Childhood asthma is a heterogeneous chronic airway disease with various clinical phenotypes [1, 2]. Its phenotypic and biologic heterogeneity contributes to the challenges clinicians face in its diagnosis and effective management [3]. It is therefore crucial to clearly define subphenotypes of asthma with homogeneous clinical characteristics in order to search for better asthma management and to develop novel therapeutic strategies. Although a large number of clinical phenotypes are often collected in childhood asthma studies, asthma genetic study has been mostly focused on case-control disease status. Such an endpoint-based analysis ignores the complexity of asthma phenotype [4–6]. In addition, although there is ample evidence for an intrinsic genetic variation among racial/ethnic populations [7, 8] suggesting possible genetic ancestry variation in childhood asthma, most genetic analyses rely on self-reported race thus do not account for the potential contribution of genetic ancestry to disease variation in diverse populations.

An approach to overcome the phenotypic heterogeneity of childhood asthma is to identify homogeneous subgroups by establishing either classical “endotype”, based on experts’ criteria, or statistical phenotype clustering on asthma clinical phenotypes. The latter has been successfully applied to identify clinically relevant subgroups of asthmatics and other airway diseases [9–17]. However, these studies differ in some key elements: variation in phenotyping, analytical approaches used and the patient population under study. These differences limit the comparability of the identified subphenotypes and pose difficulty in applying clustering results to individual patients. Furthermore, little is understood regarding the genetic ancestry of the identified subphenotypes.

The objective of the study is to investigate childhood asthma phenotypic heterogeneity and genetic ancestry variations and their relationships. Specifically, we used childhood asthma data from the NIH controlled database of Genotype and Phenotype (dbGaP) to identify homogeneous subphenotypes, determine clinical phenotypes, estimate subphenotype-specific genetic ancestry, and analyze the relationship between ancestry and subphenotypes using a stepwise approach incorporating cluster analysis, classification tree analysis, and genetic ancestry analyses [9–16, 18, 19]. Our goal is to combine both cluster and genetic ancestry to identify biologically-relevant subphenotypes in childhood asthma.

Participants from CAMP and CARE were different except in sex, exposure to in utero smoking, PC20, and FEV1/FVC ratio (Table 1). All pairwise Spearman correlation coefficients were less than 0.60, except between FEV1 percent predicted and FVC percent predicted (0.71) and between FEV1/FVC and maximum bronchodilator percent change (− 0.65). No variables had more than 10% of missing values.

Atopic dermatitis status and SPT distinguished the subphenotypes

Conditional inference trees analysis revealed that, in both CAMP and CARE data, AD and one or more positive SPT were the top two variables that best discriminated the individuals into the subphenotypes (Fig. 1, prediction accuracy 95.8%). Given the consistent findings across CAMP and CARE data, we combined the two datasets and grouped the three clusters individually identified in CAMP and CARE into three subphenotypes. One subphenotype was the moderate AD group with negative SPT and preserved lung function (subphenotype 1, n = 158), one was the high AD group with positive SPT and airway hyperresponsiveness (subphenotype 2, n = 300), and one was the low AD group with positive SPT and lower lung function (subphenotype 3, n = 753).

Genetic ancestry proportion varied at asthma GWAS SNPs among asthma subphenotypes

The three subphenotypes had 15, 49, and 103 African American individuals, respectively. Global African ancestry proportion varies from 71.2 to 100% with mean 96.6% and standard deviation (SD) 7.2%. Higher global African ancestry was associated with AD (mean ± SD of African origin is 0.96 ± 0.08 for AD free vs. 0.98 ± 0.06 for AD subjects, p-value = 0.0294), but not with other clinical phenotypes. Proportion of African ancestry at asthma GWAS SNPs was correlated with the subphenotypes (mean 64.9, 89.4 and 94.4% for subphenotypes 1, 2, and 3, respectively, p-value < 0.0001, Figs. 2 and 3(a)). The subphenotypes were associated with lung function: FEV1 percent predicted is 96.8 ± 14.1, 95.3 ± 13.9, and 93.9 ± 13.7 (p-value = 0.0083); and FEV1/FVC ratio is 81.9 ± 7.6, 80.5 ± 8.1, and 79.0 ± 8.2 (p-value < 0.0001) for subphenotypes 1, 2, and 3, respectively. Furthermore, African ancestry at asthma GWAS SNPs was inversely associated with AD (median 0.95 with IQR (0.93, 0.95) for AD free vs. 0.92 (0.89, 0.94) for AD subjects, p-value < 0.0001, Fig. 3(b)). Additionally, genetic ancestry at asthma GWAS SNPs was associated with positive SPT with median and interquartile range (IQR) 0.94 (0.92, 0.95) for positive SPT individuals vs. 0.74 with IQR (0.59, 0.78) for negative SPT individuals (p-value < 0.0001, Fig. 3(c)). The odds of one or more positive SPT was 6.3 higher (95% confidence interval: (3.4, 13.8), p-value < 0.0001) with each additional 10% of African origin at asthma GWAS SNPs. African origin at asthma GWAS SNPs was also associated with IgE levels (Spearman correlation coefficient = 0.27, p-value = 0.0004) and IgE was 1.5 fold higher with each additional 10% of African origin (Fig. 3(d)).

Current clinical practice in childhood asthma treatment tends to use average patient care strategies. Such a “one size fits all” treatment approach faces major challenges when it is becoming clearer that childhood asthma is heterogeneous in pathogenesis. Our unbiased cluster and genetic ancestry analyses pointed toward three distinct phenotypic clusters with differences in clinical characteristics, genetic ancestry, and clinical outcomes, underscoring the clinical and genetic heterogeneity of asthma [10, 13, 17, 31]. Previous studies have also identified clusters with atopic or non-atopic asthma, clusters with preserved or lower lung function, and clusters with mild asthma [13, 14, 32]. It is reassuring that the two independent studies replicated the clustering results and there are similarities with previous clustering-based childhood asthma subphenotypes.

We determined genetic ancestry [33] using genome-wide SNPs and asthma GWAS SNPs for African-American asthmatic individuals in CAMP and CARE data. Our estimate of African global ancestry in asthmatic children is higher than what has been reported in different general populations confirming the higher prevalence of asthma in individuals with higher African ancestry than others. Our results showed that genetic ancestry at asthma GWAS SNPs differed between the childhood asthma subphenotypes and was associated with lung function, SPT, IgE levels, and AD. Previous studies have also showed association between genetic ancestry and asthma prevalence and related clinical phenotypes [34–42]. To our best knowledge, our study is the first to show the association between genetic ancestry at asthma GWAS SNPs and cluster-based subphenotypes in childhood asthma. Leveraging ancestry and cluster analyses to derive genetic and phenotypic homogeneity subgroups in childhood asthma demonstrates the utility of these approaches to characterize and understand the complexity of asthma towards individual based precision medicine strategies.

This study demonstrates that genetic ancestry at asthma GWAS SNPs is more strongly associated with asthma subgroups sharing similar clinical characteristics compared to broadly defined asthma. The results suggest that validation of genetic studies are more likely to be successful for replication studies carried-out in more homogeneous asthma cohorts (sharing similar clinical characteristics) compared to the multifactorial case-control status. In addition, the results indicate that ancestry-specific genetic loci of asthma are likely to be found by focusing on better defined asthma patients. Furthermore, genetic ancestry analysis in homogeneous asthma subgroups is suitable to refine the biological role of asthma susceptibility variants from GWAS studies in a given phenotype. For example, SNPs at STARD3/PGAP3 are strongly associated with the high atopic dermatitis subgroup suggesting that STARD3/PGAP3 may act on the allergic component of asthma [43]. Another example is that ORMDL3/17q locus is associated with asthma in multiple studies in the European ancestry but not in African ancestry asthmatic individuals [44]. We also investigated associations between asthma GWAS SNPs with the identified subphenotypes in CAMP and CARE data (methods and results in Additional file 1: Table S1). Several significant associations were identified at p = 0.05, but none after multiplicity adjustment, possibly due to small sample size and limited statistical power.

Our study had several limitations. First, participants in CAMP and CARE represent studies of childhood asthma, thus the results herein may not be applicable to adulthood asthma. Second, although we identified clinically relevant subphenotypes of childhood asthma using clinical phenotypes [45], the integration of this result with molecular and physiologic phenotyping may help to better understand childhood asthma pathogenesis for possibly more personalized therapeutic strategies. Furthermore, subgroup analyses of asthma may limit sample sizes and impair statistical power. However, given asthma is a highly heterogeneous phenotype, studying homogeneous subgroups of asthma patients not only recovers power limitation, but achieves more statistically significant results. Classifying asthma patients in more homogenous groups may help us to identify new susceptibility or modifying subphenotype-specific genes. Our ability to better define subtypes might help to predict who may respond to treatment vs subjects who may not. Future studies need to elucidate the mechanisms that distinguish each ancestral and clinical clusters to facilitate the development of targeted therapies and providing personalized treatments.

The present study has notable strengths. First, we were able to dissect childhood asthma heterogeneity into subphenotypes using cluster analysis of clinical phenotypes in one study and replicate the findings in an independent study. Second, we were able to show associations between the identified subphenotypes with asthma clinical outcomes. Third, analysis of genetic ancestry at asthma GWAS SNPs in childhood asthma clinical phenotypes provide biologically relevant subphenotype-specific results. Lastly, our study used a more accurate and direct assessment of genetic ancestry instead of self-reported race to determine the relationship between ancestry and childhood asthma subphenotypes and relevant clinical phenotypes. Studies have shown that people with the same self-reported race could have drastically different levels of genetic ancestry, and self-reported race may not be as accurate as direct assessment of genetic ancestry in predicting treatment outcomes [33]. Future studies to identify genetic ancestry-specific variants associated with a specific subphenotype are important as we move towards applying precision medicine paradigm. The finding indicates that African genetic ancestry at asthma GWAS SNPs are differentially associated with the asthma clinical subphenotypes. Unraveling the reasons why individuals with higher African origin at asthma GWAS SNPs had higher IgE level or rate of positive SPT is necessary to determine the potential clinical applications of our findings. In addition, genetic analysis based on more refined phenotypes may increase the statistical power and allow for the detection of population structure-specific phenotype-genotype associations that are undetectable otherwise.

In conclusion, through our systematic clinical phenotype analysis, we identified distinct subphenotypes for childhood asthma using cluster analysis. Further genetic ancestry analysis showed correlations between African ancestry at asthma GWAS SNPs and childhood asthma subphenotypes and related clinical outcomes. Our results demonstrated that cluster analyses on clinical phenotypes followed by ancestry analysis can enhance the understanding of the phenotypic and genetic heterogeneity of childhood asthma. Our approach is distinct from previous efforts in that we developed cluster-based subphenotype and applied genetic ancestry analysis to define subphenotype-ancestry relationships which can be subsequently used as the basis of genetic ancestry based clinical risk prediction. Our findings suggest that defining asthma heterogeneous subgroups on the basis of clinical phenotypes and genetic ancestry proportion is an essential step to understand and refine patient subsets and enable more targeted therapy.