Machine Learning Maps How Environment and Immunity Shape Childhood Eczema Risk

Atopic dermatitis, the itchy, inflamed skin condition better known as eczema, is one of the most common chronic diseases of early childhood, and its origins have long frustrated researchers. A new machine learning study of South African children now suggests that the story of who develops the disease and who is protected is written not in a single gene or allergen, but in the interplay between where a child lives, the antibodies circulating in their blood, the cytokines their immune system produces, and the activity of hundreds of genes in their cells. The research, published in PLOS Medicine, analysed data from 217 AmaXhosa children aged between 12 and 36 months and identified distinct multimodal signatures, one protective and two associated with susceptibility, that could reshape how scientists think about the disease.

The team, led by Damir Zhakparov of the Swiss Institute of Allergy and Asthma Research together with collaborators including Kathleen Moriarty, Michael Levin, Carol Hlela, and Katja Baerenfaller, chose their study population with deliberate care. The AmaXhosa people of South Africa share a common ethnogenetic background, yet some children grow up in rural communities while others live in urban environments. Previous research has documented differences in atopic dermatitis prevalence and allergy sensitisation patterns between African populations, and between urban and rural settings more broadly. That combination, a shared genetic heritage paired with divergent environmental exposures, makes the AmaXhosa an unusually powerful natural experiment for disentangling the environmental and immune mechanisms that underlie the disease.

Methodologically, the study is notable for how it wrangled a genuinely complex dataset. The researchers re-analysed a previously established multimodal collection of measurements spanning four layers: environmental features describing each child’s surroundings, plasma cytokine levels reflecting immune signalling, antibody profiles including allergen-specific and total immunoglobulin E, and transcriptomic data capturing gene expression. Rather than treating these layers separately, they applied machine learning to each modality individually and then integrated them. Their toolkit included GeneSelectR, a workflow for identifying informative genes, SHAP values, an explainability technique that reveals which features drive a model’s predictions, and DIABLO, an integration method designed to find correlated patterns across different data types.

The environmental and antibody analyses produced an early clue. When the models predicted which children had atopic dermatitis, they relied on the combined effects of environmental features together with higher levels of allergen-specific and total IgE antibodies. IgE is the antibody class most closely associated with allergic responses, and its elevation in children with eczema fits the long-standing clinical observation that the skin disease frequently precedes or accompanies food allergies, asthma, and allergic rhinitis. But the SHAP-based explainability analysis showed that the antibody signal did not stand alone; it gained predictive power in combination with the environmental context in which each child was growing up.

The transcriptomic layer proved the richest single source of information. From the gene expression data, the researchers identified a subset of 560 genes whose activity discriminated between children with and without atopic dermatitis. These genes then served as the input for the downstream integration analyses. The size of this discriminating set underscores how genetically and molecularly diffuse the disease is: rather than a handful of culprit genes, the eczema phenotype appears to emerge from a broad shift in cellular programs, a pattern consistent with the involvement of skin barrier function, immune activation, and inflammatory signalling pathways.

The centrepiece of the study came when the team integrated all the modalities at once. Using DIABLO, they identified three multimodal clusters associated with disease status. The first cluster was linked to the healthy phenotype and was composed of environmental features found primarily in the rural setting. Crucially, these rural environmental features correlated with plasma cytokine levels and with the expression of autophagy-related genes. Autophagy is the cellular housekeeping process by which cells degrade and recycle damaged components, and it has been implicated in skin barrier maintenance and immune regulation. The suggestion that rural exposures might be associated with protective cytokine profiles and enhanced autophagy gene expression offers a molecular echo of the hygiene hypothesis, the idea that certain early-life exposures train the immune system away from allergic disease.

The two remaining clusters told a different story. Both were associated with atopic dermatitis, but in distinct ways. One was characterised by correlations between allergen-specific and total IgE antibodies and two cytokines, MCP-4 and TARC. Both of these signalling molecules are chemokines involved in recruiting immune cells to inflamed tissue, and TARC in particular is a well-established marker of the type 2 immune responses that dominate allergic disease. The second susceptibility cluster was defined by a transcriptomic feature signature associated with the atopic dermatitis endotype, the term researchers use for biologically distinct subtypes of a disease that may look similar on the surface. In other words, the study did not find a single molecular route to eczema but at least two, one immunological and antibody-driven, the other transcriptional.

The authors are candid about the limitations of their work. The explainable machine learning framework they employed is exploratory in nature, meaning the clusters and feature associations it surfaces are hypotheses to be tested rather than definitive causal pathways. The study also lacked validation in an independent cohort, a standard benchmark for observational research of this kind, so the identified signatures must be replicated before they can inform clinical practice. Observational designs, however carefully analysed, cannot by themselves prove that rural environments cause protection or that elevated IgE causes disease; they establish associations whose direction and mechanism require further study.

Even with those caveats, the findings carry significant weight for the field. Atopic dermatitis affects a substantial share of children worldwide and is often the first step in the so-called atopic march, the progression from eczema to food allergy and asthma. Understanding which children are on which molecular trajectory could eventually support earlier and more targeted interventions. The identification of a protective cluster rooted in rural environmental exposures, cytokine profiles, and autophagy gene expression points to concrete biological hypotheses about how environment gets under the skin, potentially through immune training and cellular maintenance pathways that could one day be mimicked therapeutically.

There is also a methodological legacy. The researchers emphasise that their feature-selection and integration workflow, combining GeneSelectR, SHAP explainability, and DIABLO integration, provides a reusable framework for analysing complex multimodal datasets in biomedical research more broadly. As studies increasingly collect layered data spanning genomes, transcriptomes, immune measurements, and environmental exposures, the bottleneck is often not data generation but interpretation. By demonstrating how machine learning can be made explainable enough to yield biologically interpretable clusters from such data, the team offers a template that extends well beyond eczema, and their AmaXhosa cohort stands as a reminder that the most informative answers may come from populations whose environments, as much as their genes, tell the story.

Subject of Research: Machine learning identification of protective and susceptibility clusters in paediatric atopic dermatitis among AmaXhosa children

Article Title: Multimodal protective and susceptibility clusters in paediatric atopic dermatitis: A machine learning-based, data-driven observational study

Article References: Zhakparov, D., Lunjani, N., Schmid, M., Moriarty, K., Roquero, D., Dreher, A., Heldstab-Kast, J. I., Nadeau, K. C., Akdis, C., Levin, M., Hlela, C., Sokolowska, M., O’Mahony, L., & Baerenfaller, K. (2026). Multimodal protective and susceptibility clusters in paediatric atopic dermatitis: A machine learning-based, data-driven observational study. PLOS Medicine, 23(9), e1004917. https://doi.org/10.1371/journal.pmed.1004917

Image Credits: AI Generated

DOI: 10.1371/journal.pmed.1004917

Keywords: atopic dermatitis, eczema, machine learning, AmaXhosa, pediatrics, IgE antibodies, cytokines, transcriptomics, autophagy, rural environment, immune regulation, PLOS Medicine