Immune Gene Signature Linked to Severe Male Infertility Identified by Machine Learning

Non-obstructive azoospermia, one of the most severe forms of male infertility, leaves the semen completely devoid of sperm despite open transport pathways, because the testes themselves fail to produce mature germ cells. For the millions of men affected worldwide, the diagnosis often arrives with little mechanistic explanation, and treatment options remain limited to surgical sperm retrieval attempts whose success rates are difficult to predict. Now a research team led by Yucheng Zhong, Songqing Ye, Hongshen Lai, Xiangyu Zeng, Ge Song and Xiaofu Qiu has turned to large-scale computational biology to illuminate what has long been a biological black box. In a study published in PLOS One, the researchers report the construction of a twelve-gene immune-related signature that distinguishes men with non-obstructive azoospermia from fertile controls with notable accuracy, and they map how immune cell populations within the testicular environment shift in the disease state.

The study’s central premise rests on an idea gaining traction across reproductive medicine: infertility is not solely a matter of sperm-producing cells failing, but may involve the broader immunological milieu of the testis. The testis is an immunologically privileged organ, a site where the immune system normally tolerates developing germ cells that carry surface molecules foreign to the rest of the body. When this delicate tolerance breaks down, immune cells may attack the very tissue they are meant to protect. Yet cataloguing which immune players matter, and which genes drive their behavior, has proven difficult because individual patient samples vary enormously and small studies frequently yield contradictory gene lists. The team behind the new work therefore chose an integrative strategy designed to separate robust signals from statistical noise.

At the foundation of the analysis lay transcriptomic data from four independent publicly available cohorts, catalogued under the identifiers GSE145467, GSE45885, GSE45887 and GSE9210. These datasets contain measurements of gene activity, typically derived from testicular tissue biopsies, in men with confirmed non-obstructive azoospermia and in controls with normal sperm production. Rather than analyzing a single cohort and hoping the findings generalize, the researchers combined all four, applying differential expression analysis to find genes whose activity consistently differs between patients and controls. This cross-cohort approach acts as a form of internal replication: a gene that appears dysregulated only because of the quirks of one laboratory’s sequencing run or one patient group’s characteristics is unlikely to survive scrutiny across four separate collections.

To move from a long list of differently behaving genes to an organized picture of the disease, the team then deployed weighted gene co-expression network analysis, a technique known as WGCNA. Instead of treating each gene as an isolated variable, WGCNA clusters thousands of genes into modules based on the similarity of their expression patterns across samples. Genes that rise and fall together are often governed by the same regulatory programs, so these modules can point to biological processes at the heart of a disease. Within the modules that the analysis linked most strongly to non-obstructive azoospermia, the researchers searched specifically for immune-related genes, arriving at sixteen candidate genes that were both differentially expressed and embedded in disease-associated co-expression networks.

Refining this sixteen-gene set into a practical diagnostic panel required a second layer of machine learning. The researchers ran five independent feature-selection algorithms: bagged trees, Bayesian methods, Boruta, least absolute shrinkage and selection operator regression, known as LASSO, and learning vector quantization. Each algorithm applies a different logic to the question of which genes carry the most information. LASSO, for example, shrinks the coefficients of less informative genes toward zero, effectively discarding them, while Boruta compares each real gene against randomized shadow features to test whether its association with disease exceeds chance. By seeking consensus across all five approaches, the team whittled the list down to a final twelve-gene panel they named Hub NOA.Sig, comprising CD14, CD19, CMTM2, CXCL16, HSPA1L, HSPA2, IGF1, IL13, POMC, RBP1, STAB2 and TYROBP.

The composition of this panel is itself revealing. Several of the genes are classic markers of immune lineages: CD14 flags monocytes and macrophages, CD19 marks B lymphocytes, and TYROBP encodes a signaling adaptor essential to the killing and activating receptors of natural killer cells and myeloid cells. CXCL16 is a chemokine that guides immune cell migration, while IL13 is a cytokine most familiar from allergy and parasite responses. Other genes speak more directly to the biology of germ cells and hormones: HSPA2 and HSPA1L are heat shock proteins with established roles in spermatogenesis, IGF1 supports cell growth and development, and POMC is a prohormone that yields biologically active peptides including beta-endorphin. CMTM2, a chemokine-like factor with testicular expression, and RBP1, a retinol-binding protein central to vitamin A transport, round out the panel. Vitamin A metabolism, it is worth noting, is indispensable for sperm production, giving RBP1 a plausible mechanistic role alongside its immune dimension.

To test whether the twelve-gene signature could actually classify samples, the researchers built a Random Forest classifier, an ensemble method that averages the votes of many decision trees trained on random subsets of the data. The classifier achieved an area under the receiver operating characteristic curve, or AUC, of 0.87. An AUC of 1.0 represents perfect discrimination between cases and controls, while 0.5 is equivalent to a coin flip, so a value of 0.87 indicates solid, though not flawless, discriminatory power. In a condition where clinical predictions currently rely on hormone measurements and histology, a transcriptomic classifier of this strength suggests that gene expression patterns capture disease-relevant information that conventional measures miss. The authors emphasize that the signature should serve as a transcriptomic reference for characterizing the heterogeneity of non-obstructive azoospermia rather than as a ready-made clinical test, and independent validation in new patient cohorts will be essential before any diagnostic application.

Beyond classification, the study probed the immune landscape itself using single-sample gene set enrichment analysis, or ssGSEA, a method that estimates the relative abundance of specific immune cell populations within a bulk tissue sample by scoring the expression of marker gene sets. This computational dissection revealed significant alterations in six immune cell populations in azoospermic testes, including effector memory CD4-positive and CD8-positive T cells, regulatory T cells and plasmacytoid dendritic cells. Effector memory T cells are the快速-response veterans of adaptive immunity, poised to react upon re-encounter with their targets, and their enrichment raises the possibility of localized immune activity against germ cells. Regulatory T cells, by contrast, are the immune system’s brakes, and shifts in their proportions could signal either an attempt to restore tolerance or a failure of that restraint. Plasmacytoid dendritic cells, potent producers of type I interferons, add another layer, hinting at innate immune activation. The strong correlations the researchers documented between individual hub genes and specific immune infiltrates tie the genetic signature directly to the cellular composition of the tissue.

The final arm of the study translated the gene list into therapeutic hypotheses. Through drug-gene interaction screening, the team identified candidate compounds predicted to act on signature genes, then used molecular docking, a computational technique that models how small molecules fit into protein binding pockets, to prioritize pairs. Three combinations emerged as leading candidates for laboratory validation: loperamide with POMC, acitretin with RBP1, and retinol with RBP1. The latter two pairs are particularly intriguing because acitretin is a synthetic retinoid and retinol is vitamin A itself, both operating in the retinoid pathway that governs germ cell differentiation. Loperamide, better known as an anti-diarrheal opioid receptor agonist, engaging the POMC axis is a more unexpected connection, but the docking predictions are explicitly framed as hypotheses requiring experimental testing rather than established drug effects on the testis.

What the study ultimately offers is a map rather than a cure. By integrating four cohorts, network analysis and consensus machine learning, the researchers have assembled a defensible, reproducible molecular handle on a condition that has resisted mechanistic study, and they have connected that handle to immune cells whose presence in the testicular environment may shape fertility outcomes. The limitations are those inherent to the approach: the analysis rests on existing datasets whose tissue composition and clinical annotation vary, computational immune deconvolution infers cell populations indirectly, and docking predictions are only as reliable as the structural models behind them. Still, the work charts a concrete path forward, from validating the twelve-gene panel in fresh cohorts to testing the flagged drug-gene pairs in cellular systems. If immune dysregulation proves to be a genuine driver of non-obstructive azoospermia rather than a passenger phenomenon, the signatures and targets described here could open routes to immunomodulatory or retinoid-based interventions for men whose infertility currently has few answers.

Subject of Research: An immune-related gene signature and immune cell alterations associated with non-obstructive azoospermia identified through integrated bioinformatics and machine learning.

Article Title: Integrated bioinformatics analysis identifies an immune-related gene signature associated with non-obstructive azoospermia

Article References: Zhong, Y., Ye, S., Lai, H., Zeng, X., Song, G., & Qiu, X. (2026). Integrated bioinformatics analysis identifies an immune-related gene signature associated with non-obstructive azoospermia. PLOS One, 21(10), e0360305. https://doi.org/10.1371/journal.pone.0360305

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360305

Keywords: non-obstructive azoospermia, male infertility, bioinformatics, machine learning, immune-related genes, WGCNA, Random Forest, ssGSEA, molecular docking, transcriptomics, testicular immunity, biomarkers