Tumor necrosis factor SNP haplotypes are associated with iron deficiency anemia in West African children

Anemia (hemoglobin < 110 g/L) is one of the commonest and most intractable public health problems of children in malaria-endemic countries.¹  In The Gambia, three-quarters of children are anemic.²  However, iron deficiency anemia (IDA) is difficult to distinguish from the anemia of chronic disease (ACD) in many malaria-endemic areas.³  There is compelling evidence that plasma levels of tumor necrosis factor-α (TNF-α) are significantly raised in malaria infection compared with other illnesses,⁴  even in children with asymptomatic infection.⁵ 

Studies show that TNF, mapped on 6p21.3 in the major histocompatibility complex (MHC), is involved in the regulation of iron metabolism. TNF-α induces hypoferremia by inhibiting iron release from macrophage storage compartments⁶  and increasing the transcriptional induction of ferritin.⁷  TNF-α also inhibits iron uptake in erythroid precursors,⁸  blocks the differentiation and proliferation of erythroid progenitor cells⁹  and causes erythrophagocytosis and dyserythropoiesis.^10,11  Recent in vitro and animal studies have demonstrated that TNF-α also directly inhibits intestinal iron absorption¹²  independently of hepcidin production.^13,14 

There are significant interindividual variations in TNF-α production¹⁵  and family studies suggest that up to 60% of the variability in TNF-α levels may be genetically determined.¹⁶  Functional studies indicate that the TNF₋₃₀₈ promoter polymorphism (TNF2)¹⁷  increases constitutive and inducible gene transcription,^18-20  although data linking the TNF₋₂₃₈ promoter polymorphism (TNFA)²¹  to altered gene transcription is less clear.^22,23  A small number of clinical studies indicate a link between TNF polymorphisms and iron status and anemia. Interestingly, the TNF₋₃₀₈ and the TNF₋₂₃₈ promoter polymorphisms were found to alter the phenotypic expression of hereditary hemochromatosis.^24,25  The TNF₋₃₀₈ polymorphism was associated with an increased risk of severe anemia in low-birth-weight children in Kenya,²⁶  although not severe malaria anemia.^26,27  The TNF₋₂₃₈ polymorphism was associated with an increased risk of severe malarial anemia in Gambian children.²⁷ 

The aim of this study was to determine whether variation in the TNF gene might be a risk factor for nutritional iron deficiency, IDA, or ACD in children in malaria-endemic areas. To achieve this aim we investigated putative functional single nucleotide polymorphisms (SNPs) and haplotypes across the TNF gene locus in relation to iron status and anemia in a cohort of 780 Gambian children across a malaria season.

A cohort of 780 children aged 2 to 6 years was recruited from 10 rural villages in the West Kiang region of The Gambia at the start of the malaria season with follow-up to the end of the malaria season, as described previously.²⁸  In brief, all children were examined by the study clinician and had a blood sample collected for full blood count, malaria slide, iron status assays, and α₁-antichymotrypsin (a marker of inflammation) at baseline (the start of the malaria season) and at the end of the malaria season. Genomic DNA was extracted from venous blood in ethylenediaminetetraacetic acid (EDTA) using standard methods²⁹  and was available for 756 children. Children with pyrexia (temperature > 37.5°C) had appropriate investigations and clinical treatment, and a blood sample was taken 2 weeks later after recovery from illness. All children were treated with a 3-day course of mebendazole at the start of the study for possible hookworm and other helminth infections. Nine villages (700 children) were of the Mandinka ethnic group and one village (80 children) was of the Fulani ethnic group.

Malaria incidence is highly seasonal in The Gambia, with the majority of malaria cases occurring between September and December.³⁰  Children were followed across the malaria season to control for multiple individual factors that may influence iron status and anemia. The study was approved by The Gambia Government and the MRC Ethics Committee, and written informed consent was obtained from children's parents or guardians in accordance with the Declaration of Helsinki before recruitment.

We used biochemical markers of iron status and hemoglobin levels to classify, according to published case definitions for iron deficiency, IDA, ACD, and ACD with coexisting iron deficiency. Iron deficiency was defined as a plasma ferritin concentration less than 12 μg/L.³¹  IDA was defined as a hemoglobin concentration less than 110 g/L, plasma ferritin concentration less than 12 μg/L, and a zinc protoporyphyrin level of more than 80 μmol/mol hemoglobin, as previously validated in a malaria-endemic population.³²  Children were defined as iron replete if they had hemoglobin levels more than 110 g/L, ferritin more than 12 μg/L, α₁-antichymotrypsin less than 0.6 g/L, no fever, and a blood slide negative for malaria parasites. We took a ferritin-based approach because ferritin metabolism, unlike soluble transferrin receptor, is not affected by clinically silent hemoglobinopathies such as HbAS or α-thalassemia.³³  However, plasma ferritin levels may be falsely raised by concomitant inflammation underestimating true levels of iron deficiency. To account for this we included inflammatory markers and malaria status in all regression-based analyses and also defined children with ACD with and without iron deficiency as previously described.³⁴ 

ACD without iron deficiency was defined as a hemoglobin concentration less than 110 g/L, ferritin concentration more than 30 μg/L, transferrin saturation less than 16% and a soluble transferrin receptor to log ferritin ratio of less than 1. ACD with iron deficiency was defined as above except with a soluble transferrin receptor to log ferritin ratio more than 2.³⁴  However, using these definitions we found only 7 children had ACD and iron deficiency and 29 had ACD without iron deficiency at the end of the malaria season. We thus defined ACD as above, but did not exclude children on the basis of their soluble transferrin receptor to log ferritin ratio.

Hemoglobin level was measured by the Medonic CA 530 Oden 16 Parameter System Hemoglobinometer and zinc protoporphyrin (ZnPP) level by the Aviv Biomedical Hematoflurometer within 24 hours of collection. Blood films were stained with Giemsa and examined for malaria parasites. Frozen plasma samples were analyzed for indicators of iron status at MRC Human Nutrition Research Laboratories (Cambridge, United Kingdom). We measured ferritin concentration by the IMx Ferritin assay, a Microparticle Enzyme Immunoassay (MEIA; Abbott Laboratories, Abbott Park, IL) and soluble transferrin receptor (sTfR) concentration by enzyme-linked immunosorbent assay (ELISA) using a commercially available kit (R&D Systems, Minneapolis, MN). Serum iron and unsaturated iron binding capacity (UIBC) were measured by Ferrozine-based photometry and colorimetry using an automated analyzer (Hitachi 911; Hitachi, Tokyo, Japan). Transferrin saturation (TS) was calculated from plasma iron and UIBC (TS = [plasma iron/(UIBC + plasma iron)] × 100). We also measured α₁-antichymotrypsin, a measure of the inflammatory response, by immunoturbidimetry (Cobas Mira Plus Bioanalyser) to aid the interpretation of the markers of iron status.

Genotypes were determined on whole-genome amplified DNA (primer extension preamplification)³⁵  by the SEQUENOM system using matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectrometry.³⁶  Primer sequences are available on request. The most informative haplotype-tagging SNPs (htSNPs) to type in Gambian subjects were identified by analyzing the pattern of linkage disequilibrium (LD) in the MHC class III genomic region including TNF and its immediate neighbors IκBL and LTA using 32 Gambian family trios in a previous study.³⁷  The PHASE program, version 1³⁸  (http://stephenslab.uchicago.edu/software.html) was used to infer haplotypes from the genotypes of the study population and estimate the frequency of each inferred haplotype. The entropy maximization method was used to identify htSNPs that described >90% of the observed haplotypic diversity in this gene region³⁷  (Figure 1). Pairwise LD statistics were estimated using the program HaploXT (http://www.sph.umich.edu/csg/abecasis/GOLD/docs/haploxt.html).

Analyses were conducted using STATA version 8.0 (Stata, Timberlake, London, United Kingdom). Pearson χ² test was used to assess associations between TNF genotypes and other genetic polymorphisms and to compare prevalence of malaria and definitions of iron status (iron deficiency, IDA, ACD) at baseline and after the malaria season. Changes in markers of iron status and hemoglobin level over the malaria season were assessed by Wilcoxon signed-rank test. Multivariate regression models included the explanatory variables of age (grouped by year), village (which also acted as a proxy for ethnic group), the presence of malaria parasites on blood film, α₁-antichymotrypsin (a marker of inflammation), hemoglobin type (HbAA/HbAS) and G6PD deficiency (A type). Because all children of the Fulani ethnic group were located in a single village and village was a powerful explanatory variable (even when excluding the Fulani village), village alone was used to adjust for both location and ethnic group.

Odds ratios (ORs) for iron deficiency, IDA, ACD, and being iron replete were derived by univariate and multivariate logistic regression models. SNPs were first analyzed individually in a multivariate model and then entered simultaneously into a multivariate model with subsequent exclusion of markers with P more than .1. Multivariate logistic regression models tested for associations with haplotype and further multivariate models then analyzed for a dose-response relationship with single haplotypes. Log-transformed markers of iron status were analyzed using univariate and multivariate linear regression models. G6PD deficiency genotype and sickle cell trait did not influence markers of iron status and were thus not included in multivariate analysis of markers of iron status.

From a cohort of 780 children recruited to the study, 756 had DNA typed for polymorphisms at the TNF gene locus. The characteristics of the study population and frequencies of the case definitions for nutritional iron status are summarized in Table 1. Over the malaria season there was a significant increase in the prevalence of iron deficiency (from 18.0% to 31.1% of children, P < .001) and IDA (from 11.9% to 21.7% of children, P < .001), and a marginal increase in the prevalence of ACD (from 10.1% to 13.7% of children, P = .03), shown in Table 1. Individual markers of iron status (ferritin, ZnPP, sTfR, and Ts) also indicated an increase in iron deficiency over the malaria season (Table 1). Multiple factors influenced the risk of iron deficiency and IDA. In multivariate models with iron deficiency and IDA as dependent variables, the following variables emerged as significant predictors: age group (P < .001), village (P = .008), malaria parasites on blood film (P < .001), α₁-antichymotrypsin levels (P < .001) and TNF genotype. Sex, ethnic group, hemoglobin type, and G6PD deficiency genotype did not significantly influence the risk of iron deficiency or IDA.

All of the TNF SNPs were in Hardy Weinberg equilibrium with the exception of the TNF₋₂₃₈ polymorphism, which had a higher than expected number of variant TNF₋₂₃₈AA homozygotes (4 observed, 0.8 expected). The gene frequency of the TNF₋₃₀₈A allele (TNF2) in this population was 0.16 in agreement with a previous study in The Gambia.³⁹  The frequency distribution of the TNF genotypes is shown in Table 2. The LTA₊₂₄₉ and IκBL₋₆₃ markers were in almost complete LD (r² = 0.94), making it difficult to distinguish separate effects of these markers. Pairwise measures of LD for the TNF markers are presented in Table S1 (available on the Blood website; see the Supplemental Materials link at the top of the online article). With the exception of ethnic group, none of the clinical characteristics assessed (sex, age, hemoglobin type [HbAA/HbAS], G6PD deficiency A type, and parasite prevalence) were significantly associated with TNF polymorphisms. The Mandinka ethnic group had a significantly higher proportion of minor alleles for the TNF₋₃₀₈, LTA₊₂₄₉, and IκBL₋₆₃ SNPs compared with the Fulani and a greater proportion of Fulani children had the variant TNF₋₂₃₈A allele compared with the Mandinka group. All of the Fulani children were located in a single Fulani village and ethnic group was accounted for in all analyses by adjusting for village, which was itself a powerful explanatory variable.

In single marker multivariate logistic regression models and in a further model of joint marker effects the TNF markers were not significantly associated with iron deficiency or IDA at baseline before the malaria season (Figure 2). However, there was a trend toward increased risk of iron deficiency with the TNF₋₃₀₈AA (adjusted OR 3.0 [95% CI 0.90-9.98]; P = .07) and the LTA₊₂₄₉CC genotypes (adjusted OR 1.96 [95% CI 1.0-3.83]; P = .05).

In single marker analysis and then in a model of joint marker effects at the end of the malaria season the TNF₋₃₀₈AA genotype was strongly associated with an increased risk of iron deficiency (adjusted OR 8.08 [95% CI 2.33-27.96]; P = .001). The TNF₋₂₃₈AG genotype (there were too few TNF₋₂₃₈AA homozygotes [n = 4] for analysis) was associated with an increased risk of iron deficiency of borderline statistical significance (adjusted OR 2.52 [95% CI 0.97-6.55]; P = .06), shown in Table 2.

The TNF₋₃₀₈AA and the TNF₋₂₃₈AG genotypes were associated with an increased risk of IDA at the end of the malaria season (adjusted OR 5.05 [95% CI 1.47-17.33]; P = .01 and adjusted OR 3.58 [95% CI 1.36-9.44]; P = .01 for each marker, respectively) in a model of joint marker effects (Table 2 and Figure 2). At the end of the malaria season 40.9% of children carrying the TNF₋₃₀₈AA genotype were classified as having IDA compared with 21.1% of children without this genotype and 30.3% of children carrying the TNF₋₂₃₈AG genotype had IDA compared with 21.9% of children carrying the TNF₋₂₃₈GG genotype (Figure 2). The TNF₋₃₀₈AA genotype (but not the TNF₋₂₃₈AG genotype) was also associated with increased iron deficiency in individual markers of iron status at the end of the malaria season (Table 3).

Haplotype analysis identified 9 haplotypes (2 with less than 1% population frequency) resolved by SNPs at nt-63 in the IκBL locus, +249 in the LTA locus and −1031, −308, −238, and +851 in the TNF locus (Table 4). Homozygotes for haplotype 3 (discriminated by the TNF₋₃₀₈A allele) had a significantly increased risk of iron deficiency (adjusted OR 5.12 [95% CI 1.73-15.16]; P = .003) and IDA (adjusted OR 3.40 [95% CI 1.10-10.44]; P = .03) in a logistic regression model analyzing for a dose-response relationship (Table 5). Heterozygotes for haplotype 3 did not have an increased risk of iron deficiency or IDA (adjusted OR 0.90 [95% CI 0.55-1.45] and adjusted OR 1.07 [95% CI 0.63-1.81], respectively). Haplotypes 4 through 7 were present at frequencies too low to allow meaningful analysis of the homozygote state.

Children carrying haplotype 2, distinguished by mutations only at IκBL₋₆₃ and LTA₊₂₄₉, had a marginally significant increased likelihood of being iron replete at the end of the malaria season in a model including all haplotypes (adjusted OR 1.51 [95% CI 1.03-2.23]; P = .04; Table 4). In a logistic regression model looking for a dose response relationship (Table 5), homozygotes (but not heterozygotes) for haplotype 2 were more frequently iron replete (adjusted OR 2.79 [95% CI 1.19-6.55]; P = .018). Thus, haplotype 2 and haplotype 3, discriminated only by the TNF₋₃₀₈A allele, appear to have opposite effects on iron status.

We also investigated whether TNF polymorphisms might be associated with ACD due to a diversion of iron into macrophage storage sites and iron-limited erythropoiesis. In single marker, multivariate analysis a reduced risk of ACD (of marginal significance) was associated with IκBL₋₆₃TT and LTA₊₂₄₉CC homozygotes (adjusted OR 0.30 [95% CI 0.10-0.91]; P = .03 and adjusted OR 0.25 [95% CI 0.07-0.85]; P = .03) at the end of the malaria season. However, in a multimarker logistic regression model of TNF SNPs and in a model including all TNF haplotypes there was no significant association between TNF polymorphisms and ACD.

We found a significant increase in the prevalence of iron deficiency (from 18% to 31% of children; P < .001) and IDA (from 12% to 22% of children; P < .001) in a large cohort of rural Gambian children over a malaria season. Because the malaria season coincides with the “hungry season” in The Gambia, it is likely that dietary iron insufficiency is a major cause. However, we also hypothesized that malaria-induced TNF-α production might play a role. There is compelling evidence that malaria (even in mild and asymptomatic infection) significantly increases plasma levels of TNF-α.^4,5  In addition to blocking recycling of macrophage iron,⁶  and inhibiting erythropoiesis,^9,11  TNF-α directly inhibits absorption of iron through the small bowel in vitro^12,14  and in mouse models.¹³  We found that IDA increased generally over a malaria season and that this effect was concentrated in individuals carrying the TNF₋₃₀₈AA and TNF₋₂₃₈AG genotypes (Figure 2). Nutritional iron status at baseline, before the malaria season, did not differ significantly by TNF genotype.

Homozygotes for the TNF₋₃₀₈A allele had an increased risk of iron deficiency (adjusted OR 8.08, [95% CI 2.33-27.96]; P = .001) and IDA (adjusted OR 5.05; [95% CI 1.47-17.33]; P = .01) at the end of the malaria season compared with other genotypes. At the end of the malaria season 40.9% of children homozygote for the TNF₋₃₀₈A allele had IDA compared with 21.1% of children with other genotypes (Figure 2). These findings were robust to adjustments for other factors influencing iron status and anemia and were also supported by analysis of individual markers of iron status. Zinc protoporphyrin levels were raised in the TNF₋₃₀₈AA genotype compared with the TNF₋₃₀₈GG genotype (P = .006) indicating an increase in iron-limited erythropoiesis. Moreover, despite known actions of TNF-α to stimulate ferritin synthesis⁷  and inhibit transferrin receptor expression,^40,41  plasma ferritin levels were reduced (P = .05) and soluble transferrin receptor levels were insignificantly raised (P = .26) in TNF₋₃₀₈AA homozygotes compared with TNF₋₃₀₈GG homozygotes.

We found that homozygotes, but not heterozygotes for the TNF₋₃₀₈A allele had an increased risk of iron deficiency and IDA at the end of the malaria season (Table 2). Similarly, a large study in The Gambia found that homozygotes for the TNF₋₃₀₈A allele, and not heterozygotes, were at increased risk of cerebral malaria.³⁹  A further study in Kenya found that homozygosity for the TNF₋₃₀₈A allele was associated with premature birth and increased risk of mortality in those born prematurely.²⁶  In low birth weight Kenyan children the TNF₋₃₀₈A allele was also associated with severe anemia.²⁶  In Kenya and The Gambia the TNF₋₃₀₈A was not found to be significantly associated with severe malaria anemia.^26,27,39  The relationship between the TNF₋₃₀₈A SNP and nutritional iron status in malaria-endemic areas has not previously been examined. Studies in patients with hereditary hemochromatosis have found differing results with the TNF₋₃₀₈A allele. One study found a trend toward lower ferritin levels and protection from liver damage,²⁴  another increased iron loading²⁵  and 2 studies found no effect.^42,43  These studies were in iron-overloaded adults rather than in children at risk of iron deficiency and the numbers of TNF₋₃₀₈AA homozygotes were few (n = 0-2).

So how might the TNF₋₃₀₈AA genotype result in iron deficiency and IDA at the end of the malaria season? Functional studies have indicated that the TNF₋₃₀₈A allele is associated with higher constitutive and inducible levels of TNF-α transcription compared with the TNF₋₃₀₈G allele.^19,20  Moreover, homozygotes for the TNF₋₃₀₈A allele and haplotype appear to have a striking increase in TNF-α production compared with TNF₋₃₀₈AG and TNF₋₃₀₈GG genotypes (7194 pg/mL for TNF₋₃₀₈AA, 4698 pg/mL for TNF₋₃₀₈AG and 3832 pg/mL for TNF₋₃₀₈GG after stimulation with chlamydial EB).⁴⁴  Recent studies have demonstrated that TNF-α directly inhibits iron absorption in the small bowel by increased enterocyte ferritin deposition and relocalization of ferroportin1 away from the basolateral enterocyte membrane via a hepcidin-independent mechanism.^12-14  We hypothesized that increased production of TNF-α in the TNF₋₃₀₈A homozygote during chronic and recurrent malaria infections, in children who are already in a precarious state of iron balance, would further block iron absorption resulting in iron deficiency. We did not find a significant association with genotype at baseline, before the malaria season. This demonstrates the importance of gene/environment interactions. In an environment of chronic malaria-induced inflammation and dietary iron insufficiency, there was an increased risk of iron deficiency and IDA among children generally, but this effect was concentrated in TNF₋₃₀₈A homozygotes. It is possible that in an iron-replete or iron-overloaded environment a TNF-α–induced block in macrophage iron export would outweigh a block in iron absorption and the TNF₋₃₀₈AA genotype might instead be a risk factor for ACD.

The TNF₋₂₃₈AG heterozygote (n = 4 for homozygotes, which did not allow meaningful analysis) was associated with an increased risk of IDA (adjusted OR 3.6; P = .01), but was not significantly associated with an increased risk of iron deficiency (adjusted OR 2.5; P = .06). Studies indicating that the TNF₋₂₃₈A allele influences transcription are conflicting: one study found reduced transcription²²  and others no effect.^23,45  The TNF₋₂₃₈A allele was associated with a reduced prevalence of hereditary hemachromatosis and less severe liver damage in one study,²⁴  although other studies did not find an effect.^42,43  The TNF₋₂₃₈A allele has also been associated with severe malarial anemia.²⁷  The TNF₋₂₃₈A allele was not significantly associated with iron deficiency in our study suggesting that other mechanisms, besides a reduction in iron absorption, are likely to be important. A study in Gabon found that the TNF₋₂₃₈A allele was significantly associated with a lower IL-10 to TNF plasma level (IL-10:TNF ratio <1)⁴⁶  and a low IL-10:TNF ratio has been associated with severe malarial anemia.^46,47  A further possibility is that the TNF₋₂₃₈A allele might be in LD with another functional mutation.

We also mapped local patterns of LD and constructed haplotypes to increase the chances of capturing a functional mutation which might reside within a given haplotype. As expected, we found that homozygotes for haplotype 3, uniquely defined by the TNF₋₃₀₈A allele, had an increased risk of iron deficiency (adjusted OR 5.1, P = .003) and IDA (adjusted OR 3.4, P = .03) in a logistic regression model looking for a dose-response relationship (Table 5). We also unexpectedly found that homozgotes for haplotype 2, distinguished by mutations only at IκBL₋₆₃ and LTA₊₂₄₉, were more likely to be iron replete (adjusted OR 2.79, [95% CI 1.19-6.55], P = .018) at the end of the malaria season in single haplotype analysis. A possible explanation is that the IκBL gene product, thought to be part of the I kappa B family of regulatory proteins, inhibits the function of nuclear factor κB (NF-κB),⁴⁸  and NF-κB appears to be essential for the transcriptional activation of TNF within monocytes.⁴⁹  The variant IκBL₋₆₃T allele has been associated with a moderate increase in transcriptional activity in vitro compared with the major IκBL₋₆₃A allele.⁵⁰  Thus, it is interesting to speculate that by inhibiting NF-κB-induced transcription of TNF, the variant IκBL₋₆₃T allele might dampen the effects of TNF-α on intestinal and macrophage iron transport. The LTA_{+ 249}C allele is in strong LD with the IκBL₋₆₃T allele. It is also possible that an undetected functional mutation may lie within this haplotype.

In summary, our results suggest that variation in the TNF gene locus is a risk factor for iron deficiency and IDA in malaria-endemic countries. Studies have highlighted the difficulty of assessing iron status in malaria-endemic countries due to chronic inflammation obscuring true levels of iron deficiency.³  However, our findings suggest that chronic malaria-induced TNF-α production is itself an important cause of iron deficiency in children in malaria-endemic areas and supports the findings of in vitro and animal studies that show that TNF-α blocks iron absorption.^12-14  The relative benefits of iron supplementation and malaria treatment on childhood anemia have been debated.⁵¹  However, treatment of malaria might also help to prevent IDA in malaria-endemic areas by unblocking iron absorption. The TNF₋₃₀₈A allele exists at a gene frequency of 0.16 in The Gambia despite its association with cerebral malaria³⁹  and it is intriguing to speculate that this might be due to balancing evolutionary pressures. There is mounting evidence that iron is pivotal to host-pathogen interactions,⁵²  and it is possible that an “iron-deficient phenotype” might represent an evolutionary adaptation for limiting iron availability for microorganisms and protecting from infectious disease.

We thank the children who took part in the study and their parents. We would also like to thank Baba Jobarteh, Sosseh Sanyang, Kabiru Ceesay, Musa Colley, Khalilu Sanneh, and Ibrima Camera for their assistance in the field and laboratory. We also thank Chris Bates and Anna Richardson for helpful discussions and assistance with, respectively, the iron assays and genotyping.

This work was supported by the United Kingdom Medical Research Council (London).

Contribution: S.H.A., P.A.B., D.P.K., and A.M.P. designed the study; S.H.A., G.M., K.R., G.S., M.A.O., and N.H. performed field and laboratory work; S.H.A., K.R., and P.A.B. analyzed results; and S.H.A., K.R., P.A.B., G.M., G.S., M.A.O., N.H., D.P.K., and A.M.P. contributed to writing the paper.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Dr Sarah H. Atkinson, University of Oxford, Department of Paediatrics, Children's Hospital (John Radcliffe), Oxford, United Kingdom; e-mail: sarah.atkinson@paediatrics.ox.ac.uk.