Machine Learning Identifies Two Gene Markers of Allergic Rhinitis

The researchers then tested the 13 hub genes using three machine-learning methods: least absolute shrinkage and selection operator regression (LASSO), random forest, and support vector machine recursive feature elimination (SVM-RFE). LASSO selected five genes at the penalty value chosen through cross-validation. Random forest ranked six genes above a MeanDecreaseGini score of 1. SVM-RFE produced its lowest classification error when using three genes. Only two genes, TNFSF14 and CD83, were selected by all three methods. In the original dataset, TNFSF14 alone had an AUC of 0.931, while CD83 alone had an AUC of 0.845. The model combining both genes reached an AUC of 0.968. The researchers also used the two markers to create a nomogram, a chart that estimates the probability of a sample belonging to a particular group. They checked how closely the predicted results matched the observed results and used decision-curve analysis to assess the model across different probability thresholds.

Testing the two-gene model in separate datasets produced different results, but these were still encouraging. In GSE46171, which contained six allergic-rhinitis samples and three control samples, the model achieved an AUC of 1.000. The authors stress that this dataset was so small that the result must be treated with caution. In GSE40037, which contained 10 cases and 12 controls, the AUC was 0.758. This means the model could still distinguish between the groups, but less accurately than in the original dataset. When the genes were tested separately, CD83 showed moderate ability to distinguish the groups in both validation datasets, while TNFSF14 performed differently across them. Both genes generally had lower activity in allergic-rhinitis samples, but the differences between cases and controls were not the same in every dataset. TNFSF14 was significantly lower in GSE46171 but not in GSE40037. CD83 was significantly lower in GSE40037 but not in GSE46171. In this analysis, the two-gene combination therefore appeared to give more consistent results than either gene on its own.

Both genes are involved in immune processes that matter in allergic inflammation. TNFSF14 provides instructions for making a protein in the tumor necrosis factor (TNF) family. This protein binds to receptors called herpesvirus entry mediator and lymphotoxin-beta receptor. These interactions can help lymphocytes multiply and can change immune responses. The lower TNFSF14 activity seen in allergic rhinitis is not easy to explain because the gene has been linked to immune activation in other situations. The authors suggest that changes in the types of immune cells present, the stage of the disease, or the body's attempts to compensate could explain the result. CD83 is linked to the development and control of dendritic cells, which help the immune system recognize threats. Both the form of CD83 attached to cell membranes and its soluble form can affect T-cell and B-cell responses. Lower CD83 activity across a tissue sample could also mean that fewer dendritic cells are present, or that the mix of cells has changed, rather than that individual cells are producing less of the protein. The researchers therefore view TNFSF14 and CD83 as gene-expression signals linked to immune activity. Their results do not directly prove that particular molecular pathways are impaired or that either gene causes allergic rhinitis.

The researchers also estimated the types and activity of immune cells in the samples. These analyses suggested that the gene-expression differences occurred in a T-cell environment that had changed. Single-sample gene set enrichment analysis found lower signals for activated dendritic cells and T follicular helper cells in allergic-rhinitis samples. At the same time, the general T-helper-cell signal increased moderately, while signals for T-cell co-stimulation decreased. CIBERSORT provided a related picture: the cases had lower estimated levels of resting memory CD4 T cells and higher estimated levels of naive CD4 T cells. Pearson correlation analysis showed that TNFSF14 and CD83 had some different, but partly overlapping, relationships with estimates of several immune-cell types. These computer-based findings fit with the idea that the cells and immune activity in the nose differ in allergic rhinitis. However, they do not show which immune cells cause the changes in gene activity. Because the study measured gene activity across whole tissue samples, changes in the number or mix of cells can look like changes in gene control within individual cell types.

Clustering the gene-expression results further divided the allergic-rhinitis samples in the training dataset into two groups for exploration. Non-negative matrix factorization suggested that the samples formed two groups, and principal component analysis showed that the groups occupied different regions in the gene-expression data. TNFSF14 activity was higher in Subtype 2, while CD83 activity was higher in Subtype 1. The groups also differed in estimated levels of plasmacytoid dendritic cells, helper T cells, and signals related to tumor-infiltrating lymphocytes. They showed differences in signals for antigen-presenting-cell co-stimulation, immune checkpoint activity, cell-killing activity, and T-cell co-stimulation. MCP-counter also suggested broader differences in the mix of immune cells. Genes that differed between the groups were linked to depyrimidination, ketone breakdown, and the breakdown of small molecules. These findings point to possible differences in metabolism as well as immunity. Because the groups were identified using only samples from the original dataset, the authors describe them as possible molecular patterns for further study, not established clinical subtypes.

The study also explored possible controls of the two genes, but these results are only starting points for future research. Predictions from NetworkAnalyst linked both genes to possible transcription factors, including members of the NF-kappa-B, STAT, and IRF families. Transcription factors help control which genes are active. The researchers also combined experimentally supported microRNA interactions from miRTarBase with relationships between long non-coding RNAs and microRNAs to build a proposed competing endogenous RNA network. Such a network describes how different RNA molecules might influence one another's activity. TNFSF14 occupied a more central position in the network and was linked to more predicted regulators made from non-coding RNA than CD83. An enrichment analysis using DSigDB also found similarities between the gene-expression patterns and patterns associated with compounds such as puromycin, menadione, sanguinarine, and betamethasone valerate. However, a similarity in a database does not prove that any of these compounds treats allergic rhinitis or changes these genes in patients. The study also did not include protein measurements or experiments to test gene function. It lacked detailed information about disease severity and treatment, did not control for other conditions such as asthma or conjunctivitis, and did not include broad representation from different regions and populations.

Overall, the study identifies a two-gene signature that could help researchers study allergic rhinitis and understand why the disease differs between people. TNFSF14 and CD83 should now be tested in independent samples from nose tissue or blood. Researchers also need to check whether the levels of RNA from these genes match the amounts of their proteins, and whether changing either gene affects dendritic cells or T cells. Larger studies that follow patients over time and record their clinical details will be needed to find out whether the two-gene model remains accurate in people receiving different treatments, people with different levels of symptoms, and people who also have other allergic diseases. For now, the main finding is that three different machine-learning methods repeatedly selected TNFSF14 and CD83 from a group of genes linked to cell adhesion and immune activity. Their biological roles and usefulness in medical care still need to be tested directly.

Subject of Research: Study of gene activity and immune patterns to find possible diagnostic markers for allergic rhinitis

Article Title: Integrative Network Analysis and Machine Learning Identify Cell Adhesion‐Related Biomarkers and Immune Heterogeneity in Allergic Rhinitis

Article References: He, G., & Zhang, L. (2026). Integrative Network Analysis and Machine Learning Identify Cell Adhesion‐Related Biomarkers and Immune Heterogeneity in Allergic Rhinitis. Immunity, Inflammation and Disease, 14(10), Article e70562. https://doi.org/10.1002/iid3.70562

Image Credits: AI Generated

DOI: 10.1002/iid3.70562

Keywords: allergic rhinitis, TNFSF14, CD83, diagnostic biomarkers, machine learning, transcriptomics, immune heterogeneity, cell adhesion, T-cell profiling, public gene-expression datasets, integrative analysis, network analysis