Atopic dermatitis (AD) is a chronic inflammatory skin disease characterised by recurrent eczematous lesions. These lesions are ill-defined and are often accompanied by erythema, early exudation, blistering and crusting; later, they may present with scaling, dehiscence and lichenisation. AD patients are also prone to allergic rhinitis, asthma, intense itching and discomfort(Reference Weidinger, Beck and Bieber1,Reference Weidinger and Novak2) . The prevalence of AD is estimated to be 15–20 % in children and 1–3 % in adults, with the incidence of AD increasing 2–3-fold in industrialised countries over the past decades(Reference Nutten3,Reference Avena-Woods4) . It is the leading non-fatal health burden attributable to skin diseases and inflicts a substantial psychosocial burden on patients(Reference Ständer5–Reference Boguniewicz, Fonacier and Guttman-Yassky9). A strong link between AD and food allergies has been established in childhood. Food allergies are also present in up to 37 % of infants with AD, whereas they are present in approximately 10 % of adults with AD(Reference Manam, Tsakok and Till10). There is some debate about the role of food allergies in AD(Reference Roerdink, Flokstra-de Blok and Blok11–Reference Niggemann, Sielaff and Beyer13). In assessing allergen triggers in AD patients, doctors typically perform a range of tests, including tests for food allergens. These tests may ask patients to avoid certain foods to determine whether they improve symptom control for AD(Reference Singh, Anvari and Hauk14). However, this ‘elimination diet’ approach does not fully determine the role of food allergies in the development of AD because there are many other factors that may affect the control of the disease, such as genetics, the environment and the microbiome.
Mendelian randomisation (MR) studies are defined as any study that uses genetic variation as a robust proxy for modifiable environmental exposures to make causal inferences about the outcomes of modifiable exposures(Reference Lawlor, Harbord and Sterne15). MR have been likened to ‘natural’ randomised controlled trials. MR exploits the fixity of genes and Mendel’s first and second laws of inheritance; namely, at meiotic gamete formation, alleles from parents are randomly assigned to offspring. MR uses genetic variants associated with risk factors of interest to explore their association with disease outcome(Reference Burgess, Davey Smith and Davies16–Reference Sanderson, Glymour and Holmes18). The process is similar to that of a traditional randomised controlled trials; in that, patients are randomised into treatment and control groups. However, MR studies use random allocation of genetic variation to avoid the interference of reverse causality and potential confounding factors encountered in traditional randomised controlled trials(Reference Hingorani and Humphries19,Reference Davey Smith and Ebrahim20) .
Previous studies have shown that different dietary patterns have varying degrees of impact on the risk of AD. However, the causal relationship between AD and dietary patterns has not been assessed systematically. Excessive restriction of dietary intake may lead to nutritional deficiencies. Our study is the first to explore the association between dietary factors and AD risk using MR. The primary objective of our research was to substantiate the existence of a causal relationship between diet and AD at the genetic level. By establishing this connection, our study aimed to provide valuable insights that can guide the development of tailored dietary recommendations for individuals with AD.
Methods
This study followed the Strengthening the Reporting of Observational Studies in Epidemiology–Mendelian randomisation reporting guidelines (online Supplementary Table S1)(Reference Skrivankova, Richmond and Woolf21). This study used only publicly available summary data, which were approved for human experimentation by an ethical standards committee. Therefore, additional ethical approval was not required for the current study. MR uses genetic variants to assess causal relationships using observational data(Reference Burgess, Scott and Timpson22,Reference Davey Smith and Hemani23) . The principle of MR is based on Mendelian laws of inheritance and instrumental variable (IV) estimation methods, which enable the inference of causal effects in the presence of unobserved confounding factors(Reference Sanderson, Glymour and Holmes18). The relationships among exposure factors (dietary habits), IV and the outcome (AD) are detailed in Fig. 1. MR analysis must satisfy the following three model assumptions(Reference Didelez and Sheehan24): (1) IV are strongly correlated with intermediate phenotypes or exposure factors; (2) IV are not correlated with confounding factors; and (3) conditional independence between IV and disease outcome, that is, IV affect outcome only through exposure
Fig. 1. The relationship between exposure factors (dietary habits), instrumental variables (IV) and outcomes (AD).
Data sources
The diet-related exposure factors used in this study included meat intake (processed meat, poultry, beef, non-oily fish, oily fish, pork and lamb/mutton intake), staple food intake (white bread and cereal intake), vegetable intake (salad/raw vegetable intake and cooked vegetable intake), beverage intake (weekly number of alcoholic beverages consumed and frequency of alcohol consumption, tea intake, milk intake and coffee intake), fruit intake (dried fruit intake and fresh fruit intake) and intake of other foods (cheese intake, whole egg intake and unsalted peanut intake). Genome-wide association study (GWAS) summary-level data were extracted directly or indirectly from UK Biobank using the IEU Open GWAS project, and details of the design, such as quality control procedures and statistical analyses, are available on the website (http://www.nealelab.is/UK-biobank/c). Dietary intake was measured by questionnaires as the average intake over the past year. GWAS summary-level data for AD were extracted from the FinnGen consortium (https://r5.finngen.fi)(Reference Kurki, Karjalainen and Palta25). The R5 release data of the FinnGen consortium include 7024 AD patients and 198 740 control participants. Cases were ascertained by hospital discharge and cause of death (code L20 of the ICD-10), 6918 (with the exclusion of 6918X) of the ICD-9, and 691 of the ICD-8. FinnGen consortium researchers define AD as ‘A chronic inflammatory genetically determined disease of the skin marked by increased ability to form reagin (Ig E), with increased susceptibility to allergic rhinitis and asthma, and hereditary disposition to a lowered threshold for pruritus. It is manifested by lichenification, excoriation, and crusting, mainly on the flexural surfaces of the elbow and knee’. Table 1 shows the baseline characteristics of AD patients in the FinnGen consortium.
Table 1. Baseline characteristics of atopic dermatitis patients in FinnGen consortium
Selection of instrumental variables
All the statistical analyses were performed using the ‘MendelianRandomization’ and ‘TwoSampleMR’ packages in R (3.6.1). We performed analyses using a two-sample MR model with dietary factors as the exposure variables and AD as the outcome variable. In MR analysis, IV are utilised as mediators between exposure factors and outcomes to explore the causal relationship between exposure and outcomes and are generally genetic variations, among which SNP are the most commonly used(Reference Smith and Ebrahim26). In this study, SNP sites associated with significant dietary factors were selected in advance based on the criterion (P1 = 5e-8); dietary factors with no significant sites were extracted, and the criterion was adjusted to 5e-6. In the first step, we used standard parameters (aggregation window of 10 000 kb and an r2 cut-off of 0·001) to eliminate the interference of linkage disequilibrium(Reference Slatkin27). In the second step, SNP with genome-wide significance (P < 5 × 10 − 5) for AD risk were excluded. Then, SNP with the same alleles in the dietary factor data and AD data were collated to harmonise the effect sizes of exposure and outcome(Reference Hartwig, Davies and Hemani28). To avoid potential weak instrumental bias, the F statistic (F = beta2/se2) was used to assess the strength of the IV; if F > 10, the correlation between IV and exposure was considered strong enough that the results of the MR analysis could be protected from weak instrumental bias(Reference Burgess and Thompson29).
Mendelian randomisation analysis
We performed a two-sample MR analysis to assess the causal relationship between exposure (dietary factors) and outcome (AD). We utilised the inverse-variance weighting (IVW), weighted median, weighted mode and MR‒Egger methods for the main MR analysis(Reference Nikolakopoulou, Mavridis and Salanti30–Reference Bowden, Del Greco and Minelli32). First, associations were made for each SNP, and the IVW method was used to combine Wald ratios to assess associations between dietary factors and AD. The IVW method adopted in this study mainly uses the inverse of variance of each IV as a weight for weighting calculation. The results of the random-effects IVW method were regarded as the predominant analysis(Reference Zuber, Colijn and Klaver33). This process is carried out under the premise of ensuring that all IVs are valid to evaluate horizontal pleiotropy(Reference Burgess, Bowden and Fall34). MR‒Egger regression uses the inverse of the variance of the outcome as a weight for fitting, which adds an intercept term to the regression(Reference Slob, Groenen and Thurik35). The weighted median method(Reference Bowden, Davey Smith and Haycock31) involves the median of a weighted empirical density function defined as a ratio estimate that still provides consistent effect estimates when the proportion of invalid IV is as high as 50 % and the accuracy of the estimates varies widely among IV.
Sensitivity analysis
To ensure the reliability of our conclusions, six sensitivity analyses were used to verify whether heterogeneity and pleiotropy in genetic variables biased the MR results. First, MR‒Egger regression was applied to detect and adjust for the effect of potential horizontal pleiotropy among the selected IV by assessing the intercept(Reference Burgess and Thompson36,Reference Verbanck, Chen and Neale37) . Second, MR pleiotropy residual and outlier (MR-PRESSO) tests were used to detect potential horizontal pleiotropy, and MR-PRESSO global heterogeneity tests were performed to identify potential horizontal pleiotropy, which were immediately removed once outliers were found. Third, the asymmetry of the funnel plot was used as an indicator of horizontal pleiotropy. Fourth, the Cochran Q test was used to quantify the heterogeneity of the selected genetic instruments(Reference Burgess and Thompson36,Reference Verbanck, Chen and Neale37) . Fifth, leave-one-out sensitivity analysis was implemented by deleting each SNP, which ensured that the MR estimates were not driven by certain strong SNP. Finally, after removing outliers, the MR analysis was performed again.
Results
The flow chart of the study design, data collection, data processing and analysis, and exclusion of SNP numbers is shown in Fig. 2
Fig. 2. Study flow diagram. *p1 = 5e-8, p1 changed to 5e-6 if tool variables are insufficient.
Selection of instrumental variables
The causal relationships between dietary factors and AD were analysed using twenty-one different exposure factors. The number of European-descent individuals included in the study ranged from 335 394 to 462 346. The outcomes included 7024 European-descent AD patients and 198 740 European-descent control participants from the FinnGen consortium, and there was little overlap between the populations involved in the exposures and outcomes. The number of SNP used for different exposures in this study ranged from 7 to 90. The F statistic for all SNP was greater than 10, which indicates that the IV used in our study satisfied the requirements of strong associations with the exposures. Table 2 provides detailed information on the twenty-one exposure factors.
Table 2. Information of the exposures and outcome datasets
IEU, Integrated Epidemiology Unit; GWAS, genome-wide association study;
* P1 = 5e-6. In addition to the marked exposure, the remaining P-values are 5e-8.
Causal effect in the main analysis and sensitivity analyses
In this study, we found that beef intake (OR = 0·351; 95 % CI 0·145, 0·847; P = 0·020) was associated with a decreased risk of AD, which was further verified by the results of the MR‒Egger model (OR = 0·002; 95 % CI 0·000, 0·341; P = 0·036) despite the negative results of the weighted median and weighted mode models. The IVW results for white bread suggested that white bread might be a protective factor against AD (OR = 0·141; 95 % CI 0·030, 0·656; P = 0·012). However, the wide CI indicates uncertainty about the protective effect of white bread on AD. Therefore, we cannot verify the protective effect of white bread on AD.
This study also found that the number of alcoholic drinks per week (OR: 0·721; 95 % CI 0·441, 1·179; P = 0·192), alcohol consumption frequency (OR: 1·027; 95 % CI 0·856, 1·233; P = 0·771), processed meat intake (OR: 0·638; 95 % CI 0·300, 1·353; P = 0·241), poultry intake (OR: 2·565; 95 % CI 0·668, 9·847; P = 0·170), non-oily fish intake (OR: 1·125; 95 % CI 0·299, 4·241; P = 0·862), oily fish intake (OR: 0·980; 95 % CI 0·647, 1·485; P = 0·923), pork intake (OR: 0·920; 95 % CI 0·313, 2·710; P = 0·880), lamb/mutton intake (OR: 0·828; 95 % CI 0·403, 1·702; P = 0·607), cheese intake (OR: 0·856; 95 % CI 0·598, 1·224; P = 0·393), cooked vegetable intake (OR: 1·586; 95 % CI 0·636, 3·952; P = 0·322), tea intake (OR: 0·714; 95 % CI 0·495, 1·030; P = 0·071), fresh fruit intake (OR: 0·618; 95 % CI 0·338, 1·129; P = 0·118), cereal intake (OR: 1·283; 95 % CI 0·720, 2·286; P = 0·399), salad/raw vegetable intake (OR: 1·732; 95 % CI 0·606, 4·952; P = 0·306), coffee intake (OR: 0·902; 95 % CI 0·593, 1·373; P = 0·631), milk intake (OR: 0·994; 95 % CI 0·493, 2·001; P = 0·986), whole egg intake (OR: 1·238; 95 % CI 0·631, 2·426; P = 0·535), unsalted peanuts intake (OR: 1·404; 95 % CI 0·555, 2·426; P = 3·553) and dried fruit intake (OR: 0·864; 95 % CI 0·448, 1·666; P = 0·662) were not associated with AD. Additional MR analysis results are shown in Table 3.
Table 3. The results of Mendelian randomisation analyses
*P1 = 5e-6. In addition to the marked exposure, the remaining P-values are 5e-8.
The analysis of beef intake showed that the P value of the IVW method was less than 0·05, and the direction of the β-coefficient was consistent for the remaining methods. Therefore, we concluded that beef intake was a protective factor against AD. To ensure the robustness of the positive results, we performed a sensitivity analysis. Heterogeneity was not detected in the exposures (Cochran’s Q test P > 0·05), and the results of the MR‒Egger intercept suggested that no directional pleiotropy was present (Table 4). A scatter plot of the individual causal effect estimates is shown in Fig. 3. The diagnostic funnel plot showed visually apparent symmetry, which excluded the possible influence of directional pleiotropy on our estimates (Fig. 3). A forest plot is shown in Fig. 3. Leave-one-out analysis indicated that the causality of the associations was robust (Fig. 3). As shown in Table 3, the results of the MR-PRESSO analysis were consistent with those of the IVW model.
Table 4. The results of sensitivity analyses
Fig. 3. MR results of beef intake and atopic dermatitis (AD): (a) scatter plot of genetic correlations of beef intake and AD using different MR methods. The slopes of line represent the causal effect of each method, respectively; (b) forest plot of the causal effects of beef intake associated SNP on AD. The red and black dot/bar indicate the causal estimate of beef intake on risk of patients with AD;(c) MR results of leave-one-out sensitivity analysis for beef intake and AD; (d) funnel plot the causal effects of beef intake associated SNP on AD.
Discussion
An increase in the incidence of AD is associated with an increase in food allergies, but the relationship is complex. Opinion is divided on the use of dietary elimination as an effective treatment option for AD because of growing evidence that other factors are driving the development of AD, independent of food consumption(Reference Oykhman, Dookie and Al-Rammahy38). Furthermore, excessive dietary restriction may cause nutritional deficiencies. To our knowledge, the present study is the first to verify the causal relationship between diet and AD at the genetic level. The most important finding of this MR analysis was that beef intake was associated with a reduced risk of AD. Therefore, when beef allergies are excluded, AD patients can eat some beef appropriately, which may help to improve their skin condition. Due to the wide 95 % CI (0·030–0·656), we could not determine the exact protective effect of white bread against AD. The width of the CI was influenced by the sample size, the variability of the data or the statistical methods used. Therefore, further studies are needed to verify this conclusion. In addition, this study revealed that the number of alcoholic drinks per week, alcohol consumption frequency, processed meat intake, poultry intake, non-oily fish intake, oily fish intake, pork intake, lamb/mutton intake, cheese intake, cooked vegetable intake, dried fruit intake, fresh fruit intake, tea intake, cereal intake, salad/raw vegetable intake, whole egg intake, unsalted peanuts intake, milk intake and coffee intake were not associated with AD. Therefore, after the exclusion of foods that trigger an allergic response, rational food intake should be used to avoid excessive use of dietary avoidance to treat AD. Excessive avoidance of certain foods may lead to nutritional deficiency, which is detrimental to the health of patients. For patients with food allergies, we recommend regular food allergy monitoring combined with one-on-one consultation from a dietician on food intake for better management and treatment of AD. In conclusion, this study provides comprehensive dietary guidance for AD patients and helps them improve their skin condition while maintaining balanced nutrition. Moreover, this study provides more specific suggestions and treatment options for patients with food allergies.
However, studies on beef allergies are rare. In a study by Ogle et al., beef was considered hypoallergenic. In contrast, Chandra et al. considered beef to be a strong allergen(Reference Chandra, Puri and Suraiya39). Our study suggested that beef may be a protective factor against AD. Unlike observational studies, MR studies using genetic variants (mainly SNP) as IV are not affected by confounding factors or reverse causality. We hypothesised that beef type, ethnicity and age would have some effect on the results.
There is currently some controversy about the relationship between alcohol use and AD. A study of dietary habits in Japanese adult patients with AD revealed a negative association between alcohol consumption and the development of AD(Reference Ito, Morita and Okazaki40). However, several epidemiological studies suggest that increased alcohol consumption is also associated with increased risk of developing AD(Reference Sawada, Saito-Sasaki and Mashima41,Reference Lim, Lim and Ng42) . Nonetheless, we did not find a clear causal relationship between alcohol consumption and AD according to the MR analysis. Therefore, further research and discussion of the relationship between alcohol consumption and AD are still needed. Alcohol consumption and AD appear to be associated to some extent, but the specific mechanism of this association is not clear. Several studies have shown that high alcohol consumption is associated with elevated IgE levels but does not necessarily indicate an increased risk of allergic disease(Reference Lomholt, Nielsen and Nordestgaard43). Another study revealed a sex difference between alcohol consumption and IgE sensitisation. However, the specific effects of alcohol on allergic diseases may vary by individual and alcohol type(Reference Roh, Lee and Lee44). Observational studies are susceptible to confounding factors, such as alcohol consumption, which may be accompanied by other confounding factors such as smoking and excessive energy intake. These confounding factors may influence the accuracy of the results. In contrast, MR analysis is able to remove these confounders and more accurately reflect the relationship between the variables. However, due to sample size limitations in the database, our study only verified the association between alcohol consumption and AD in the European population. Therefore, our results only apply to European populations and cannot be generalised to other populations. In addition, the relationship between alcohol consumption and AD still needs further investigation. Future research should focus on the specific effects of alcohol consumption on allergic diseases and how that relationship is impacted by sex and quantity and type of alcohol consumed to attain a more accurate assessment of that relationship.
A study from the Swedish Prospective Birth Cohort showed that fish consumption during infancy was associated with the development of allergic disease by the age of 12 years and that regular fish consumption during infancy reduced the risk of allergic disease up to the age of 12 years(Reference Magnusson, Kull and Rosenlund45). Another study of a primary population of patients over 14 years of age with AD showed that allergy to fish in AD patients was associated with persistent eczematous skin lesions(Reference Čelakovská, Josef and Vaneckova46). However, our study showed no causal relationship between fish intake (neither oily nor non-oily) and AD. A cross-sectional study of dietary modifications in patients with AD showed that the consumption of vegetables and fruits improved the skin condition of patients(Reference Nosrati, Afifi and Danesh47). This study used MR analysis methods and did not find a causal relationship between vegetable intake or fruit intake and AD. Notably, previous studies have shown that the consumption of meat(Reference Yang, Byun and Kim48), wheat(Reference Yang, Byun and Kim48), processed foods(Reference Park, Choi and Bae49,Reference Li, Su and Luo50) and coffee is associated with the onset of AD(Reference Uenishi, Sugiura and Uehara51). In contrast, our study did not find a causal relationship between the above factors. Cows’ milk, eggs, peanuts, wheat, soya, nuts and fish are responsible for > 90 % of food allergies in children with AD(Reference Sicherer and Sampson52). Previous studies support the association between milk, egg, and nut intake (e.g. peanuts) and an elevated risk of AD(Reference Yang, Byun and Kim48,Reference Estrada-Reyes, García-Hernández and Martínez-Gimeno53–Reference Du Toit, Roberts and Sayre57) . Using a significance threshold of P1 = 5e-8, we were unable to perform effective studies on milk, egg or peanut intake due to the lack of sufficient IV for MR analysis. To address this problem, we relaxed the significance threshold to P2 = 5e-6. After using this relaxed threshold, our IVW results showed no significant associations between milk, eggs, or peanuts and the risk of AD. However, future studies with larger sample sizes are needed to verify the accuracy of our conclusions.
Although some studies have shown that food allergies may play a role in the development of AD in some patients, not all AD patients exhibit food allergy reactions. Furthermore, even if a person is allergic to a certain food, this does not mean that the food is the cause of AD(Reference Oykhman, Dookie and Al-Rammahy38). We speculate that there is most likely a common pathogenesis between food allergy and AD. This study provides strong new evidence for the relationship between dietary factors and AD. A possible pathway for dietary factors to affect AD is through the gut microbiome, as the intake of different foods can affect the composition of bacteria in the gut, thereby affecting nutrient metabolism(Reference Kau, Ahern and Griffin58,Reference Muegge, Kuczynski and Knights59) . There is growing evidence that the establishment of the gut microbiome early in life influences the development of AD(Reference Chu, Meng and Yu60). A two-sample MR study demonstrated bidirectional causality between the gut flora and AD(Reference Jin, Ren and Dai61). Therefore, these studies revealed a role for the gut flora in the pathogenesis of AD.
The strengths of this study are that it is the first large-scale MR analysis to explore the causal relationship between twenty-one dietary factors and AD. Our findings provide stronger evidence than traditional observational studies. This study has several limitations, and the results of MR need to be interpreted with caution. First, the causal relationship found in the MR analysis reflects the effects of long-term exposure; thus, short-term exposure may not be clinically meaningful. Second, we were unable to distinguish between causal relationships at different time points; for example, allergies in AD patients usually occur in childhood, and patients gradually tolerate sensitisation with age. Therefore, our findings may not be meaningful for infants. Furthermore, the univariable MR analysis showed only an overall effect between exposure and outcome and not a direct effect between them. Extremely complex mechanisms may exist between exposure and outcome. The most important point is that due to the database sample size and sample type, we were unable to study the AD population by age and disease severity. In future studies with suitable samples, more detailed analysis should be performed. Finally, this study mainly included data from individuals of European ancestry, preventing us from extrapolating these findings to other ethnic groups. We were unable to further subdivide the different types of dietary intake or distinguish the effects of different dietary combinations. Therefore, further studies with larger sample sizes are needed to validate our findings.
Conclusions
The study showed that beef intake may be associated with a reduced risk of AD. Due to the wide 95 % CI( (0·030–0·656), we could not determine the exact protective effect of white bread against AD. In addition, the number of alcoholic drinks per week, alcohol consumption frequency, processed meat intake, poultry intake, non-oily fish intake, oily fish intake, pork intake, lamb/mutton intake, cheese intake, cooked vegetable intake, dried fruit intake, fresh fruit intake, cereal intake, salad/raw vegetable intake, whole egg intake, unsalted peanut intake, milk intake and coffee intake were not associated with AD. Our study revealed a causal relationship between diet and AD at the genetic level, with beneficial implications for AD patients. Through personalised dietary advice, they can better cope with and manage diseases.
Acknowledgement
The authors want to acknowledge all participants and investigators from the UK Biobank and FinnGen studies. The authors are particularly grateful to the MRC Integrated Epidemiology Unit (IEU) at the University of Bristol for developing the IEU Open GWAS Project.
This work was supported by the National Natural Sciences Foundation of China (grant number 82074443), ‘Young Qi Huang Scholar’ of State Administration of Traditional Chinese Medicine (grant number 2022–256), Science and Technology Research Project of Sichuan Administration of Traditional Chinese Medicine (grant number 22CP1423), Sichuan Administration of Traditional Chinese Medicine (grant number 2021MS307), Chengdu University of Traditional Chinese Medicine ‘Xing Lin Scholars’ discipline talent Research Promotion Program (grant number QJJJ2021001), Chengdu University of Traditional Chinese Medicine Hospital ‘100 People Program’ (grant number 21-L03 and 22-B09) and Chengdu University of Traditional Chinese Medicine ‘Foundation Thickening’ action plan (grant number 2023–42). The study funders/sponsors had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.
Y. C., J. Y. and Q. W. designed the study, drafted the manuscript and verified the data; Y. C., M. L., D. T. and M. C. conducted the statistical analyses; J. Z. and J. G. drafted the manuscript and revised it critically for important intellectual content; X. X., Y. P. and S. C. substantially contributed to the interpretation of data; and W. H. and Z. X. were accountable for all aspects of the work. All authors critically reviewed and approved the final manuscript.
The authors declare that this research was conducted in the absence of any commercial or financial relationships that could be construed as potential.
This study used only publicly available summary data, which were approved for human experimentation by an ethical standards committee. Therefore, additional ethical approval was not required for the current study.
Supplementary material
For supplementary material/s referred to in this article, please visit https://doi.org/10.1017/S0007114524000436
