Genome-Wide Risk Assessment and Clinical Prediction of Iodinated Contrast Media–Induced Anaphylaxis

Abstract

Purpose

Iodinated contrast media (ICM) are widely used in diagnostic imaging, but may trigger immediate hypersensitivity reactions, including anaphylaxis. Although human leukocyte antigen (HLA) associations have been proposed, genome-wide evidence remains limited. To identify genetic risk variants associated with ICM-induced anaphylaxis and evaluate their clinical utility.

Methods

We conducted a genome-wide association study (GWAS) in 91 Korean patients with confirmed ICM-induced anaphylaxis (discovery, n = 67; replication, n = 24) and 1,565 controls. Genotyping used a Korean-specific single nucleotide polymorphism (SNP) array. Significant variants were replicated. HLA alleles were imputed in the combined dataset. Linkage disequilibrium (LD) among significant loci was analyzed. Cumulative genetic burden was assessed by summing risk alleles per individual. A Least Absolute Shrinkage and Selection Operator-based prediction model incorporated genetic and clinical variables.

Results

Nineteen SNPs reached genome-wide significance and were replicated, including rs2229629 in HK2 (odds ratio, 5.02; P = 9.6 × 10⁻²³), suggesting altered glycolysis. Additional variants were found near GABBR1, ZNF391, ATAT1, and HLA genes. Imputed HLA-DRB1*15:02 was significantly enriched in patients (41.8% vs. 5.9%, P = 5.21 × 10⁻¹⁷). LD analysis showed haplotypic clustering within HLA and non-HLA loci. A gene dosage effect was observed. The combined genetic-clinical model showed the highest predictive performance (area under the curve [AUC] = 0.888), exceeding genetic (AUC = 0.800) and clinical (AUC = 0.835) models.

Conclusions

This first GWAS of ICM-induced anaphylaxis identifies susceptibility loci across metabolic, immune, and neuroimmune pathways. HK2 was most strongly associated, and HLA-DRB1*15:02 was robustly replicated.

INTRODUCTION

Iodinated contrast media (ICM) are essential in modern diagnostic imaging, commonly used in computed tomography (CT) and angiographic procedures to enhance the visualization of vascular structures and soft tissues. The introduction of non-ionic low-osmolar ICM has markedly reduced the overall incidence of adverse drug reactions (ADRs), with current rates estimated at 1%–3% for all ADRs and 0.01%–0.04% for serious hypersensitivity reactions (HSRs), including anaphylaxis.¹

Despite improved safety profiles, the absolute number of ICM-related ADRs has increased substantially, driven by the rising use of imaging technologies. In South Korea, reported ICM-related ADRs increased more than tenfold from 1,688 in 2009 to 20,604 in 2019.² Among these, moderate to severe reactions remain a major concern due to their potential for life-threatening outcomes. In particular, ICM-induced anaphylaxis poses a significant clinical challenge, underscoring the need for improved strategies to predict, manage, and prevent such events.³,⁴,⁵

Several risk factors for ICM-induced anaphylaxis have been reported. A prior history of ICM HSRs, especially if severe, is a strong predictor, as reactions often recur with similar severity. Other clinical risk factors include allergic conditions (e.g., asthma, allergic rhinitis, and chronic urticaria) and HSRs to other drugs.⁶,⁷ Skin test positivity has been more frequent in patients with severe reactions than in those with milder forms.⁸ However, these factors alone are insufficient to identify all at-risk individuals, highlighting the need for more robust predictive tools.

A family history of ICM hypersensitivity is associated with up to a 14-fold increased risk,⁹ suggesting a potential genetic component. Human leukocyte antigen (HLA) alleles are established genetic markers for drug hypersensitivity, especially in delayed-type T cell–mediated reactions such as severe cutaneous adverse reactions.¹⁰,¹¹,¹² Emerging evidence also suggests a genetic basis for immediate HSRs to ICM. In a previous study, we identified a strong association between HLA-DRB1*15:02 and ICM-induced anaphylaxis in Koreans, with the allele five times more frequent in patients than in controls.¹³ Another study reported HLA-B*58:01 to be associated with mild to moderate immediate HSRs.¹⁴ While these findings highlight the potential of HLA alleles as genetic risk markers for ICM HSRs, large-scale replication studies are limited, and the contribution of non-HLA loci remains largely unexplored.

Genome-wide association studies (GWAS) provide an unbiased approach to identifying common variants linked to complex traits, including immune-mediated drug hypersensitivity.¹⁵,¹⁶ Compared with candidate gene studies, GWAS enables systematic interrogation of the genome and has revealed multiple risk alleles across various drug reactions. Microarray-based gene chip technology further allows efficient, high-throughput genotyping of hundreds of thousands of single nucleotide polymorphisms (SNPs), supporting large-scale pharmacogenetic research.

Given this background, we conducted the first GWAS using microarray-based genotyping to identify susceptibility loci for ICM-induced anaphylaxis in Koreans. We hypothesized that both immune-related and non-immune genetic variants contribute to disease risk. By integrating genetic data with clinical features from discovery and replication cohorts, we aimed to evaluate their combined value for risk prediction.

MATERIALS AND METHODS

Study design and participants

This GWAS study was designed to minimize potential sources of bias through a two-phase approach comprising independent discovery and replication cohorts. Strict eligibility criteria were applied to ensure diagnostic accuracy, and cases and controls were matched by age and sex, which were included as covariates to adjust for residual confounding. Standardized definitions of ICM-induced anaphylaxis were used across both phases. To reduce population stratification, only ethnically homogeneous Korean individuals were included.

For the discovery phase, newly identified cases of ICM-induced anaphylaxis were prospectively enrolled from Seoul National University Hospital and Seoul National University Bundang Hospital between 2018 and 2023. For the replication phase, patients with ICM-induced anaphylaxis were selected from the Korean Registry of Iodinated Contrast Media Adverse Reactions, which included cases enrolled by seven tertiary hospitals between March and October of 2017 as part of a previous multicenter study conducted by our group.¹³ Only patients for whom both biospecimens and clinical data were available and suitable for analysis were included.

Demographic and clinical data (including age, sex, culprit contrast media, clinical symptoms, and skin test results) were extracted from electronic medical records. Anaphylaxis diagnoses were confirmed by board-certified allergists according to the established diagnostic criteria of the World Allergy Organization.¹⁷

Controls were randomly selected from the general population cohort of the National Biobank of Korea using a computer-generated random number algorithm, with frequency matching to cases by age and sex. A total of 1,224 and 341 controls were matched to 67 and 24 cases in the discovery and replication phases, respectively. The use of population-based controls aligns with recommended practices for rare disease studies and is supported by precedents from large-scale biobanks.¹⁸,¹⁹ Defining true tolerance to ICM is particularly challenging, as HSRs may occur not only at first exposure but also after multiple exposures, even after the tenth administration over many years. Given the difficulty in reliably identifying tolerant individuals, the use of general population controls was considered appropriate. The enrollment and grouping process is illustrated in Fig. 1.

Fig. 1. Flowchart of the enrollment process.

SNUH, Seoul National University Hospital; SNUBH, Seoul National University Bundang Hospital; ICM, iodinated contrast media; ADR, adverse drug reaction.

Genotyping and quality control

This study adhered to the STrengthening the REporting of Genetic Association Studies (STREGA) reporting guidelines. Genomic DNA was extracted from venous blood, and 200 ng per individual was genotyped using the Affymetrix Axiom KORV1.1-96 Array (Affymetrix, Santa Clara, CA, USA). DNA quality and concentration were confirmed using NanoDrop spectrophotometry and Qubit fluorometric quantification. Genotyping was performed using the Korean Chip (K-CHIP) v1.0, a SNP array optimized for the Korean population based on whole-genome sequencing of 2,576 individuals. Hybridization and scanning were carried out on the GeneTitan^® MC Instrument following the Axiom^® 2.0 Assay protocol, and data were analyzed with the Axiom Analysis Suite v4.0. Genotyping procedures and reporting followed the Minimum Information About a Microarray Experiment (MIAME) guideline. Quality control was performed using PLINK v1.9, excluding samples with sex inconsistency or call rate ≤ 97%, and SNPs with call rate < 95%, minor allele frequency (MAF) ≤ 0.01, or Hardy–Weinberg P < 1 × 10⁻⁶. Missing genotypes were excluded based on call rate thresholds, and phenotype data were complete. Cases and controls were genotyped separately but using identical platforms, protocols, and pipelines to minimize batch effects. Genotype data is available at reasonable requests to the corresponding author.

Genetic association analysis

SNPs were annotated using dbSNP build 151 (GRCh37) and analyzed using logistic regression in PLINK v1.9 under an additive model. Genotypes were coded as 0, 1, or 2 based on risk allele count. Although cases and controls were matched by age and sex to reduce demographic imbalance, these variables were also included as covariates in the model to account for residual confounding, as age and sex are generally considered relevant to disease susceptibility in many conditions, including anaphylaxis; age was continuous, sex binary (male vs. female). Results are reported as odds ratios (ORs) with 95% confidence intervals (CIs). Significant SNPs in the discovery phase were tested in the replication phase (P < 0.05) and results combined. Study size was determined by case availability; post hoc power was evaluated using Quanto v1.2.4. Manhattan plots were generated using fastman (R v4.3.1). Due to the homogeneous Korean ancestry of participants, population stratification was considered minimal, and principal component analysis was not applied. To further assess potential population stratification, we generated a quantile–quantile (QQ) plot (Supplementary Fig. S1) and calculated the genomic inflation factor (λGC). The QQ plot showed that the observed test statistics closely followed the expected null distribution, with deviation only in the extreme tail. The genomic inflation factor was λGC = 0.95, indicating no evidence of test statistic inflation.

Functional annotation

To assess biological relevance, in silico functional annotation was performed using the National Center for Biotechnology Information Gene database and the Ensembl genome browser to identify nearby genes and evaluate SNP functions.

HLA Imputation and Association Analysis

We extracted genotype information from 29 Mb to 34 Mb on chromosome 6 using PLINK (v1.90b6.10).²⁰ HLA alleles were then imputed for 91 case samples and 1,565 control samples using the CookHLA program (v1.1)²¹ with default settings. Imputation quality was evaluated using the confidence scores generated by CookHLA, and their distribution across loci is provided in Supplementary Fig. S2 and Supplementary Table S1. The 1000 Genomes Project data²² was used as the reference panel for HLA imputation and genetic map construction. Fisher’s exact test was performed using the SciPy library (v1.13.1) in Python.

Linkage disequilibrium (LD) analysis

LD analysis was performed in PLINK. Regional association plots were generated by LocusZoom.

Cumulative genetic burden analysis

Cumulative genetic risk was assessed by summing risk alleles per individual across SNPs with significant associations in discovery. Genotypes were scored as 0 (homozygous reference), 1 (heterozygous), or 2 (homozygous risk). Distributions were compared between cases and controls using descriptive statistics. Heatmaps visualized SNP co-occurrence, with individuals as rows and SNPs as columns. The genotypes were color-coded: black (homozygous risk), dark gray (heterozygous), light gray (non-carrier), and white (missing).

Prediction modeling of ICM-induced anaphylaxis

To explore the combined effects of clinical and genetic factors, a least absolute shrinkage and selection operator (LASSO) regression model was developed using the combined dataset. The regularization parameter (λ) was optimized through 10-fold cross-validation using the caret package in R v4.4.2. Variables included in the model (age, sex, history of asthma, underlying comorbidities, and risk SNPs identified through GWAS) were selected based on their availability in both cases and controls. To avoid circular reasoning, other allergic diseases that could be influenced by or indicative of ICM-induced anaphylaxis were excluded from the model. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity.

Statistical analysis

General comparisons were conducted in R. Continuous variables were tested using t-tests, and categorical variables using the χ² test or Fisher’s exact test.

Ethics statement

This study was conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained from all participants. The protocol was approved by the Institutional Review Boards (IRBs) of Seoul National University Bundang Hospital (B-1203-147-401) and Seoul National University Hospital (1705-012-851) for discovery-set cases, and by the Seoul National University Hospital IRB for replication-set cases and controls.

RESULTS

Clinical characteristics of the study population

A total of 91 patients with ICM-induced anaphylaxis were analyzed (discovery: n = 67; replication: n = 24). Baseline characteristics were largely similar, with a non-significant age difference (P = 0.132) and comparable rates of male sex, respiratory symptoms, and hypotension. However, organ involvement varied: skin symptoms were more common in the discovery group (P = 0.004), while gastrointestinal and cardiovascular symptoms were more frequent in the replication group (P = 0.001 and P = 0.018, respectively). No significant differences were observed in allergic disease history, underlying comorbidity, culprit ICM, or skin test outcomes. Detailed data are summarized in Table 1, and no variables had missing values, supporting replication reliability.

Table 1. Clinical characteristics of patients with ICM-induced anaphylaxis in the discovery and replication cohorts.

Characteristics		Discovery group (n = 67)	Replication group (n = 24)	P value
Age (yr)		57.3 ± 13.5	63.0 ± 16.2	0.132
Male		26 (38.8)	11 (45.8)	0.719
Anaphylaxis symptoms by organ system
	Skin	49 (76.6)	10 (41.7)	0.004
	Respiratory	33 (49.3)	9 (37.5)	0.452
	Cardiovascular	54 (80.6)	24 (100.0)	0.018
	Hypotension	49 (87.5)	21 (100.0)	0.180
	Gastrointestinal	5 (7.5)	9 (37.5)	0.001
Underlying allergic disease		23 (35.4)	4 (16.7)	0.148
Allergic disease type
	Asthma	3 (4.5)	0 (0.0)
	Drug allergy	16 (23.9)	4 (16.7)
	Food allergy	3 (4.5)	0 (0.0)
	Others	3 (4.5)	0 (0.0)
Underlying comorbidity		63 (94.0)	18 (75.0)	0.232
	Malignancy	43 (64.2)	16 (66.7)
	Autoimmune diseases	1 (1.5)	1 (4.2)
	Cardiovascular diseases	18 (26.9)	2 (8.3)
	Metabolic diseases	9 (13.4)	1 (4.2)
	Kidney and liver diseases	5 (7.5)	1 (4.2)
	Cerebrovascular diseases	4 (6.0)	0 (0.0)
	Others	5 (7.5)	0 (0.0)
Culprit ICM				0.184
	Iohexol	19 (28.4)	4 (16.7)
	Iopamidol	13 (19.4)	7 (29.2)
	Iomeprol	10 (14.9)	1 (4.2)
	Iopromide	8 (11.9)	6 (25.0)
	Iobitridol	7 (10.4)	2 (8.3)
	Ioversol	4 (6.0)	0 (0.0)
	Iodixanol	1 (1.5)	1 (4.2)
	Two or more causative ICM	3 (4.5)	0 (0.0)
	Unknown	2 (3.0)	3 (12.5)
Previous ICM hypersensitivity history		18 (26.9)	4 (16.7)	0.469
ICM skin test		56 (83.6)	16 (66.7)	0.145
	Positive response to at least one ICM	46 (82.1)	13 (81.2)	1.000
	Positive response to culprit ICM	32 (59.3)	10 (58.8)	1.000

Data are presented as the mean ± standard deviation or number (%).

ICM, iodinated contrast media.

Genome-wide association analysis

Following quality control procedures, 424,242 autosomal SNPs were retained for analysis. The discovery set (67 cases, 1,224 controls) had 80% power to detect an OR ≥ 2.01 for MAF = 0.1 and OR ≥ 1.68 for MAF = 0.3 under a log-additive model. In the discovery phase, 20 SNPs surpassed the Bonferroni-corrected threshold for genome-wide significance (P < 1.18 × 10⁻⁷). Excluding rs9269878, 19 SNPs were successfully replicated (P < 0.05) in the independent replication set. The most significant signal was observed for rs2229629 within or near the HK2 gene (OR, 5.775; P = 1.97 × 10⁻¹⁹ in the discovery phase; OR, 3.299; P = 1.67 × 10⁻⁴ in the replication phase). Genes involved in glucose metabolism (HK2), neuroimmune signaling (GABBR1), cytoskeletal function (ATAT1), and transcriptional regulation (ZNF391, ZSCAN23 and ZSCAN12, TCF19) were identified as significant susceptibility factors in our study.

In addition, associated loci were identified not only in non–immune-related genes as above but also in classical immune-related genes, including HLA-B, HLA-C, HLA-DRB1, HLA-DQB1, HLA-DQA2. The Manhattan plot of GWAS results is shown Fig. 2 and the full list of replicated SNPs is provided in Table 2.

Fig. 2. Manhattan plot of genome-wide association study results. The x-axis represents single nucleotide polymorphism markers across chromosomes. The y-axis shows the −log10 P value from logistic regression. The red horizontal line indicates the Bonferroni-adjusted significance threshold (P = 1.18 × 10⁻⁷), while the blue horizontal line represents the genome-wide suggestiveness threshold (P = 1.0 × 10⁻⁴).

Table 2. Logistic regression analysis results of the genome-wide association study for iodinated contrast media anaphylaxis.

SNP	Chr	Position	Risk allele (M)	Discovery set			Replication set			Combined set
				MAF	OR (95% CI)	P value	MAF	OR (95% CI)	P value	MAF	OR (95% CI)	P value
rs2229629	2	75116527	A	0.49/0.13	5.78 (3.94–8.46)	1.97 × 10−19	0.35/0.13	3.30 (1.77–6.14)	1.67 × 10−4	0.45/0.13	5.02 (3.64–6.92)	9.58 × 10−23
rs58690463	6	27368874	C	0.10/0.02	6.46 (3.24–2.86)	1.11 × 10−7	0.12/0.02	10.26 (3.35–31.44)	4.62 × 10−5	0.10/0.02	7.09 (3.99–12.61)	2.57 × 10−11
rs3757181	6	28381896	G	0.10/0.02	6.44 (3.24–12.79)	1.10 × 10−7	0.12/0.02	9.30 (3.08–28.05)	7.51 × 10−5	0.10/0.02	6.99 (3.94–12.42)	3.32 × 10−11
rs55650674	6	29571180	A	0.12/0.03	5.65 (3.04–10.52)	4.63 × 10−8	0.15/0.03	5.50 (2.11–14.28)	4.72 × 10−4	0.13/0.03	5.57 (3.32–9.34)	7.73 × 10−11
rs2076482	6	29573254	A	0.12/0.03	5.73 (3.07–10.66)	3.80 × 10−8	0.15/0.03	5.79 (2.20–15.20)	3.65 × 10−4	0.13/0.03	5.72 (3.40–9.60)	4.48 × 10−11
rs2076489	6	29576437	G	0.11/0.02	7.05 (3.67–13.54)	4.38 × 10−9	0.12/0.02	7.86 (2.68–23.05)	1.71 × 10−4	0.12/0.02	7.10 (4.09–12.34)	3.67 × 10−12
rs117993568	6	30612861	G	0.12/0.02	7.43 (3.93–14.04)	6.78 × 10−10	0.15/0.02	12.35 (4.25–35.87)	3.85 × 10−6	0.13/0.02	8.12 (4.74–13.93)	2.69 × 10−14
rs17190769	6	31127037	T	0.14/0.04	4.54 (2.60–7.94)	1.12 × 10−7	0.15/0.03	6.01 (2.25–16.04)	3.45 × 10−4	0.14/0.04	4.73 (2.92–7.66)	2.90 × 10−10
rs2844586	6	31318024	A	0.14/0.04	4.59 (2.64–7.96)	6.18 × 10−8	0.15/0.04	4.59 (1.75–12.01)	1.94 × 10−3	0.14/0.04	4.53 (2.81–7.29)	5.18 × 10−10
rs549182	6	32205045	A	0.22/0.08	3.52 (2.22–5.57)	8.70 × 10−8	0.25/0.09	3.17 (1.58–6.34)	1.13 × 10−3	0.23/0.08	3.42 (2.33–5.01)	3.35 × 10−10
rs549270	6	32205072	T	0.22/0.08	3.52 (2.22–5.58)	8.28 × 10−8	0.25/0.09	3.17 (1.58–6.34)	1.13 × 10−3	0.23/0.08	3.42 (2.33–5.02)	3.20 × 10−10
rs58302219	6	32430368	A	0.22/0.07	5.06 (3.07–8.33)	1.97 × 10−10	0.23/0.09	3.16 (1.50–6.64)	2.41 × 10−3	0.23/0.07	4.38 (2.90–6.61)	2.04 × 10−12
rs9269910	6	32551350	T	0.28/0.12	3.01 (2.00–4.52)	1.15 × 10−7	0.33/0.11	3.90 (1.98–7.68)	8.16 × 10−5	0.30/0.12	3.21 (2.27–4.54)	4.73 × 10−11
rs9271413	6	32587768	G	0.28/0.11	3.15 (2.09–4.74)	4.29 × 10−8	0.33/0.10	4.16 (2.12–8.18)	3.48 × 10−5	0.30/0.11	3.40 (2.40–4.82)	6.55 × 10−12
rs6915986	6	32683898	C	0.16/0.03	9.17 (5.10–16.47)	1.23 × 10−13	0.17/0.02	13.78 (4.93–38.51)	5.60 × 10−7	0.16/0.02	9.73 (5.88–16.08)	7.76 × 10−19
rs3104398	6	32685685	A	0.19/0.07	3.91 (2.39–6.41)	6.31 × 10−8	0.25/0.05	5.64 (2.59–12.26)	1.31 × 10−5	0.21/0.06	4.28 (2.83–6.46)	4.78 × 10−12
rs3104401	6	32687358	G	0.20/0.07	3.98 (2.45–6.46)	2.28 × 10−8	0.25/0.06	5.41 (2.50–11.71)	1.88 × 10−5	0.21/0.06	4.29 (2.86–6.44)	2.28 × 10−12
rs79517313	6	32713044	T	0.15/0.02	9.07 (5.00–16.45)	3.95 × 10−13	0.15/0.02	11.30 (3.94–32.35)	6.32 × 10−6	0.15/0.02	9.23 (5.52–15.41)	2.18 × 10−17
rs9784876	6	32788878	A	0.15/0.02	9.36 (5.15–17.01)	2.11 × 10−13	0.15/0.02	9.65 (3.44–27.06)	1.66 × 10−5	0.15/0.02	9.27 (5.54–15.48)	1.98 × 10−17

For each significant SNP, minor-allele frequencies (case/control) and ORs with 95% CIs are presented for the discovery, replication, and combined datasets.

SNP, single-nucleotide polymorphism; Chr, chromosome; MAF, minor allele frequency; OR, odds ratio; CI, confidence interval.

LD among associated SNPs

LD analysis of 18 SNPs on chromosome 6 revealed distinct clustering within the HLA class II region. Notably, rs6915986 (HLA-DQB1, DQA2), rs79517313 (HLA-DQA2), and rs9784876 (HLA-DOB, TAP2) exhibited strong pairwise LD (r² > 0.700), forming a broad haplotype block. Similarly, rs9269910 (HLA-DRB1) and rs9271413 (HLA-DRB1, DQA1) were tightly linked (r² = 0.898). These clusters corresponded to dense regional signals in association plots (Supplementary Fig. S3).

Outside the HLA region, strong LD was also observed between rs58690463 (ZNF391) and rs3757181 (ZSCAN23, ZSCAN12), and between rs2076489 (GABBR1) and rs117993568 (ATAT1), all with r² > 0.700. These results suggest that susceptibility alleles may act in haplotypic units rather than independently. Detailed LD pairings are listed in Supplementary Table S2.

HLA imputation and association analysis

Comparison of HLA allele frequencies between 91 cases and 1,565 controls identified several significant associations (Supplementary Table S3). DRB1*15:02 showed the strongest association (41.8% vs. 5.9%; OR, 8.70; Pc = 5.21 × 10⁻¹⁷), followed by B*52:01 (OR, 7.94), C*12:02 (OR, 6.95), and DQB1*06:01 (OR, 2.67). All of these variants retained significance after Bonferroni correction for 122 loci. Conversely, several alleles, including DQB1*03:02, C*04:01, and B*13:01, showed reduced frequencies in patients, suggesting potential protective effects; however, these associations did not remain significant after correction.

Cumulative genetic burden analysis

A cumulative burden of risk alleles was observed in affected individuals. The mean number of risk alleles per patient was 7.20 ± 7.79 in the ICM-anaphylaxis group, compared to 2.00 ± 3.51 in controls. This gene dosage effect was visualized in a genotype heatmap demonstrating the aggregate genetic load associated with disease susceptibility (Fig. 3).

Fig. 3. Heatmaps of genotype frequencies among subjects: (A) patients with ICM-induced anaphylaxis (n = 91), and (B) control individuals (n = 1,565). Each row represents an individual, and each column corresponds to one of the 19 SNPs significantly associated with ICM anaphylaxis. The genotype at each SNP position is indicated by color: black for homozygous risk allele, dark gray for heterozygous, light gray for homozygous reference or non-carrier, and white for missing data. These heatmaps illustrate the distribution and co-occurrence patterns of risk alleles across individuals, suggesting a stronger gene dosage effect in patients with ICM-induced anaphylaxis.

ICM, iodinated contrast media; SNP, single nucleotide polymorphism.

Prediction modeling of ICM-induced anaphylaxis

Genetic model

A prediction model incorporating only the presence or absence of risk SNPs demonstrated moderate discriminative ability, with an AUC of 0.800. The model showed high overall accuracy (0.95) and specificity (1.00), but sensitivity remained low (0.11). The most influential genetic predictors in this model included rs2229629 (HK2), rs6915986 (HLA-DQB1 and DQA2), rs9271413 (HLA-DRB1 and DQA1), and rs58302219 (HLA-DRB5 and DRA) (Fig. 4A).

Fig. 4. Variable importance plots for prediction models. (A-C) Variable importance plots derived from the least absolute shrinkage and selection operator model, ranked by standardized coefficients: (A) genetic model, (B) clinical model, and (C) combined genetic-clinical model. Genetic predictors (HK2, HLA-DQB1, HLA-DRB1, ZNF391) and clinical factors (malignancy, renal disease) contributed most prominently to model discrimination.

Clinical model

A prediction model based solely on clinical variables demonstrated moderate discriminative ability, with an AUC of 0.835. Despite its high overall accuracy (0.95) and specificity (1.00), the model was limited by extremely low sensitivity (0.01). Malignancy, renal diseases, and old age emerged as the most influential clinical predictors (Fig. 4B).

Combined genetic-clinical model

The combined genetic-clinical model demonstrated the highest overall performance. Among clinical predictors, malignancy and renal disease were identified as key factors. The most influential genetic variant was rs2229629 (HK2), followed by rs6915986 (HLA-DQB1 and DQA2), rs9271413 (HLA-DRB1 and DQA1), and rs58690463 (ZNF391) (Fig. 4C). The combined model achieved an AUC of 0.888 and an accuracy of 0.96, outperforming both the genetic-only model and the clinical-only model (Fig. 5). While sensitivity remained moderate (0.40), the combined model exhibited specificity of 1.00 at the selected threshold. Calibration analysis demonstrated reasonable agreement between predicted and observed risks, with a calibration slope of 1.24 (Supplementary Fig. S4). Decision-curve analysis further showed that the combined model provided higher net benefit than the genetic-only model and comparable benefit to the clinical model across threshold probabilities of 5%–60%, representing clinically relevant decision-making ranges (Supplementary Fig. S5).

Fig. 5. Receiver-operating characteristic curves for the three predictive models. The combined genetic-clinical model (red) demonstrated the highest discrimination (AUC = 0.888), outperforming the genetic-only (blue, AUC = 0.800) and clinical-only (green, AUC = 0.835) models. Perfect specificity (1.00) and improved sensitivity (0.40) were observed in the combined model.

AUC, area under the curve.

DISCUSSION

This GWAS delineates the complex genetic architecture underlying ICM-induced anaphylaxis, a rare but potentially life-threatening HSR. We identified 19 SNPs associated with disease risk, spanning immune, metabolic, neuroimmune, and cytoskeletal pathways, consistent with a polygenic susceptibility model. Genotyping was conducted using the Korean-specific K-CHIP platform, which has been validated in over 100 peer-reviewed SCI-indexed studies as of May 2024.²³,²⁴ Given the exploratory nature of this study and its modest sample size, these findings should be interpreted as hypothesis-generating rather than confirmatory.

Among the top signals, rs2229629 in or near HK2 (encoding hexokinase-2, a key glycolytic enzyme) showed the strongest association and emerged as the most influential variable in the LASSO prediction model. Although the functional relevance of this variant remains uncertain, HK2 is known to regulate cellular energy metabolism and has been implicated in immune cell activation in experimental studies.²⁵,²⁶,²⁷,²⁸ While our findings alone cannot establish a causal link between glycolysis and anaphylaxis, they suggest a potential metabolic component that warrants further investigation. The observation is broadly consistent with recent multi-omics reports describing glycolytic reprogramming in severe allergic inflammation.²⁸ However, this interpretation remains hypothesis-generating, and future studies incorporating pathway or enrichment analyses will be needed to clarify whether metabolic dysregulation contributes mechanistically to ICM-induced hypersensitivity.

We also identified three significant SNPs in or near GABBR1, which encodes a gamma-aminobutyric acid receptor subunit involved in inhibitory neuroimmune signaling. While GABAergic activity has been linked to anti-inflammatory responses in asthma and airway allergy,²⁹,³⁰ its role in anaphylaxis is less defined. Thus, our results should be viewed as exploratory evidence suggesting, but not confirming, a possible neuroimmune component in ICM hypersensitivity.

Variants in or near ZNF391, ZSCAN23, ZSCAN12, TCF19, and ATAT1 implicate transcriptional regulation and cytoskeletal remodeling. Although these genes have not been previously associated with allergy, both transcriptional control and cytoskeletal dynamics are known to modulate mast cell degranulation,³¹,³² suggesting plausible mechanistic links.

Multiple SNPs were located in the HLA region, spanning class I (e.g., rs2844586 near HLA-B and C) and class II loci (e.g., rs9269910 and rs9271413 near HLA-DRB1). Although class I alleles are typically associated with delayed-type hypersensitivity,¹⁰,¹¹,¹²,³³,³⁴,³⁵ growing evidence, including our prior study,¹³ points to a role in immediate ICM reactions, possibly through direct drug-HLA binding or innate-like T cell activation.³⁶,³⁷,³⁸ Consistent with prior data, class II variants in HLA-DRB1 and DQA2 are implicated in immunoglobulin E- and T cell–mediated HSRs.³⁹,⁴⁰,⁴¹,⁴²,⁴³ Notably, HLA-DRB1*15:02, previously identified by our group,¹³ was robustly replicated and showed the strongest enrichment among HLA alleles, underscoring its relevance as an immunogenetic marker. Nevertheless, as HLA genotypes were imputed rather than directly sequenced, these associations require validation using high-resolution genotyping in future studies.

LD was observed among several non-HLA loci, including GABBR1, ATAT1, and TCF19, which are functionally related to neuroimmune signaling, cytoskeletal regulation, and transcriptional control. These observations suggest potential shared pathways of immune dysregulation, but additional replication and experimental work are needed to confirm their biological significance.

To assess the functional relevance of each SNP, we performed in silico annotation using publicly available databases, along with biological interpretation of nearby genes. Most variants were located in intronic or non-coding regions, suggesting roles in transcriptional regulation, splicing, or mRNA stability. Table 3 summarizes variant-level functional annotations and proposed mechanisms through which nearby genes may contribute to ICM-induced anaphylaxis.

Table 3. Summary of risk SNPs and gene functions in ICM anaphylaxis.

SNP	Closest gene	Functional annotation of SNPs	Putative function of the gene in ICM anaphylaxis
rs2229629	HK2	Missense variant; intron variant; non- coding transcript variant	May alter immune cell energy metabolism by regulating HK2 expression and glycolytic activity, promoting enhanced pro-inflammatory activation of T cells and macrophages, thereby increasing susceptibility to anaphylaxis.
rs58690463	ZNF391	Missense variant	May regulate transcriptional programs involved in mast cell activation.
rs3757181	ZSCAN23, ZSCAN12	Intron variant; non-coding transcript variant	May regulate transcriptional programs involved in mast cell activation.
rs55650674	GABBR1	3’ UTR variant; intron variant	Impaired GABAergic anti-inflammatory signaling via neuroimmune pathways may shift immune response toward a proinflammatory state, increasing susceptibility to anaphylaxis.
rs2076482	GABBR1	Intron variant
rs2076489	GABBR1	Missense variant
rs117993568	ATAT1	Intron variant	May contribute to ICM anaphylaxis by affecting cytoskeletal dynamics involved in mast cell degranulation.
rs17190769	TCF19	5’ UTR variant	May regulate transcriptional programs involved in mast cell activation.
rs2844586	HLA-B, HLA-C	Intron variant; non-coding transcript variant	May facilitate rapid T-cell activation through direct drug-HLA interactions
rs549182	NOTCH4, C6orf10	Intergenic variant	May regulate transcriptional programs involved in mast cell activation.
rs549270	NOTCH4, C6orf10	Intergenic variant
rs58302219	HLA-DRB5, HLA-DRA	Intron variant; non-coding transcript variant	May contribute to ICM anaphylaxis by presenting drug-derived peptides to CD4+ T cells, promoting IgE production or memory T-cell activation.
rs9269910	HLA-DRB1	Intron variant
rs9271413	HLA-DRB1, HLA-DQA1	Intron variant; non-coding transcript variant
rs6915986	HLA-DQB1, HLA-DQA2	Intergenic variant
rs3104398	HLA-DQB1, HLA-DQA2	Intergenic variant
rs3104401	HLA-DQB1, HLA-DQA2	Intergenic variant
rs79517313	HLA-DQA2	Missense variant
rs9784876	HLA-DOB, TAP2	Intron variant; non-coding transcript variant

SNP, single nucleotide polymorphism; ICM, iodinated contrast media; UTR, untranslated region; HLA, human leukocyte antigen; IgE, immunoglobulin E.

Integrating genetic and clinical risk factors has shown promise in other allergic diseases. For example, a Korean study demonstrated improved prediction of pediatric asthma and atopic dermatitis when combining GWAS-based SNPs with family history.⁴⁴ Similarly, we observed a higher cumulative burden of risk alleles among patients with ICM anaphylaxis, supporting a polygenic model. Three prediction models were evaluated: genetic, clinical, and combined genetic-clinical. The genetic model alone (AUC = 0.800) did not outperform the clinical-only model (AUC = 0.835), indicating that genetic information is not sufficient as an independent predictor. However, the genetic model reflects biologically relevant variation that is not fully captured by clinical factors, suggesting that genetic susceptibility may still contribute complementary information for identifying high-risk individuals.

The combined genetic-clinical model demonstrated the best performance (AUC = 0.888, accuracy 96%), with improved sensitivity while maintaining perfect specificity. Calibration was also reasonable (slope ≈ 1.24), showing improvement over the genetic model and approaching that of the clinical model. Decision-curve analysis further indicated that the combined model provides modest but meaningful net benefit across clinically relevant thresholds. Although the absolute gain was small (≈ 0.02), identifying one or two additional high-risk patients per 100 may have clinical relevance given the rarity and potential severity of ICM-induced anaphylaxis.

At this stage, however, the clinical utility of these findings remains limited. As discussed earlier, translation into practice would not require full genome-wide testing but rather a targeted, low-cost genetic panel developed after robust replication of the identified variants. Such tools could eventually complement clinical predictors and HLA typing, particularly in cases where clinical information is limited.

Nevertheless, the current models were derived from a modest sample size and internally validated only, and the broader significance of the findings should be viewed as preliminary. Larger, multi-center, and multi-ethnic cohorts, along with functional studies, will be essential to determine which variants have true mechanistic importance and whether genetic information provides meaningful improvement beyond existing clinical models.

This study has limitations. First, the relatively small sample size limits statistical power, particularly for rare variants, and may have led to overestimation of effect sizes for some loci. Large ORs (often > 5) should therefore be interpreted with caution until validated in larger, independent, and multi-ethnic cohorts. Minor clinical differences between the discovery and replication cohorts, such as the distribution of symptoms, may also have influenced effect estimates. However, the main genetic associations were consistently reproduced in both cohorts, indicating that the findings were robust despite clinical variability. Future large-scale studies with harmonized clinical definitions and multi-center replication will be essential to confirm these associations and to generalize the genetic architecture of ICM-induced anaphylaxis across diverse populations. Second, controls lacked confirmed ICM exposure, which could introduce exposure misclassification. However, national data indicate that about 15 million CT scans are performed annually in Korea among 51.7 million people, corresponding to an ICM exposure prevalence of roughly 30% (2023 Korean Society of Radiology; OECD Health Data). Given this high exposure rate and the very low incidence of immediate ICM anaphylaxis (~0.01%), the number of undetected cases among controls is expected to be negligible. Consistent with this assumption, our simulation analysis (Supplementary Table S4) using f = 0.00003 showed ≤ 0.02% change in observed ORs across all loci, confirming minimal bias. Therefore, while the use of population-based controls may slightly limit how precisely we could interpret the risk among exposed individuals, it is unlikely to affect the validity of the genetic associations. Third, functional inference was based on in silico predictions without experimental validation. Fourth, our ethnically homogeneous Korean sample limits generalizability. Finally, although our findings highlight biologically plausible pathways and suggest potential future applications, the clinical implications of these GWAS results should be interpreted with particular caution. The results remain exploratory and are not ready for clinical translation. Meaningful application will require robust replication in larger, multi-center, and multi-ethnic cohorts, as well as functional studies to determine the mechanistic relevance of the implicated variants. In addition, the prediction model was internally validated using limited clinical features; external validation will be essential to determine its generalizability and clinical utility.

In conclusion, this study provides preliminary evidence suggesting genetic susceptibility to ICM-induced anaphylaxis, involving immune, metabolic, neuroimmune, and cytoskeletal pathways. These findings should be interpreted as exploratory and hypothesis-generating due to the limited sample size, the use of imputed HLA data, and the absence of external validation. The combined genetic-clinical prediction model demonstrated good performance and calibration, indicating the potential for the future development of precision prevention strategies once validated in larger, independent cohorts.

SUPPLEMENTARY MATERIALS

Supplementary Table S1

Summary statistics of CookHLA imputation confidence scores across major HLA loci

Supplementary Table S2

LD among risk SNPs on chromosome 6

Supplementary Table S3

Associations of HLA alleles with iodinated contrast media-induced anaphylaxis

Supplementary Table S4

Simulation analysis for potential control misclassification

Supplementary Fig. S1

Quantile–quantile plot of genome-wide association results.

Supplementary Fig. S2

Distribution of CookHLA imputation confidence scores for major HLA loci. The distribution of CookHLA confidence scores demonstrated generally high imputation quality across loci, with most alleles showing scores > 0.75. HLA-A, HLA-C, and HLA-DQB1 showed particularly high confidence, whereas HLA-B and DRB1 exhibited broader variability, consistent with known polymorphism at these loci.

Supplementary Fig. S3

Regional plot for the significant loci in the genome-wide association study test set: (A) around the HLA-DRB1 locus on chromosome 6, (B) around the HLA-DQB1/HLA-DQA2/HLA-DOB locus on chromosome 6. Purple diamonds indicate associated SNPs according to joint analyses. Nearby SNPs are color-coded based on their level of disequilibrium with the top SNP. The left y-axis shows the significance of the association based on the −log10 P value (logistic regression), while the right y-axis shows the recombination rate across the region. Estimated recombination rates from the 1000 Genomes Project Asian database and the hg19 database are plotted with the blue line to reflect the local linkage disequilibrium structure.

Supplementary Fig. S4

Calibration plot of the prediction model for iodinated contrast media-induced anaphylaxis. (A-F) Calibration plots for the three prediction models: genetic (A, B), clinical (C, D), and combined genetic–clinical model (E, F). (A, C, E) Display bin-based calibration using grouped predicted probabilities (dots, with 95% confidence intervals). (B, D, F) Show LOESS-smoothed calibration curves, plotted against the ideal 45° reference line. The combined model demonstrated reasonable calibration performance, showing improvement over the genetic model, although not exceeding the calibration of the clinical model.

Supplementary Fig. S5

Decision-curve analysis for the prediction model of iodinated contrast media-induced anaphylaxis. Decision-curve analysis comparing the genetic, clinical, and combined genetic–clinical models. The combined model showed higher net benefit than the genetic model across clinically relevant threshold ranges and comparable performance to the clinical model, supporting its potential clinical applicability.