Genetic predisposition for immunoglobulin E production explains atopic risk in children: Tohoku Medical Megabank cohort study

Summary

The atopic march lacks early identification methods for high-risk children. In this study, we assessed whether the risk of atopic diseases in infants could be predicted using a polygenic score (PGS) for total immunoglobulin E (IgE) levels. The PGS estimated using the polygenic model generated by PRS-CS was significantly correlated with log-transformed IgE levels (ρ = 0.200, p < 2.2 × 10⁻¹⁶). Assessment of the risk from birth to 2 years of age in a Japanese birth cohort (n = 17,154) applying the estimated PGS revealed significantly elevated incidence risk ratios in the highest PGS quintile (Q5) compared with those in the reference quintiles (Q1–Q3) for food allergy (1.51-fold; 95% confidence interval: 1.30–1.76), atopic dermatitis (1.30-fold; 1.12–1.51), and both conditions (1.88-fold; 1.46–2.43). These findings address critical gaps in allergy and PGS research among non-European populations, suggesting the contribution of genetic predisposition to IgE production in early-onset allergic diseases and supporting the use of PGS in early intervention.

We report that genetic predisposition influencing serum immunoglobulin E (sIgE) levels is associated with the development of allergies in children, including food allergy and atopic dermatitis. Polygenic scores for sIgE may help identify children at high genetic risk, enabling early intervention to reduce allergy-related disease burden.

Introduction

Allergic diseases are common health issues in children, profoundly affecting their quality of life.^1,2 Food allergies (FAs), for instance, affect approximately 10% of infants in the United States³ and Australia,⁴ with evidence suggesting rising prevalence rates in non-Western countries, although data from these regions remain limited.⁵

In childhood, allergic conditions often follow a sequential progression known as the “atopic march,” which reflects the natural history of allergic diseases.^6,7 This progression typically begins with atopic dermatitis (AD [MIM: 603165]) and advances to immunoglobulin E (IgE)-mediated FAs, asthma (AS [MIM: 600807]), and allergic rhinitis (AR [MIM: 607154]).^7,8 In particular, AD is strongly correlated with an increased risk of developing food-specific IgE antibodies and subsequent FAs,^9,10 as well as other allergic conditions.¹¹

The risk of allergic diseases varies among individuals and is significantly influenced by genetic predispositions. Family history is a well-established risk factor for atopic diseases.^12,13 Recent genome-wide association studies (GWASs) have identified associated genetic factors and indicated the existence of a shared genetic background across allergic diseases.^14,15 Among these, IgE—a key mediator in allergic diseases—has shown substantial heritability, ranging from 36% to 78% in segregation analyses¹⁶ and twin studies.¹⁷ Genetic variants associated with IgE levels have been identified in IL4R (MIM: 147781), FCER1A (MIM: 147140), RAD50 (MIM: 604040), STAT6 (MIM: 601512), IL13 (MIM: 147683], and HLA loci (HLA-A [MIM: 142800], HLA-G [MIM: 142871], and HLA-DQA2 [MIM: 613503]) using GWASs.^18,19,20 IL4R encodes the alpha chain of the receptors for interleukin (IL)-4 and IL-13 and interacts with the JAK1/2/3-STAT6 pathway, which regulates IgE production and is implicated in the development of AS and other allergic diseases.^21,22,23 These findings underscore the genetic overlap between total IgE levels and various allergic conditions.

Given the central role of IgE in allergic disease development, we hypothesized that evaluation of the genetic predisposition to elevated IgE levels would offer a promising strategy for predicting the overall risk of atopic diseases. To test this hypothesis, in this study, we aimed to develop and validate a polygenic score (PGS)-based model for the assessment of IgE levels and evaluate its potential for allergy risk prediction. Our study used data from one of the largest Asian birth cohorts, enabling the classification of children based on their genetic predisposition for IgE production and estimation of their allergy risks. This approach is particularly relevant in children, in whom genetic influences are pronounced due to limited environmental exposure, offering an opportunity for early and comprehensive risk assessment. By addressing the lack of data from non-European cohorts, our findings contribute to the growing field of personalized medicine and provide a foundation for targeted prevention strategies for childhood allergies, aiming to improve the quality of life of high-risk children through optimized early interventions.

Methods

Study population

The Tohoku Medical Megabank Community-Based Cohort (TMM CommCohort) study design has been described in previous studies.^24,25 Briefly, residents from Iwate and Miyagi prefectures, located on the Pacific coast of northeast Japan, were recruited between May 2013 and March 2016. This study included three datasets within this cohort: TMM67K, TMM18K, and TMM9K. TMM67K included participants recruited during public health checkups, while TMM18K and TMM9K included individuals who visited assessment centers for the cohort study. The TMM9K dataset was used for training, and the remaining datasets were used for validation. Overlapping samples in TMM9K were excluded from other datasets. Phenotypes were determined using a baseline survey, and total IgE concentrations were measured using ImmunoCAP total IgE reagents.²⁰ Affected individuals were identified based on self-reported medical histories obtained from self-administered questionnaires.²⁵

The TMM Project Birth and Three-Generation (TMM BirThree) cohort study has been described previously.²⁶ Briefly, pregnant women residing in Miyagi and Iwate prefectures were recruited into the cohort from 2013 to 2016, and genotypic and phenotypic data were collected from the participants, their newborns, and their families. The parent-reported medical histories of the newborns surveyed at approximately 1, 6, 12, and 24 months after birth were used for phenotyping. If a parent affirmed the question, “Has your child ever been diagnosed with any diseases by a clinician?” and selected immunological diseases like “allergic conjunctivitis, AR, pollinosis,” “AD,” “FA,” or “bronchial asthma,” the child was classified as an affected individual. Children with other responses were classified as control subjects, excluding instances where the parent did not answer the first question. The observation period started 1 month before the first survey date (approximately 1 month post-birth) and ended either when the event was reported or on the last survey day. Questionnaires submitted 36 months after birth were excluded from the study.

The study was approved by the institutional review boards of Iwate Medical University (HG H25-2 and MH2023-024) and the Tohoku University Tohoku Medical Megabank Organization (ToMMo; first edition: 2012-4-617, latest edition: 2021-4-113). Written consent was obtained from all participants in the TMM study. Personal genotype and phenotype data analyses were conducted exclusively using the ToMMo supercomputer, which is a standalone computing system. This study adhered to the ethical guidelines of the Declaration of Helsinki.

The datasets used in this study are not publicly accessible due to ethical constraints. However, access may be obtained upon approval from the Ethical Committee of Iwate Medical University, Ethical Committee of the Tohoku University Tohoku Medical Megabank Organization (ToMMo), and Materials and Information Distribution Review Committee of the TMM Project.

Genotyping and imputation

Participants from the TMM CommCohort were genotyped using the Axiom Japonica Array v.2 (JPAv2), as described in previous studies,^27,28 whereas those in the TMM BirThree cohort were genotyped using either JPAv2 (“BirThree 1”) or the Axiom Japonica Array NEO (JPA-NEO; “BirThree 2”).²⁹ Participants with a call rate < 0.95 and gender discrepancies between genotypic and cohort data were excluded. Genotypes failing to meet the quality control criteria (call rate < 0.99, Hardy-Weinberg equilibrium exact test p < 0.05, and minor-allele frequency [MAF] < 0.05) were also omitted. Pre-phasing and imputation were performed using SHAPEIT2 (r837) and IMPUTE4 (r300.3) with 3.5KJPNv2 haplotype reference panels.^28,30,31 The variant positions in the genome were based on GRCh37.

PGS

To generate the PGS model, single-nucleotide polymorphism weights were estimated using a publicly available GWAS summary of 9,260 Japanese residencies²⁰ in jMorp³⁰ (ID: TGA000004). Candidate models were constructed as previously described³² using three methods: linkage disequilibrium (LD) pruning and thresholding (P+T), LDpred,³³ and PRS-CS.³⁴ For the P+T method, 24 models were generated under different LD pruning conditions using one of the six parameters for p values (1, 5 × 10⁻¹, 5 × 10⁻², 5 × 10⁻⁴, 5 × 10⁻⁶, or 5 × 10⁻⁸) and one of the four parameters for R² values (0.2, 0.4, 0.6, or 0.8). For LDpred, seven models were generated for ρ values (1, 0.3, 0.1, 0.03, 0.01, 0.003, and 0.001). For PRS-CS, six models were generated for each ϕ value (1 × 10⁻¹, 1 × 10⁻², 1 × 10⁻⁴, 1 × 10⁻⁶, 1 × 10⁻⁸, and “auto”). Individual PGSs were calculated using Plink2 (v.2.00a2LM)³⁵ for each model and were standardized to Z scores. Variants with low imputation quality (R² < 0.3) were excluded using the “--minimac3-r2-filter 0.3” option. In addition, to avoid potential sample overlap with the GWAS of total IgE,²⁰ individuals genotyped with the Illumina Infinium OmniExpressExome-8 BeadChip in the TMM CommCohort were excluded. Model fitness was assessed using Spearman’s rank correlation coefficient (ρ) between the PGS from each candidate model and the residuals of log-transformed and Z score-standardized total IgE levels, obtained from a linear regression model adjusted for age and sex in the TMM9K dataset.

For comparison, PGSs based on publicly available models from the PGS Catalog³⁶ were calculated using PLINK2 (v.2.0.0-a.6.9). PGSs based on GRCh38/hg38 positions were converted to GRCh37 positions using CrossMap.³⁷ Associations between each PGS and allergic phenotypes were evaluated using logistic regression adjusted for age, sex, and principal components (PCs) 1–10; in the analysis using the birth cohorts (BirThree 1 and 2), age was excluded from the adjustment.

Statistical analysis

A phenome-wide association study (PheWAS) of standardized PGSs was conducted using logistic regression for binomial phenotypes adjusted for age, sex, and the top 10 genotype PCs in R (v.4.2.1). The list of phenotypes was based on our previous study²⁵ and updated to characterize the PGS for IgE levels, including immune-related, inflammatory, and other diseases relevant to internal medicine. In contrast, relatively rare cancers, diseases of the sensory organs (e.g., macular degeneration and deafness), obesity, and hypertension were excluded, as they were considered less relevant to this objective. The results from each validation dataset were pooled using the R package “meta” (v.6.2-1) with a fixed-effect model. To account for multiple testing in the PheWAS analysis, p values were adjusted using the Benjamini-Hochberg procedure to control the false discovery rate (FDR). Associations with FDR-adjusted p values (q values) below 0.05 were considered statistically significant. The receiver operating characteristic (ROC) curve and area under the ROC curve (AUC) for PGS to detect phenotypes were estimated using the R package “pROC” (v.1.18.0) without adjustment. The confidence interval (CI) for AUC was estimated using the DeLong method.

For birth cohort data, the odds ratio (OR) was estimated for each PGS quintile using logistic regression with covariates, such as sex and the top 10 genotype PCs. A trend analysis was conducted for the standardized PGS using the same regression approach. The Venn diagram was drawn using the R package “gplots” (v.3.1.3.1). Incidence rates per 1,000 person months were calculated for the lowest-to-middle PGS groups (Q1–Q3), Q4, and Q5. Incidence rate ratios (IRRs) with 95% CIs were estimated using the metainc function in the R package “meta” (v.6.2-1) based on the summed number of events and person months in the exposure (Q4 or Q5) and the control (Q1–Q3) groups.

Mendelian randomization (MR) analysis was performed using the ivreg function in the “AER” package (v.1.2.14) on R (v.4.2.1), with adjustments for age, sex, and PC1–10. Instrumental variables (IVs) for serum IgE were selected from a previous GWAS²⁰ using LD clumping in PLINK2 (v.2.0.0-a.6.9) with the following options: “--clump-p1 5e-8 --clump-p2 5e-8 --clump-r2 0.001 --clump-kb 10000.” All IVs were confirmed to be significantly associated (F-statistics > 10) with log-transformed and standardized total IgE levels in the training dataset (TMM9K) using a linear regression model adjusted for sex, age, and PC1–10.

Causal mediation analysis was performed using the “mediation” package (v.4.5.0) in R (v.4.2.1). The mediator model was constructed using a binomial generalized linear model (GLM) with a probit link function, adjusting for PGS (exposure), sex, and the top PC1–10. The outcome model had the same structure as the mediator model but included the mediator as an additional covariate. The 95% CIs for the effect sizes (β) were estimated using a quasi-Bayesian approximation via Monte Carlo simulation with 2,000 iterations.

HLA imputation and association study

Genotyped data from JPAv2 were used for human leukocyte antigen (HLA) imputation following the procedure described in a previous study.³¹ Briefly, quality control was conducted using Plink2,excluding participants and genotypes with call rates of <0.99. Genotypes with a Hardy-Weinberg equilibrium exact test p < 1 × 10⁻¹⁰ and MAF < 20 were also omitted. The genotyped data were pre-phased using Eagle (v.2.4.1). The 54KJPN-HLA reference panel constructed by TMM,³⁸ formatted using the NomenCleaner and MakeReference programs in SNP2HLA from the HLA-TAPAS pipeline,^39,40 was used in this study. After converting the file format using bcftools (v.1.18)⁴¹ and savvy (v.2.0.1), the reference panel was used for the imputation of the pre-phased genotype data using Minimac4 (v.4.1.6).⁴² The HDS value, which is an estimated phased haploid alternate allele dosage, of the imputed alleles was used for downstream association.

For the association study, participants with call rates < 0.99 were excluded. Additionally, genotypes with a call rate < 0.99, Hardy-Weinberg equilibrium exact test p < 1 × 10⁻⁵, and MAF < 0.01 were excluded. Participants with close kinship within the second degree were excluded using the KING method,⁴³ and principal-component analysis (PCA) was conducted using LD-pruned variants in Plink2 using the “--indep-pairwise 1500 150 0.03” option.

Association tests were performed using BOLT-LMM (v.2.3.6)⁴⁴ for quantitative traits (such as common logarithmic total IgE) or SAIGE (v.0.44)⁴⁵ for binary traits (other phenotypes). Subjects estimated to be monozygotic (MZ)-twin or the same individual were excluded from the analysis. Imputed HLA alleles with an MAF of ≥0.01 and INFO >0.3 were analyzed, with age, sex, and 10 genotype PCs as covariates. To represent the association study, the direction of the effect was aligned with that of the minor allele as the effect allele. To account for multiple testing in the HLA association test, p values below the Bonferroni-corrected threshold (p = 0.05 divided by the number of alleles tested) were considered statistically significant.

Results

Model selection and validation

The study design is shown in Figure S1. The characteristics of the dataset used in this study are presented in Table 1. To evaluate the genetic predisposition for IgE production, we constructed a suitable polygenic model for the present study. A total of 37 models were generated for PGS using three different methods (Figure 1A). The PGS estimated from the model generated using PRS-CS with ϕ = 1 × 10⁻⁶ correlated best with the logarithmic total IgE levels after adjusting for age and sex (ρ = 0.200, p < 2.2 × 10⁻¹⁶). Therefore, we selected this model for the subsequent analysis.

Table 1.

Characteristics of the study population at baseline

Role	Model selection	Validation		Application to birth cohort
	TMM9K	TMM67K	TMM18K	BirThree 1	BirThree 2
n	7,482	47,766	11,698	9,425	7,729
Female, %	66.5	60.3	75.6	48.5	48.5
Age, year	56.4 ± 13.2	60.3 ± 11.4	58.0 ± 12.9	0	0
IgE, IU/mL	190.6 ± 568.0	189.9 ± 578.4	169.0 ± 514.6	N/D	N/D
log10IgEb	1.77 ± 0.67	1.77 ± 0.65	1.72 ± 0.64	N/D	N/D
Array	JPAv2	JPAv2	JPAv2	JPAv2	JPA-NEO
Residencya	Iwate	Iwate and Miyagi	Miyagi	Miyagi	Miyagi
Recruited	assessment center	health checkup	assessment center	assessment center	assessment center

N/D, no data available; IgE, immunoglobulin E; JPA-NEO, Japonica Array NEO; JPAv2, Japonica Array v.2; PGS, polygenic score; TMM, Tohoku Medical Megabank.

^aPrefecture in Japan.

^bCommon logarithmic IgE value.

Figure 1.

Optimized polygenic model for IgE correlates with total IgE levels

(A) The model with the highest Spearman’s ρ coefficient with logarithmic total IgE levels among the 37 candidate models identified as the best model. The horizontal dotted line in the PRS-CS (right) image indicates the ρ value for the model generated using the “auto” mode. The model generated by PRS-CS at ϕ = 1 × 10⁻⁶ showed the highest ρ.

(B) Bimodal distribution of standardized PGS calculated using the best model in the training dataset (TMM9K) and two validation datasets (TMM18K and TMM67K). The dotted curve indicates a normal distribution.

(C) Correlation of PGS in the validation datasets. PGS correlated significantly (p < 2.2 × 10⁻¹⁶) with logarithmic total IgE levels adjusted for age, sex, and the top ten genotype principal components (PCs) and then was standardized. Dotted lines indicate the diagonal. The color indicates the density of the subject.

IgE, immunoglobulin E; PGS, polygenic score.

The PGS distribution was bimodal (Figure 1B). In the TMM18K and TMM67K validation datasets, the PGS showed a significant correlation with log-transformed total IgE levels (ρ = 0.189 and 0.174, respectively; p < 2.2 × 10⁻¹⁶ in Spearman’s rank correlation test) after adjusting for age, sex, and the top 10 genotype PCs, followed by standardization (Figure 1C). In the best model, the proportions of variants within the major histocompatibility complex (MHC) region were 1.35% and 1.01% in the top and bottom percentile groups of effect weights, respectively (Figure S2A). Stratification of the TMM9K dataset by allele type with the top three effect weights (both positive and negative) clearly shifted the PGS distribution (Figures S2B and S2C).

PheWAS

A PheWAS of 63 phenotypes (Table S1) and subsequent meta-analysis of the results from two validation datasets (TMM67K and TMM18K) identified eight associations for PGS (Benjamini-Hochberg q < 0.05) (Figure 2A; Table S2). Among these, three were allergic disorders (cedar pollinosis, hives, and FAs), and two were autoimmune diseases (Graves’ disease [MIM: 275000] and Hashimoto’s disease [MIM: 140300]). A one standard deviation (SD) increase in PGS increased the risk of allergic disorders, such as cedar pollinosis (OR per SD: 1.10 [95% CI: 1.09–1.12]), hives (1.04 [1.02–1.07]), and FAs (1.10 [1.04–1.15]). Conversely, Graves’ and Hashimoto’s diseases exhibited protective effects (0.84 [0.78–0.90] and 0.90 [0.85–0.96], respectively).

Figure 2.

Phenome-wide association study of PGS and disease risk

(A) The association between binary phenotypes and standardized PGS analyzed using a logistic regression model adjusted for age, sex, and the top 10 genotype PCs in the genotype data. Phenotypes are colored according to the ICD-11 category. The q value indicates the adjusted p value obtained using the Benjamini-Hochberg method. The red horizontal line indicates q = 0.05. An increasing or decreasing odds ratio is indicated using an upward or downward arrowhead, respectively.

(B) Comparison of the risk analyzed using the same method based on the standardized common logarithmic total IgE instead of the PGS.

PC, principal component; IBD, inflammatory bowel disease; IgE, immunoglobulin E; PGS, polygenic score.

Subsequently, an analysis was conducted using total IgE levels instead of PGS for comparison (Figure 2B; Table S3). The three allergic disorders were also significantly correlated with the total IgE levels: cedar pollinosis (OR: 2.71 [95% CI: 2.65–2.77]), hives (1.35 [1.32–1.39]), and FAs (1.59 [1.51–1.68]). However, other conditions like Graves’ disease (1.07 [1.00–1.14]), Hashimoto’s disease (0.95 [0.90–1.01]), dental cavities (1.00 [0.98–1.01]), colon cancer (MIM: 114500) (1.02 [0.95–1.10]), and inflammatory bowel disease (IBD [MIM: 266600]) (0.98 [0.86–1.12]) did not show significant associations.

Furthermore, MR analyses were conducted using IVs significantly associated with total IgE levels (F-statistics > 10; Table S4). The pooled results indicated significant causal effects (Benjamini-Hochberg q < 0.05) of IgE levels on FAs, AR, cedar pollinosis, AS, and hives in the positive direction and on Graves’ disease and Hashimoto’s disease in the negative direction (Figure S3; Tables S5 and S6). In contrast, no causal association between IgE levels and colon cancer, dental cavities, or IBD was identified, suggesting different roles of IgE in these phenotypes.

For these diseases, PGS showed modest AUC values in both validation datasets (Figures S4 and S5), such as FAs (AUC in TMM18K and TMM67K: 0.549 [95% CI: 0.523–0.576] and 0.517 [0.500–0.535], respectively), cedar pollinosis (0.528 [0.517–0.539] and 0.527 [0.522–0.532]), and hives (0.518 [0.504–0.532] and 0.515 [0.506–0.523]).

HLA alleles influence PGS and allergic disease risk

To assess the contribution of HLA alleles to the polygenic model, we estimated their effects on the PGS and related phenotypes. HLA alleles were imputed with high quality (mean INFO value of approximately 1.0) for an MAF > 0.01 (Figure S6).

We then performed association analyses between imputed HLA alleles and various phenotypes. To validate the reliability of our association results prior to detailed interpretation, we compared them with previously reported findings identified through literature searches (Table S7). The direction of association identified in our study was consistent with that reported in previous studies for diseases such as Graves’ disease, Hashimoto’s disease, and IBD, supporting the robustness of our HLA association analyses.

The effects (β) of HLA alleles on total IgE levels, PGS, and disease risk were estimated through meta-analysis of the validation datasets (Figure 3; Tables S8–S10). A significant correlation was observed between the effects of PGS and total IgE levels (Figure 3A; Spearman’s ρ coefficient for all HLA allele types, ρ_n = 0.824). Most HLA alleles were significantly associated with IgE levels and PGS. Additionally, the effects of HLA alleles on PGS significantly correlated with their effects on phenotypes (Figure 3B), although a few HLA alleles were directly associated with specific phenotypes. These correlations were more pronounced when the alleles were significantly associated with both total IgE levels and PGS (ρ_s).

Figure 3.

Influence of HLA alleles on PGS and total IgE levels

(A) Effect sizes (β) of HLA alleles for common logarithmic total IgE or standardized PGS estimated using a linear mixed model with BOLT-LMM, adjusted for age, sex, and the top 10 genotype PCs. Dark blue and larger points represent effect sizes significantly associated with PGS and IgE levels, respectively, under Bonferroni correction. The shapes of the points indicate HLA classes. ρ_n and ρ_s indicate Spearman’s ρ coefficients across all HLA alleles in the analysis (89 alleles) and those significantly associated with PGS and total IgE levels (48 alleles), respectively.

(B) Effect sizes for the phenotypes estimated using SAIGE under the same conditions described above. The blue and gray areas indicate the approximate line obtained using the least-squares method and its 95% CI, respectively. ^∗p < 0.05, Spearman’s correlation test.

PC, principal component; HLA, human leukocyte antigen; IBD, inflammatory bowel disease; IgE, immunoglobulin E; MHC, major histocompatibility complex; PGS, polygenic score; CI, confidence interval.

Next, we simultaneously compared the effects of HLA alleles on the various phenotypes (Figure S7). Heatmap analysis revealed that phenotypes with similar correlations (positive Spearman’s ρ) with PGS, such as FAs, hives, and cedar pollinosis, shared common profiles. Clustering analysis showed that HLA alleles with higher effect sizes for these allergic diseases, total IgE levels, dental cavities, and IBD included A^∗24:02, B^∗52:01, C^∗12:02, DRB1^∗15:01 or DRB1^∗15:02, DQA1^∗01:02 or DQA1^∗01:03, DQB1^∗06:01 or DQB1^∗06:02, DPB1^∗09:01, and DPA1^∗02:01. In contrast, the relevance of these haplotypes was reduced in conditions such as colon cancer, Graves’ disease, and Hashimoto’s disease.

PGS predicts allergy risk in birth cohort study

The model developed for allergy risk prediction was applied to the TMM BirThree birth cohort (Figure 4A; Table S11). This cohort comprised two datasets, designated BirThree 1 and BirThree 2, which differed in genotyping arrays (JPAv2 and JPA-NEO, respectively).

Figure 4.

Allergy risk stratified by PGS quintile

(A) Standardized PGS distribution in the TMM BirThree cohort is represented by the timing of each survey after birth. The dotted vertical line and curve indicate the highest PGS quintile (Q5) and normal distribution, respectively.

(B) Allergy prevalence by PGS quintile in the 24 month (2 year) survey (left and middle). The OR was estimated using a logistic model with sex and the top 10 genotype PCs as covariates, with results pooled by meta-analysis using a fixed-effects model (right). Bars indicate 95% CIs.

(C) Incidence rate per 1,000 person months for the lowest-middle PGS groups (Q1–Q3), Q4, and Q5 (left and middle). Error bars indicate 95% CIs. IRRs were calculated using the Q1–Q3 group as the baseline (right), with 95% CIs indicated in square brackets.

(D) Prevalence of food allergies. The color codes show the PGS quintile as shown in (B). † indicates significance in trend analysis: †p < 0.05, ††p < 0.005, and †††p < 0.0005. ^∗ indicates significance in meta-analysis: ^∗p < 0.05, ^∗∗p < 0.005, and ^∗∗∗p < 0.0005.

CI, confidence interval; IRR, incidence rate ratio; OR, odds ratio; PC, principal component; PGS, polygenic score.

In both datasets, the prevalence of FAs in the 24 month survey exhibited a significant positive trend with increasing PGS quintiles (Figure 4B). Individuals in the highest PGS quintile (Q5) demonstrated a significantly elevated risk of FAs (OR: 1.95 [95% CI: 1.57–2.43]) and AD (1.33 [1.09–1.62]) in the meta-analysis of the two datasets (Figure 4B; Table S12). The incidence of affected individuals increased in subsequent surveys, with AD-affected individuals rising in the 6 month survey and FA-affected individuals increasing around the 12 month mark (Figure S8).

The incidence rates of FAs (per 1,000 person months) were 4.45 (95% CI: 3.74–5.24) in BirThree 1 and 4.26 (3.50–5.14) in BirThree 2 (Figure 4C; Table S13). The highest PGS group showed significant IRRs for FAs and AD: 1.51-fold (95% CI: 1.30–1.76) and 1.30-fold (1.12–1.51), respectively, compared with the lowest-to-intermediate (Q1–Q3) groups. Additionally, FA prevalence according to allergen type showed a significant trend in the PGS quintile (Figure 4D; Table S14). In the highest PGS group, the prevalence rates for egg white allergy were 7.39% (95% CI: 5.84–8.94) in BirThree 1 and 6.24% (95% CI: 4.66–7.83) in BirThree 2.

Furthermore, individuals in the highest (Q5) PGS group showed an increased risk of dual phenotypes during the follow-up period compared with those in the intermediate (Q3) group (Figure 5A). The highest PGS group exhibited significantly elevated risks for FAs and AS (OR in meta-analysis: 3.00 [95% CI: 1.30–6.93]), AD and FAs (1.89 [1.33–2.69]), and AS and AR/allergic conjunctivitis/pollinosis (2.40 [1.07–5.37]) compared to the lowest PGS group (Q1) (Figure 5B; Table S15). Compared with the lowest-to-intermediate PGS groups, the highest PGS group showed significant IRRs for FAs and AS, AD and FAs, and AD and AS (Figure 5C; Table S16).

Figure 5.

Risk of dual phenotypes in PGS quintiles

(A) Venn diagram of affected individuals during the follow-up period. The increased number of affected individuals in Q5 compared with Q3 is indicated in bold.

(B) OR estimated for individuals who developed both phenotypes during the OR follow-up period and were treated as affected individuals .

(C) Incidence and IRRs of both phenotypes. The incidence rates and IRRs of individuals with both phenotypes were calculated. Affected individuals were identified when the second phenotype was reported, and all other individuals were included in the control group. The estimation method and representation are the same as those shown in Figure 4. ^∗p < 0.05, ^∗∗p < 0.005, and ^∗∗∗p < 0.0005.

AD, atopic dermatitis; AR, atopic rhinitis; AS, asthma; CI, confidence interval; FA, food allergy; IRR, incidence rate ratio; OR, odds ratio; PGS, polygenic score.

The PGS effectively predicted the risk of FAs within a 2 year follow-up period, with AUC values of 0.559 (95% CI: 0.534–0.584) for BirThree 1 and 0.546 (95% CI: 0.518–0.574) for BirThree 2 (0.518–0.574) (Figure S9). Similarly, PGS predicted the risk of dual allergies (AD and FAs), with AUC values of 0.573 (95% CI: 0.528–0.618) for BirThree 1 and 0.553 (0.504–0.601) for BirThree 2 (Figure S10).

Interactions among the risks of atopic diseases were assessed using causal mediation analysis (Figure S11; Table S17). In models with significant total effects (q < 0.05), FAs demonstrated a significant mediation effect in the PGS-to-AD pathway, suggesting a potential causal interplay between these two allergic phenotypes.

To further characterize the PGS developed in this study, we compared it to publicly available PGS models (Figures S12–15; Tables S18–22). The resulting correlation matrix and phenotype association patterns indicated that PGS in this study showed a relatively low correlation with those in other studies and lacked distinctive associations with most individual phenotypes, except for FAs in the birth cohort and IgE levels.

Discussion

The present study demonstrated that PGS correlates significantly not only with total IgE levels but also with various immune-related phenotypes, including allergies. An analysis using birth cohort data indicated that PGS was associated with the risk of FAs and AD. Furthermore, individuals in the highest PGS quintile exhibited an elevated risk for both phenotypes. These findings suggest that the PGS for IgE levels may serve as a promising early predictor of atopic diseases.

Currently, few risk factors that predict allergy risk earlier than eczema have been identified.^46,47,48,49 A study of IgE sensitization at 4 years and comorbidity incidence at age 8 years found a 38% attributable risk, indicating that IgE levels alone may not fully explain the development of eczema, AR, and AS despite their shared mechanisms. The atopic march might involve a complex interplay of environmental and genetic factors rather than a straightforward increase in IgE levels.⁵⁰ This complexity may account for the moderate AUC of our PGS for atopic phenotypes despite its significant correlation with IgE levels.

Previous studies have explored the associations of genetic risk scores for pediatric allergies, such as those for AD^15,51 and AS¹⁴; however, longitudinal studies on non-European populations are limited. To our knowledge, this study represents the largest evaluation of PGS for allergies in a birth cohort. These disease-specific PGSs are often associated with the risk of other atopic diseases; however, their cross-disease applicability and potential biases introduced by shared environmental factors remain unexplored. In contrast, our PGS, which was specifically tuned to correlate with IgE levels, demonstrated unique phenotype associations differing from those observed with previously reported PGSs. These findings support that the model reported in this study offers a statistically robust "IgE-based" estimate of allergy risk without requiring further phenotype-specific optimization.

A population-based study conducted in the United States from 2015 to 2016 found that 7.6% of 38,408 children had FAs (95% CI: 7.1–8.1) based on parent-proxy responses.³ Among them, 10.0% (95% CI: 7.2–13.7) of children aged 2 years were reported to have at least one type of FA. Reliable epidemiological data have been limited in other regions, including Asia.¹³ In the present study, the FA prevalence in the TMM BirThree cohort in the 2-year-old survey was approximately 5.5% (see details in Table S8), which is considerably lower than that in the United States study.³ Differences in common allergens between the studies—such as milk (4.3% [95% CI: 2.2–8.3]), peanut (2.4% [95% CI: 1.6–3.6]), and egg (1.4% [95% CI: 0.8–2.4]) in the United States³ compared to lower rates of milk (∼1.5%), peanut (∼0.5%), egg white (∼5.3%), and egg yolk (∼3.3%) allergies in the BirThree 1 and 2 datasets—suggest regional variations. Our study suggests an association between the risk of egg white allergy and PGS, highlighting the need for comprehensive surveys that include genetic risks and recent food environments around children to better address allergy prevalence.

Current guidelines recommend that women should consume a healthy diet without eliminating or increasing the consumption of potentially allergenic foods during pregnancy or breastfeeding to prevent the development of FAs.^13,52 Additionally, the early introduction of allergens, such as peanuts⁵³ and hen eggs,⁵⁴ reduced the risk rather than avoidance.^55,56 The present PGS could help identify children at a higher genetic risk for FAs before severe symptoms develop, potentially guiding early intervention strategies.

Our analysis also evaluated the contribution of HLA alleles to PGS. We observed significant associations between the HLA alleles and PGS, total IgE levels, and phenotypes. This indicates that our model incorporates not only conventional genetic variants but also HLA allele effects. Moreover, total IgE levels and allergic diseases showed similar allele profiles. Among these, A^∗24:02-B^∗52:01-C^∗12:02-DRB1^∗15:02-(DRB5^∗0102, not determined in this study)-DQA1^∗01:03-DQB1^∗06:01-DPB1^∗09:01 and DRB1^∗15:01-DQB1^∗06:02 have been identified as haplotypes in a previous study⁵⁷ and could serve as targets for identifying causal genes.

Recent multiethnic studies have reported associations between HLA alleles and total IgE levels, such as A^∗02:01 (β: −0.06 [95% CI: −0.08 to −0.03] and −0.05 [−0.08 to −0.02] in discovery and replication analyses, respectively) and DQA1^∗03:01 (−0.07 [−0.10 to −0.03] and −0.08 [−0.12 to −0.05]).⁵⁸ These alleles were replicated in the present analysis, with β: −0.029 (95% CI: −0.034 to −0.024) and −0.032 (−0.035 to −0.029) for the meta-analysis of A^∗02:01 and DQA1^∗03:01, respectively. Considering the antigen specificity of HLA alleles and their requirement for antigen peptides to stimulate the immune response, these alleles may influence total IgE levels transethnically and transculturally in distinct antigen pools in different environments.

Future research should explore the association between PGS and nonallergic phenotypes. For example, dental anomalies and severe susceptibility to caries have been reported in patients with hyper-IgE syndrome (MIM: 147060).⁵⁹ IBD, colon cancer, Graves’ disease, and Hashimoto’s disease were not significantly associated with total IgE levels but were associated with PGS, coinciding with their mechanisms that do not directly lead to a Th2 response, accompanied by IgE production.

This study has several limitations. The TMM CommCohort dataset includes related individuals,⁶⁰ which may have influenced the results to a certain extent. The best model was selected and validated using Japanese genotype data imputed by the Japanese genotype panel. Therefore, additional validation is recommended before using other populations in future studies. Additionally, our phenotyping was based on self- or parent-proxy-reported surveys, which could lead to an overestimation of allergy prevalence and make it impossible to distinguish non-IgE-mediated FAs.

In conclusion, our study evaluated an approach using PGS for total IgE levels to estimate future atopic risks and elucidate the genetic contributions to these phenotypes. These findings offer important insights into the genetic factors that influence atopic diseases and provide a potential tool for early prediction and intervention.

Despite these promising results, further validation in other populations is required to ensure the robustness of the model. This approach could help inform early interventions for individuals at a high genetic risk of developing allergies, potentially guiding clinical decisions and preventive measures. Future research should focus on refining the PGS model for broader applications, including considering diverse populations and non-IgE-mediated allergies.

Acknowledgments

This study was supported by the Tohoku Medical Megabank (TMM) Project (Special Account for the Reconstruction of the Great East Japan Earthquake) of the Ministry of Education, Culture, Sports, Science and Technology (MEXT) and the Japan Agency for Medical Research and Development (AMED) under grant no. JP23tm0124006. The supercomputer system in the TMM Project under grant number JP21tm0424601 was used for the data analysis. Y.S. and A.S. were supported by the Japan Society for the Promotion of Science (JSPS) KAKENHI grant JP23K09747.

Author contributions

Conceptualization, Y.S. and A.S.; design of the work, Y.S. and T.H.; acquisition of data, Y.O.-Y., S. Komaki, S.M., H.O., N.N., A.N., M.O., M.I., T.O., K.T., A.H., and S. Kuriyama; analysis of data, Y.S.; interpretation of data, Y.S., T.H., Y.O.-Y., S. Komaki, S.M., H.O., and A.S.; writing – original draft, Y.S.; writing – review & editing, Y.S., T.H., Y.O.-Y., S. Komaki, S.M., H.O., K.T., A.H., N.N., A.N., M.O., M.I., T.O., S. Kuriyama, M.S., and A.S.

Declaration of interests

T.H. is a board member of Genome Analytics Japan, Inc.

Published: July 21, 2025

Web resources

• jMorp, https://jmorp.megabank.tohoku.ac.jp

• OMIM, https://www.omim.org

• Savvy, https://github.com/statgen/savvy