Two-Gene Blood Signature Tracks Immune States in Sepsis

Researchers have built a compact two-gene signature from growth hormone–associated transcriptional programs that distinguishes sepsis from non-septic controls across independent public transcriptomic cohorts, and the same two genes—JAK1 and GATA3—appear to report on the immune state of the patient rather than serving only as binary disease classifiers. The study, published in Immunity, Inflammation and Disease, combined weighted gene co-expression network analysis with four machine-learning algorithms to narrow more than a thousand candidate genes down to a reproducible pair whose lower expression in sepsis coincided with weaker lymphocyte activity and relatively stronger myeloid and neutrophil signals.

Sepsis, defined as organ dysfunction caused by a dysregulated host response to infection, is biologically far from uniform. Inflammatory and counter-regulatory programs can be active at the same time, while antigen presentation, leukocyte abundance, lymphocyte function, cellular metabolism, and tissue homeostasis all change as illness evolves. Lymphocyte apoptosis and dysfunction, reduced antigen-presenting capacity, altered co-stimulatory signaling, T-cell exhaustion, and expansion of immunoregulatory or myeloid populations can produce sustained immune suppression in a subset of patients. This variation in host response is one reason clinically similar patients can diverge substantially in their subsequent course, and molecular markers that capture host immune state may provide information not fully represented by conventional clinical assessment.

The biological rationale for examining growth hormone–related genes rests on the close intertwining of endocrine and metabolic adaptation with the immune response during critical illness. Canonical signaling through the growth hormone receptor is centered on JAK2 and downstream STAT, PI3K, and ERK pathways, mechanistically distinct from the JAK1-dependent signaling used by many cytokine receptors. During experimental endotoxemia, inflammatory stress can produce postreceptor growth hormone resistance, with reduced GH-stimulated JAK2/STAT5 activity and increased suppressors of cytokine signaling. Growth hormone can also influence immune-cell cytokine production. The authors stress that these observations do not imply altered GH signaling causes sepsis; indeed, randomized evidence has shown increased mortality after high-dose GH administration in adults with prolonged critical illness.

To assemble the training data, the team merged two Gene Expression Omnibus microarray datasets: GSE57065, contributing 107 samples (82 sepsis and 25 non-septic controls), and GSE95233, contributing 124 samples (102 sepsis and 22 controls). After batch adjustment with limma, the merged training set contained 231 samples: 184 sepsis cases and 47 controls. Two further datasets, GSE26378 with 82 sepsis and 21 control samples and GSE28750 with 10 sepsis and 20 controls, were reserved exclusively for external validation. Growth hormone–associated genes were retrieved from GeneCards by searching “Growth hormone” with a relevance-score cutoff above 15, yielding 1205 genes.

Differential expression analysis in the merged cohort identified 2552 genes with adjusted p-values below 0.05 and absolute log2 fold change above 0.585, including 1239 upregulated and 1313 downregulated genes. Weighted gene co-expression network analysis, using the 5000 most variable genes and a soft-thresholding power of β = 6 that achieved a scale-free topology fit of R = 0.85, then partitioned the transcriptome into modules. Four module eigengenes were positively related to sepsis status—blue (r = 0.88), black (r = 0.38), green (r = 0.36), and salmon (r = 0.31)—together containing 1873 genes. Functional annotation placed many of these genes in adaptive-immune and lymphocyte programs, including antigen-receptor signaling, T-cell differentiation, and lymphocyte differentiation, with KEGG results emphasizing T-cell receptor signaling, antigen processing and presentation, hematopoietic cell lineage, and Th17 and Th1/Th2 differentiation.

Intersecting the 1873 module genes with the 2552 differentially expressed genes and the 1205 growth hormone–related genes left 40 shared genes, termed DEGHRGs. A protein interaction network built in STRING and screened with MCODE identified one densely interconnected subnetwork of 23 genes, retained as candidate hubs. These hubs were annotated to hematopoietic regulation, alpha-beta T-cell differentiation, myeloid differentiation, intrinsic apoptotic signaling, and JAK-STAT signaling pathways. The 23 genes were then challenged under four feature-selection procedures with different inductive biases: LASSO retained 10 genes, XGBoost selected 5 with GATA3 showing the highest feature importance followed by FYN, ETS1, PARP1, and JAK1, SVM-RFE converged on an 8-gene set, and random forest retained 11 genes by mean decrease in the Gini index. Their intersection contained only JAK1 and GATA3. Because the discovery datasets contain longitudinal samples from the same patients, the machine-learning analyses were repeated under a patient-level 10-fold cross-validation scheme, with all samples from the same patient assigned to the same fold, and this again yielded JAK1 and GATA3.

Both markers discriminated sepsis from controls in the training cohort with AUC values above 0.70, and the combined JAK1/GATA3 model performed best. Both transcripts were lower in sepsis than in controls, and the same direction of change was retained in GSE26378 and GSE28750, where the combined classifier also remained informative. A nomogram translated the fitted model into individual diagnostic probabilities, decision-curve analysis showed positive net benefit across a broad threshold range, and calibration indicated close correspondence between fitted and observed probabilities. The authors caution, however, that the exceptionally high AUCs require restraint: the datasets are retrospective case–control transcriptomic cohorts, and GSE28750 contains only 30 samples, so performance under these conditions may overestimate what could be achieved in prospective clinical populations where sepsis must be distinguished from heterogeneous inflammatory and infectious mimics rather than from predefined controls.

Immune deconvolution from three complementary perspectives—single-sample gene set enrichment analysis, CIBERSORT, and MCP-counter—converged on a broadly similar shift from lymphoid-associated activity toward stronger myeloid and neutrophil-related signals in sepsis. Scores for activated dendritic cells, B cells, CD8+ T cells, natural killer cells, T-helper cells, and other T-cell subsets were generally lower than in controls, whereas dendritic-cell, Treg, macrophage, and neutrophil signatures tended to be higher. JAK1 expression rose with many lymphocyte-associated features, including T-cell co-stimulation, HLA activity, checkpoint scores, CD8+ T cells, and cytolytic activity, but decreased as macrophage-related estimates increased, and GATA3 showed a similar pattern with a strong inverse association with macrophage signatures. Single-gene enrichment analysis additionally linked JAK1 to mitochondrial electron-transfer and bioenergetic programs alongside immune pathways, though the data cannot establish whether JAK1 regulates these programs or whether both signals simply track disease severity.

Consensus clustering of the septic training samples using only JAK1 and GATA3 expression favored a two-group solution. Subtype 1 was JAK1/GATA3-high and more adaptive-immune active, showing higher scores for B cells, CD8+ T cells, NK cells, T-helper cells, Th1 and Th2 cells, and Tregs, along with stronger co-stimulation, cytolytic activity, HLA, and type I interferon-response scores, whereas Subtype 2 displayed weaker adaptive signatures and a vesicle/granule-oriented transcriptional pattern. The authors emphasize that these classes were generated from only two variables and should be described as JAK1/GATA3-defined expression subgroups rather than a new universal sepsis taxonomy, since published sepsis endotypes generally rely on much broader signatures and independent replication.

Experimental support came from an LPS-treated human umbilical vein endothelial cell model, in which RT-qPCR showed that 300 ng/mL LPS exposure for 24 hours markedly reduced the mRNA levels of both JAK1 and GATA3 compared with untreated cells (both p < 0.0001), a direction consistent with the transcriptomic cohorts. The authors note this model represents endothelial inflammation rather than the full spectrum of sepsis and does not directly confirm the genes’ proposed roles in T-cell–mediated adaptive immunity. They also stress that JAK1 should not be described as the canonical kinase directly mediating growth hormone receptor signaling, since the classical GH receptor signals primarily through JAK2; JAK1 is better interpreted as part of a broader GH-related molecular context. Database-derived regulatory and compound associations, including predicted transcription factors such as E2F1 and candidate compounds from DSigDB, should remain explicitly hypothesis-generating, as they indicate molecular connectivity only and provide no evidence of treatment efficacy or safety. The signature, the authors conclude, should be regarded as an externally reproducible transcriptomic classifier rather than a clinically validated diagnostic test, with prospective multicenter validation and cell-resolved functional work needed before any clinical translation.

Subject of Research: A machine-learning analysis of growth hormone–related gene expression identifies JAK1 and GATA3 as a diagnostic signature and markers of immune heterogeneity in sepsis.

Article Title: Growth Hormone‐Related Gene Signatures Reveal Diagnostic Biomarkers, Immune Landscapes, and Potential Therapeutic Targets in Sepsis Through WGCNA and Machine Learning Analyses

Article References: Li, X., Liu, A., Ding, Y., Chen, G., & Zhang, B. (2026). Growth Hormone‐Related Gene Signatures Reveal Diagnostic Biomarkers, Immune Landscapes, and Potential Therapeutic Targets in Sepsis Through WGCNA and Machine Learning Analyses. Immunity, Inflammation and Disease, 14(10), Article e70566. https://doi.org/10.1002/iid3.70566

Image Credits: AI Generated

DOI: 10.1002/iid3.70566

Keywords: sepsis, JAK1, GATA3, diagnostic biomarkers, gene expression, machine learning, WGCNA, immune deconvolution, growth hormone, transcriptomics, host response, molecular subtypes