Ethics statement

The study received the University of Cape Town Faculty of Health Sciences Human Research Ethics Committee approval (HREC 451/2014) and was conducted in accordance with the Declaration of Helsinki [14]. Written informed consent was obtained from parents or guardians of all participants prior to inclusion in the study. Additional information regarding the ethical, cultural, and scientific considerations specific to inclusivity in global research is included in the Supporting Information (S1 Checklist: Inclusivity in Global Research).

Study description and questionnaires

As detailed in [8,15] this study employed a single-visit, cross-sectional case-control design involving AmaXhosa children in South Africa who were between 12–36 months old, lived in either a rural or urban environment, and were either healthy or had a dermatologist’s diagnosis of moderate to severe AD, fulfilling the UK Working Party criteria for AD. Exclusion criteria included other considerable health conditions, immune-mediated diseases other than allergic comorbidities and positive stool and serum IgE tests for helminths. Urban children with moderate to severe AD were enrolled at the Paediatric Dermatology Clinic, Red Cross Children’s Hospital, Cape Town. Healthy nonallergic, nonfood sensitised (on skin prick test (SPT) screening) participants were enrolled from randomly selected crèches in the Cape Town metropole. Similarly, children from a rural setting with AD were enrolled at the dermatology clinic at the Nelson Mandela Academic Hospital, Umthatha. Their nonallergic healthy control counterparts were randomly selected at study sites based at village clinics in the Mqanduli district of Umthatha. All participants were seen at one time point for questionnaire, SPT, clinical assessment, and blood sample collection. In total, standard questionnaire data on environmental exposures and living conditions of the children were obtained for 217 AmaXhosa children (rural_AD: 60, rural_healthy control (HC): 52, urban_AD: 56, urban_HC: 49). For 152 children, we additionally had antibody data, for 149 children we had bulk RNA-Seq data, and for 159 children we had plasma cytokine data (Table A, Fig A in S1 Appendix).

Allergen sensitisation, allergen antibody and cytokine measurements

To establish the antibody data that form part of the clinical dataset, SPTs were performed as previously reported in [8], using Alk-Abello (Alk-Abello, Denmark) reagents for sensitisation to cow’s milk, egg white, peanut, hazelnut, wheat, soy, fish, raw egg white, fresh cow’s milk, and fresh peanut. ImmunoCap assays (Thermo Fisher Scientific, Sweden) were used for quantifying the antibody levels of total IgE, specific IgE against D. pteronyssinus, milk, fish, wheat, peanut, soy, hazelnut, egg, and CCD, and specific IgG4 against D. pteronyssinus, fish, wheat, peanut, soy, hazelnut, casein, and egg.

The levels of plasma analytes and cytokines were assessed using the Meso Scale Discovery multi-spot assay system and QuickPlex SQ120 platform (Meso Scale Discovery, USA). The cytokine data were subjected to statistical analysis testing for differences between HC and AD groups using a Mann–Whitney test, followed by correction for multiple testing with Bonferroni (Table B in S1 Appendix).

ML on questionnaire and antibody data

As reported previously, structured questionnaires were used to collect data on family history of atopy and allergy, medical history, household demographics, early-life exposures and various environmental factors [8] (S1 File). The questionnaire features were pre-processed to exclude features mainly consisting of missing values or limited information, as well as those that were highly imbalanced between the two groups. Features directly associated with AD, such as the SCoring Atopic Dermatitis (SCORAD) score or medication with antihistamines, were marked for exclusion from the analyses, while features with more than two possible entries, such as the kind of fuel used for heating or cooking, were one-hot encoded as separate features. The combined pre-processed questionnaire and antibody data were subjected to statistical analysis to identify differences between HC and AD groups. A hypergeometric test was used for categorical variables, and a Mann–Whitney test was used for numerical features. The results were corrected for multiple testing using the Bonferroni method (Table C in S1 Appendix).

In the ML classification, AD versus HC was set as the target variable. All data were included to build the models using cross-validation (CV) as the main aim was not to build a single best model, but to extract the features that contributed most to predictive performance. To conduct the hyperparameter search for the models, the Bayesian optimisation-based Optuna framework was used, which employs the Tree-structured Parzen Estimator (TPE) to navigate the hyperparameter search space in an informed manner. Unlike classical random or grid searches, Optuna samples combinations of hyperparameters, including algorithm choice, that yield better performance, while deprioritising less optimal ones. This approach enabled the selection of hyperparameter settings that were most likely to yield a high classification score [16]. As the stopping criterion, limits of n_trials = 150 and a timeout of 36,000 s were defined, and the timeout was reached first after 78 trials were executed. The hyperparameter search space also included selection of the ML algorithms RandomForestClassifier [17,18], Multi Level Perceptron (MLP) [18,19], Gaussian Naive Bayes (GNB) [18,20] or eXtreme Gradient Boosting (XGBOOST) [21]. In running the ML pipeline, missing data were initially imputed using the miceforest package [22], which employs RandomForest to estimate the missing values. This approach was selected because of the uneven distribution of missing values between AD and HC, which could otherwise influence the classification. The features were subsequently scaled, and quasi-constant features were removed using a variance threshold of 0.15. To avoid overfitting, the feature selection procedure was constrained to select a minimum of 5 and a maximum of 30 features. The pipeline was then evaluated for every hyperparameter combination in a 5-fold CV process, using mean accuracy as the classification performance measure, as the distribution between AD and HC in the data is nearly equal. As described in detail in S2 File, all data preprocessing steps were applied separately within each CV fold. After the thorough model building process, 78 models were trained and the best model performance was reached in Model 52 (M52). This best-performing Random Forest model had a CV mean accuracy score of 0.81. Other classification metrics were similarly high (approximately 0.80) for both CV and the test set (Fig B in S1 Appendix). After obtaining the CV accuracy scores, all models were refitted on the full dataset with their respective parameter settings to generate the final SHAP [11] values to interpret the outputs of the ML models. Inspecting the importance of the different hyperparameter search space features revealed that the choice of model had the greatest impact on the objective function, while scaling and imputation contributed little to the classification (Fig C in S1 Appendix).

Feature selection in RNA-Seq data

As described previously in [15], total mRNA was isolated from PBMCs and sequenced using HiSeq 40,000. The FASTQ sequencing data were preprocessed using the ARMOR pipeline with integrated differential gene expression (DGE) analysis using edgeR [23]. From the original 150 RNA-Seq samples one outlier was identified in the Principal Component Analysis and excluded from analyses when rerunning the pipeline. The final RNA-Seq data were from PBMCs of 149 AmaXhosa children living in urban or rural environments and being healthy (HC) or suffering from AD (groups: urban_HC = 29; urban_AD = 31; rural_HC = 44; rural_AD = 45). These four groups were set as comparisons in the design matrix of edgeR. Transcripts were considered significantly changed between two conditions if they had a false discovery rate (FDR) < 0.05 and a fold change (FC) > 1.5.

To select relevant genes for subsequent data integration, we utilised the GeneSelectR R workflow that we recently developed [12], designed to overcome shortcomings associated with classical DGE analysis in complex datasets. Initially, the input transcript count data were normalised using a between-sample normalisation approach from the edgeR R package [23]. Subsequently, the workflow was used to select informative genes (features) based on ML metrics and their biological relevance across several steps:

1. (1) A standard ML pipeline for feature selection incorporating four ML methods (Logistic Regression with L1 Penalty (Lasso), Random Forest, Boruta and Univariate filtering (Univariate)), with hyperparameter adjustment was applied.
2. (2) Features were ranked based on their inbuilt method-specific feature importance scores. Additionally, feature ranking across different iterations was reported.
3. (3) Genes meeting the filtering criterion to appear in at least 2 out of 10 iterations were assembled into a gene list for every method.
4. (4) A GO enrichment analysis was conducted on each gene list. GO categories with an adjusted p-value < 0.05 were considered enriched and subsequently subjected to semantic similarity analysis [24].

The comprehensive breakdown of the analysis is available in S3 File.

Weighted Gene Co-expression Network (WGCNA)

The weighted gene co-expression network analysis (WGCNA) was performed using the official R package, as described in [25]. In summary, a correlation matrix between genes in the dataset was calculated to detect co-expression patterns among the genes, and genes with similar patterns were clustered into modules. These gene modules were then correlated with the ML-selected questionnaire features using the cor() function.

ML on integrated data

To combine different data modalities, we adopted a late data integration strategy [26,27]. In this approach, feature selection is performed on each sub-dataset separately to reduce dimensionality and avoid introducing redundant noise into the combined dataset. The sub-datasets were then matched using a complete-case approach, resulting in 132 samples with patient information, antibody data, RNA-Seq data, and quantitative cytokine data (Fig A in S1 Appendix).

To integrate the RF560 RNA-Seq, the combined antibody and questionnaire, and the cytokine datasets, the DIABLO method from the mixOmics package was used [13]. In brief, DIABLO is a supervised ML method aimed at decomposing the dataset into components projected into latent space and correlated with the label of interest. Using DIABLO, we observed that the datasets had high pairwise correlation scores, suggesting that there is a common pattern across the data modalities (Fig D in S1 Appendix). This is also reflected in the latent space projection sample plot, where samples are distinctly grouped based on the labels AD or HC (Fig E in S1 Appendix).

ML applied to questionnaire and antibody data reveals key variables associated with promotion of AD

The study cohort consisted of 217 AmaXhosa children for whom questionnaire data on environmental exposures and living conditions were available, divided into four groups according to whether they lived in rural or urban areas and whether they were healthy or had a dermatologist-confirmed diagnosis of moderate-to-severe AD (Table A in S1 Appendix). Statistical analysis of the pre-processed questionnaire data combined with the antibody data revealed that, among the environmental variables, the use of electricity or gas as fuel for cooking differed significantly between HC and AD (Table C in S1 Appendix). To go beyond these univariate tests, we employed various ML algorithms with HC versus AD as the target variable in a classification process to identify distinguishing features between these groups. By calculating Jaccard scores to assess the similarities between the feature lists selected by the top three models of each ML algorithm, we found that Random Forest achieved the highest feature stability (Fig F in S1 Appendix). Inspecting the features that were selected in these three best-performing Random Forest models revealed an overlap of 18 features. These features include total IgE antibody levels, as well as the levels of specific IgE and IgG4 antibodies against house dust mites (sIgE_Der and sIgG4_Der) and specific IgE antibodies against different food. Additional features include exposure to food such as eggs, the type of fuel used for heating or cooking, and the duration of sunlight exposure (Table D in S1 Appendix).

To better explain the model output, SHAP values were calculated, which measure how much each of these features contributes to the prediction of the best-performing model. The SHAP values show that the levels of sIgE antibodies against house dust mites (sIgE_Der) and against egg (sIgE_egg), as well as using electricity or gas as fuel for cooking are the three most important features with the greatest influence (Fig G in S1 Appendix). While feature importance quantifies the contribution of each feature to the model prediction, it does not specify whether high or low feature values correlate with the prediction of AD. To address this, the influence of individual features on predicting AD is illustrated in Fig 1A. The findings reveal that higher levels of specific IgE against house dust mites, egg, fish, milk, and total IgE are associated with the prediction of AD, as shown in the SHAP dependence plots (Figs 1B, H–K in S1 Appendix). Similarly, higher levels of specific IgG4 antibodies against house dust mites were positively associated with the prediction of AD (Fig L in S1 Appendix), as well as no regular exposure to hen egg (Fig 1A). Of the living environment features, using electricity or gas for cooking positively influences the prediction of AD, whereas not using them contributes towards predicting HC. Conversely, using a paraffin stove or fires outside the house for cooking, or wood or coal for heating contributes towards predicting HC (Fig 1A). Finally, a higher age of first exposure to paracetamol is associated with the prediction of HC. These analyses demonstrate that, although individual features in the questionnaire data may not show significant differences in univariate statistical tests, their combined effect contributes to predicting AD.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. SHAP values of the importance and the impact of the individual features on the output of the best-performing model applied to the combined antibody and questionnaire data.

A) The SHAP values of the 18 Random Forest-selected features with the highest impact in target prediction indicate the degree of change towards predicting AD in log odds. Each dot represents a row of data from the questionnaire or antibody data, with the colour of each point indicating the value of the corresponding feature through a colour gradient from yellow (high value), transitioning through red, to purple (low value). B) SHAP dependence plot for the levels of specific IgE antibodies against house dust mites.

https://doi.org/10.1371/journal.pmed.1004917.g001

Adopting a novel ML-based RNA-Seq analysis pipeline revealed a transcript signature of AD

The dataset-wide normalised RNA-Seq data of 149 AmaXhosa children were first analysed with DGE analysis testing for statistical differences between HC and AD, and between rural and urban. As expected from the results of the principal component analysis (Fig M in S1 Appendix), significantly more transcripts changed between locations than between diagnoses (Table E in S1 Appendix), indicating a strong influence of the environment on the data distribution. For the differences between AD and HC, we found only one gene, Proline Rich Coiled-Coil 2A (PRRC2A), which was present in both the 36 DEGs for the rural samples and the 82 DEGs for the urban samples (Table E in S1 Appendix). In addition, we found no significant correlation in the FCs between AD and HC for these genes in urban and rural samples. (Fig N in S1 Appendix). Combined, these findings indicate the limited capability of DGE analysis to identify differential features within a complex dataset where two variables contribute differently to the data distribution.

Considering the univariate nature and limitations of DGE, we used the GeneSelectR workflow [12] to identify a subset of genes distinguishing between AD and HC using ML, as this approach also considers complex nonlinear relationships between features. We employed the four ML-based feature selection methods implemented in GeneSelectR, which are Random Forest (RF), Lasso, Boruta and Univariate Filtering (Univariate). Feature selection based on the inbuilt feature importance resulted in lists with varying numbers of transcripts (RF: 560, Lasso: 372, Univariate: 198, Boruta: 53). The classification performance according to different metrics was similar, with a CV mean score of around 0.7. However, RF scored slightly higher than the other ML methods (Fig O in S1 Appendix). Calculating three different overlap coefficient metrics implemented in GeneSelectR revealed that there was only minimal overlap and no consensus gene signature shared across the lists (Fig P in S1 Appendix).

As neither the classification performance metrics nor the overlap coefficients allowed for determining which feature list should be used in downstream analyses, the next step was to evaluate their biological relevance through an overrepresentation analysis of Gene Ontology Biological Process (GOBP) terms. In all the lists, a total of 379 enriched GOBP terms were identified. These functional terms were clustered with binary cut based on their semantic similarity using the topology information from GO. Of the 10 identified clusters, three stood out as more specific to the diagnosis of AD because they were related to immune processes: cluster 1 (lymphocyte development and proliferation), cluster 6 (inflammatory response and oxidative stress), and cluster 7 (metabolism and immune cell processes) (Fig 2A). The 560 features selected with the RF method, termed RF560 list, prominently featured terms from clusters 6 and 7, establishing its significance in AD biomarker identification. In conclusion, the RF-derived list emerged as the most suitable for delineating AD characteristics in the cohort for the subsequent analyses.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Gene lists associated with AD identified by semantic similarity clustering of enriched GOBP terms.

Heatmap of the semantic similarity clustering of the 379 GOBP terms passing an unadjusted p-value threshold of 0.01 in a functional overrepresentation analysis of the feature lists selected by different ML methods and in the DEGs. The composition of each of the 10 clusters is described by the word clouds on the right. At the left, the smaller heatmap shows the significance of the enrichment of the different GOBP terms, with a gradient from yellow to purple, where purple indicates a low adjusted p-value (padj), and the histograms demonstrate the number of enriched GOBP terms in each of the clusters.

https://doi.org/10.1371/journal.pmed.1004917.g002

Weighted co-expression gene network analysis demonstrates a high correlation of the RF560 list with significant questionnaire and antibody features

The weighted co-expression network analysis (WGCNA) of the genes in the RF560 list revealed 11 clusters of co-expressed genes (S1 Table). These co-expressed gene clusters were correlated with the 18 previously RF-selected questionnaire and antibody features, as well as the two main experimental variables, location (urban versus rural) and diagnosis (HC versus AD), to identify significantly associated features. This association can indicate that the gene clusters play a role in influencing these features. Most gene clusters were significantly correlated with location, underscoring the strong influence of this variable on gene expression. In contrast, only the two clusters (the magenta and yellow clusters shown in Fig 3), which were associated with rural location, also showed a significant positive correlation with healthy status. Furthermore, one cluster (the blue cluster) was significantly positively correlated with levels of specific IgE antibodies against milk, one cluster (the black cluster) with any reaction to food, and one cluster (the grey cluster) was significantly negatively correlated with the use of wood or coal as fuel for heating. In summary, the correlation of specific questionnaire and antibody features with specific gene clusters again emphasises the complexity of the dataset.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Correlations of co-expressed ML-selected gene clusters with significant questionnaire and antibody features.

Heatmap of the correlation between the 11 co-expressed gene clusters identified using WGCNA applied to the RF560 gene list (y-axis) and the 18 questionnaire and antibody features present in the three best RF models plus the major experimental variables location and diagnosis (x-axis). The direction of the correlation is indicated in brackets. For example, Diagnosis (HC | AD) means that the positive correlation of a gene cluster with HC is > 0, while the correlation with AD is < 0. The value of the correlation is indicated with a colour gradient from green to purple, where green indicates a negative correlation and purple indicates a positive correlation. For correlations that passed a p-value threshold of 0.05, the correlation score is indicated along with the p-value in brackets.

https://doi.org/10.1371/journal.pmed.1004917.g003

DIABLO integration of all datasets identifies key features related with AD

To further explore features of the four datasets significantly associated with AD, we applied the DIABLO data integration framework [13] to investigate relationships among the combined questionnaire and antibody data, the cytokine data, and the RF560-selected transcripts. To determine which samples drive the separation of HC and AD in the latent space projection plot (Fig E in S1 Appendix), the contributions of variables from the different datasets were visualised on the first two components of the DIABLO model. Since the positive or negative sign of the contribution scores is relevant only in the context of the projection plot, the absolute values of the component contribution scores reflect the importance of each feature (Fig 4).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Contributions of variables from the A) combined questionnaire and antibody data, B) the cytokine data, and C) the RF560 selected transcripts to the prediction of AD (blue) or HC (orange) on the first two components of the DIABLO model.

The contribution scores of the components are on the x axis, and the feature names are on the y- axis. The higher the absolute value of the contribution score, the more important a feature is. Features without a component score in components 1 or 2 have a tie between the two groups, indicating equal contribution in AD and HC.

https://doi.org/10.1371/journal.pmed.1004917.g004

In the questionnaire and antibody data, the use of electricity, gas, or paraffin as fuel for cooking or heating, higher counts of monocytes, eosinophils, and lymphocytes, as well as elevated levels of specific and total IgE antibodies contributed to the prediction of AD in the first two components (Fig 4A). The questionnaire features contributing most to the prediction of HC were contact with farm animals of the mother or the child, higher exposure to sunlight in summer and winter, and keeping a dog (Fig 4A). These features are all significantly more prevalent in rural living conditions [8].

In the cytokine data, the major contributors to predicting AD in the first two components were IL-1β, IL-8, IL-6, Flt1, MIP-1β, bFGF, TARC, MCP-4, and IL-16 (Fig 1B). The DIABLO model therefore enabled the identification of a greater number of cytokines associated with the diagnosis of AD compared to the univariate statistical test, which identified only Eotaxin-3 and TARC as significantly increased in AD patients compared to HC individuals (Table B in S1 Appendix). Furthermore, DIABLO allowed the identification of transcripts in the RF560 list that contributed most to the prediction of AD or HC. In both components 1 and 2, the major contributors were transcripts predicting AD, including EMC2, IFT27, RASSF7, MAP1LC3A, MRI1, CLC, LTF, MRPL28, CD99, MYL6, RANBP1, and PRMT5. By contrast, the transcripts contributing most to the prediction of HC included C5, MAMLD1, MYCBP2, NRDC, ZFAND6, UBR2, EVI5, and TEAD3 (Fig 4C).

Finally, we used the DIABLO model to investigate how the variables from the four different datasets interact with each other. As depicted in the correlation circle plot in Fig 5A, three distinct clusters of correlated features can be identified, highlighted by the violet, orange, and green areas. The violet cluster (Fig 5B) represents a HC protective cluster. It features the cytokines VGEF D and Tie 2, which contributed most to the prediction of HC in the first component of the DIABLO model for the cytokine data (Fig 4B). These cytokines were correlated with the clinical questionnaire features exposure to sunlight, contact with farm animals, and keeping a dog, which contributed most to the prediction of HC in the first component of the DIABLO model (Fig 4A) and are reported to be protective [29–32]. A network analysis of the six correlated features from the RNA-Seq dataset using STRING [33] resulted in the identification of a subcluster comprising ZFAND6 and MAP1LC3A, which was associated with the UniProt keywords autophagy, protein transport and Ubl conjugation pathway (Fig Q in S1 Appendix).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Correlations among features across datasets displayed in a circle plot reveal three distinct clusters.

A) Circle plot showing correlations between variables from the RF560-selected transcripts (red squares), the questionnaire (blue triangles), the cytokine (cyan diamonds), and the antibody (black circles) datasets. The angles between two vectors drawn to the variable point from the origin show the correlation score (<90° = positive correlation, >90° = negative correlation). The distance from the origin indicates the importance of a variable for the respective component, with the inner circle indicating a correlation score threshold of 0.5. The areas highlighted in violet, orange, or green indicate clusters of correlated variables. B) Circos plots of the correlated features in the violet, orange, or green clusters, respectively.

https://doi.org/10.1371/journal.pmed.1004917.g005

In contrast, the orange cluster was associated with AD, as it contains the cytokines TARC, MCP4, and IL-16, which are the cytokines contributing to the prediction of AD in the second component of the cytokine DIABLO model (Fig 4B), as well as the levels of total IgE antibodies and of specific IgE antibodies against peanuts, egg and house dust mites (Fig 4A). Submitting the genes of the orange cluster to a network analysis in STRING [33] resulted in three different sub-clusters). One of these sub-clusters primarily comprises nuclear proteins involved in transcription and chromatin organisation, while another sub-cluster, centred around LTF, shows enrichment for the Reactome pathways Antimicrobioal peptides and Innate immune system (Fig R in S1 Appendix).

The green cluster highlighted in the circle plot in Fig 5A was also correlated with AD. It contains the environmental variable to use electricity or gas as fuel for cooking (Fig 5B), which is the variable that contributes most to the prediction of AD in the first component of the DIABLO model of the clinical questionnaire data (Fig 4A). Of the 13 transcript features in the green cluster, the 9 with a FC of at least 1.08 were considered a signature and subjected to a similarity search against the human mRNASeq dataset in Genevestigator [34] to identify comparisons between conditions or treatments that yield similar transcript signatures. Interestingly, the highest similarity was observed in peripheral blood samples from patients with moderate to severe AD, either untreated or after three months of treatment with Dupilumab, a monoclonal antibody that blocks IL-4 and IL-13 and is used to treat AD. In that study, clinical improvement following Dupilumab administration was accompanied by a decrease in innate immune responses and an increase in B cell and natural killer cell activation [35]. The dataset with the second-highest similarity score was from an asthma study that compared peripheral blood samples collected at baseline and between four and six days after the onset of cold symptoms, without progression to exacerbation [36]. These findings underscore the importance of this gene signature in allergic diseases.