant. Among the uncatalogued variants there are promising signals of association, including in the Rv1218c promoter binding site of the transcriptional repressor Rv1219c (associated with MIC for isoniazid) upstream of the vapBC20 operon that cleaves 23S rRNA (linezolid) and in the region encoding a helix lining the active site of cyp142 (clofazimine). These variants would benefit from further investigation via replication studies in independent populations, experimental exploration of proposed resistance mechanisms, or both.

We elected to classify significant variants as catalogued versus uncatalogued, rather than known versus novel, for several reasons. The catalogues represent a concrete, preexisting knowledgebase collated by expert groups for use in a clinical context [11,17]. We chose [11,17] as they were the most recent and up-to-date catalogues available for the drugs we investigated. The inclusion criteria for variants to be considered catalogued are therefore stringent; it follows that a class of variants exist that have been reported in the literature but not assimilated into the catalogues [11,17]. The literature is vast and heterogenous, with evidence originating from molecular, clinical, and GWAS. Inevitably, some uncatalogued variants in the literature will be false positives, while others will be real but did not meet the standard of evidence or clinical relevance for cataloguing. Evidence from CRyPTIC that supports uncatalogued variants in the latter group is of equal or greater value than the discovery of completely novel variants, because it contributes to a body of independent data supporting their involvement. This seems to apply to the most significant signals of association for BDQ, CLO, DLM, LZD, and RFB (Table 1), none of which appeared in our reference catalogues, but all of which appear in credible reports in the literature (e.g., [85,86]). To take another example, gyrB did not appear in the catalogues we used for moxifloxacin [11,17]. Yet our rediscovery of gyrB 501D complements published reports associating the substitution with moxifloxacin resistance [74–76], strongly enhancing the evidence in favour of inclusion in future catalogues. Indeed, the recent WHO prediction catalogue, published after the completion of this study and which draws on the CRyPTIC data analysed here includes the E501D resistance-associated variant [77]. Moreover, of the 5 new genes added to the forthcoming WHO catalogue [77] but not featuring in the catalogues [11,17] used here—eis (amikacin), ethA (ethionamide), inhA (ethionamide), rplC (linezolid), gyrB (moxifloxacin)—we identify all as containing significant variants by GWAS except one, eis (amikacin).

The combination of a very large dataset exceeding 10,000 isolates and quantification of resistance via MIC enabled the CRyPTIC study to attribute a large proportion of fine-grained variability in antimicrobial resistance in M. tuberculosis to genetic variation. Compared to a parallel analysis of binary resistance phenotypes in the same samples, we observed an increase in sample heritability of 26.1% to 37.1% for the new and repurposed drugs bedaquiline, clofazimine, delamanid, and linezolid. The improvement was most striking for delamanid, whose heritability was not significantly different to zero for the binary resistance phenotype. In the case of delamanid, the MIC analysis detected a surprisingly large number of signals. Since few isolates were strongly delaminid resistant, this indicates we were picking apart fine-grained differences in MIC, a phenotype which may be more polygenic than binary resistance/sensitivity. In contrast, the scope for improvement was marginal for the better-studied drugs isoniazid and rifampicin, where MIC heritabilities of 94.6% to 94.9% were achieved. This demonstrates the ability of additive genetic variation to explain almost all the phenotypic variability in MIC for these drugs. Nevertheless, we were still able to find uncatalogued hits for these drugs. The very large sample size also contributed to increased sample heritability compared to previous pioneering studies. Compared to Farhat and colleagues [22] who estimated the heritability of MIC phenotypes in 1,452 isolates, we observed increases in heritability of 2.0% (kanamycin), 3.3% (amikacin), 14.0% (isoniazid), 10.8% (rifampicin), 11.2% (ethambutol), and 19.4% (moxifloxacin), although these sample heritabilities depend on the idiosyncrasies of sampling. Furthermore, many of the uncatalogued signals we report here as significant detected rare variants at below 1% MAF, underlining the ability of very large-scale studies to improve our understanding of antimicrobial resistance not only quantitatively, but to tap otherwise unseen rare variants that reveal new candidate resistance mechanisms.

Materials and methods

Author contributions

Conceptualisation: Camilla Rodrigues, David Moore, Derrick W. Crook, Daniela M. Cirillo, Philip W Fowler, Zamin Iqbal, Nazir A. Ismail, Nerges Mistry, Stefan Niemann, Tim E.A. Peto, Guy Thwaites, A. Sarah Walker, Timothy M Walker, Daniel J. Wilson

Methodology: Sarah G. Earle, Daniel J. Wilson, Clara Grazian, A Sarah Walker

Software: Sarah G. Earle, Daniel J. Wilson

Formal analysis: Sarah G. Earle, Daniel J. Wilson, Clara Grazian, A Sarah Walker

Investigation: The CRyPTIC Consortium

Resources: The CRyPTIC Consortium

Data curation: Martin Hunt, Jeff Knaggs, Zamin Iqbal, Philip W Fowler, Zamin Iqbal

Writing – original draft preparation: Sarah G. Earle, Daniel J. Wilson, A Sarah Walker, Philip W Fowler

Writing – review & editing: The CRyPTIC Consortium

Visualization: Sarah G. Earle

Supervision: Daniela M Cirillo, Derrick W. Crook, Tim E.A. Peto, Daniel J. Wilson, A Sarah Walker, Zamin Iqbal, Philip W Fowler

Project administration: The CRyPTIC Consortium;

Funding acquisition: Camilla Rodrigues, David Moore, Derrick W. Crook, , Daniela M. Cirillo, Zamin Iqbal, Nazir A. Ismail, Nerges Mistry, Stefan Niemann, Tim E.A. Peto, Guy Thwaites, A. Sarah Walker, Timothy M Walker, Daniel J. Wilson

Supporting information

Acknowledgments

We thank Faisal Masood Khanzada and Alamdar Hussain Rizvi (NTRL, Islamabad, Pakistan), Angela Starks and James Posey (Centers for Disease Control and Prevention, Atlanta, USA), Juan Carlos Toro and Solomon Ghebremichael (Public Health Agency of Sweden, Solna, Sweden), and Iñaki Comas and Álvaro Chiner-Oms (Instituto de Biología Integrativa de Sistemas, Valencia, Spain; CIBER en Epidemiología y Salud Pública, Valencia, Spain; Instituto de Biomedicina de Valencia, IBV-CSIC, Valencia, Spain). Computation used the Oxford Biomedical Research Computing (BMRC) facility, a joint development between the Wellcome Centre for Human Genetics and the Big Data Institute supported by Health Data Research UK and the NIHR Oxford Biomedical Research Centre.