Uncovering sodium overload-associated gene signatures in chronic obstructive pulmonary disease through integrated bioinformatics and machine learning Authors: Zhongzhou Zou, Jietao Chen, Zutong Mai, Ningcheng Wu, Yuxuan Wang, Ruihong Xu, Zhonghui Jiang, Guoyong Chen, Lei Xian Published: 2026 ABSTRACT BackgroundAirway epithelial cells depend on tightly regulated ion transport, and disturbed sodium handling has been linked to mucus dehydration, oxidative stress and epithelial injury in obstructive lung disease. How sodium/ion-homeostasis-related transcription is organised in chronic obstructive pulmonary disease (COPD), and whether it converges with the disease’s broader expression architecture, has not been systematically examined. This study aimed to identify sodium-overload-associated genes in COPD and evaluate a compact epithelial gene panel across multiple validation platforms.MethodsTwo bulk lung-transcriptomic cohorts were integrated and analysed with differential expression, weighted gene co-expression network analysis (WGCNA) and a locked sodium-overload-related gene (SORG) set defined a priori from GeneCards. Candidate genes were characterised by functional enrichment and protein-protein interaction analysis, benchmarked across twelve machine-learning algorithms, and then examined in an independent external cohort, a mucus-score sensitivity analysis, single-cell RNA sequencing and quantitative polymerase chain reaction (qPCR; 13 COPD and 12 non-COPD controls).ResultsSeven genes lay at the intersection of differential expression, the COPD-associated co-expression module and the sodium-overload set, enriched for xenobiotic-metabolism and oxidative-stress pathways. A reduced three-gene panel (ALDH3A1, MUC5AC and PHEX) gave moderate but reproducible discrimination [area under the curve (AUC) 0.686], with acceptable calibration and positive net benefit over the threshold range examined, and reached an AUC of 0.809 in the external cohort. The three genes co-varied with multiple immune-cell fractions without implying direction. Single-cell analysis and qPCR localised the changes mainly to airway epithelial cells: ALDH3A1 and MUC5AC were consistently higher in COPD across datasets, whereas PHEX showed directional heterogeneity across platforms.ConclusionsSodium/ion-homeostasis-related transcription is associated with a stressed airway-epithelial state in COPD. The three-gene panel is best read as a compact, biologically interpretable framework—annotation-guided rather than evidence of causal sodium overload, and not a stand-alone diagnostic—that warrants validation in larger, prospectively matched cohorts. Highlight box Key findings • Integrating bulk transcriptomics, co-expression network analysis and a sodium-overload-related gene set identified a sodium/ion-homeostasis-associated epithelial-stress signal in chronic obstructive pulmonary disease (COPD), and a three-gene panel (ALDH3A1, MUC5AC, PHEX) with moderate, reproducible discrimination. What is known and what is new? • Disturbed ion transport and epithelial dysfunction are established features of COPD, but sodium/ion-homeostasis-related transcriptional changes had not been systematically mapped. • This study links sodium-overload-annotated genes to COPD co-expression networks, localises the signal to airway epithelial cells, and tests a compact three-gene panel across multi-algorithm modelling, external transcriptomic validation, a mucus-score sensitivity analysis, single-cell re-analysis and quantitative polymerase chain reaction—while showing that “sodium overload” here is best read as an annotation rather than a demonstrated mechanism. What is the implication, and what should change now? • The panel is a biologically interpretable framework for studying epithelial stress in COPD, not a stand-alone diagnostic; ALDH3A1 and MUC5AC behaved consistently across datasets, whereas PHEX showed cross-platform directional heterogeneity and needs protein-level confirmation. • Future work should validate the signal in larger, prospectively matched cohorts and in accessible specimens (sputum, nasal or bronchial brushings, exhaled breath condensate), with standardised assays and thresholds. Introduction Chronic obstructive pulmonary disease (COPD) is a leading cause of chronic respiratory morbidity and death, sustained by persistent airway inflammation, epithelial injury and progressive remodelling (1-4). Two decades of lung transcriptomics have catalogued large numbers of COPD-associated genes, yet a long list of differentially expressed transcripts rarely tells us which changes drive epithelial dysfunction and which are bystanders. The difficulty is partly interpretive: individual genes are easy to rank but hard to read as a coherent biological story (5,6). One way to narrow this gap is to combine differential expression with co-expression structure and an externally defined, pathway-informed gene set, so that the analysis is anchored to a biological theme rather than to isolated transcripts (7). Among the candidate themes, ion handling is central to the airway epithelium (8), which depends on tightly regulated transport for surface-liquid hydration, mucociliary clearance and barrier integrity (9,10). Sodium is of particular interest. Altered epithelial sodium channel (ENaC) activity, acquired cystic fibrosis transmembrane conductance regulator (CFTR) dysfunction and impaired Na+/K+-ATPase function have each been tied to mucus dehydration (11,12), oxidative stress and epithelial damage in obstructive airway disease (13,14), and experimental work on sodium-overload-related injury frames sodium stress as a plausible context for epithelial vulnerability (15). How this ionic dimension is wired transcriptionally in COPD, and whether it overlaps with the disease’s broader expression architecture, have not been examined systematically. We approached the question conservatively. Rather than assuming that sodium overload is a causal driver in COPD, we treated sodium-overload-related genes (SORGs) as an ion-homeostasis-oriented annotation space—a biologically motivated lens through which to read the disease transcriptome—and asked a single, testable question: do these annotations intersect with COPD-associated transcriptomic structure, and can that intersection prioritise epithelial stress-response genes worth validating? Framing the problem this way keeps the analysis interpretable without overstating what an annotation-based overlap can establish. To address this, we integrated two independent bulk lung-transcriptomic cohorts and combined differential expression with weighted gene co-expression network analysis (WGCNA) and a locked SORG set. The resulting candidates were characterised by functional enrichment and protein interactions, benchmarked across twelve machine-learning algorithms, and then examined in an external transcriptomic cohort, a mucus-score sensitivity analysis, single-cell re-analysis and quantitative polymerase chain reaction (qPCR). The aim throughout was modest and explicit: to define a compact, biologically interpretable candidate panel and to test how consistently it behaves across complementary layers of evidence, not to prove a causal role for sodium handling in COPD. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0843/rc). Methods Bulk transcriptomic datasets and preprocessing Two microarray datasets were used for discovery: GSE20257 (GPL570; 23 COPD, 112 controls; bronchial-brushing/airway-epithelial samples) and GSE76925 (GPL10558; 111 COPD, 40 controls; lung tissue). An independent small-airway cohort, GSE11906 (GPL570; 20 COPD, 79 controls), was held out for external validation and was not used for gene selection or model training. Within each dataset, probes were mapped to official symbols, collapsed by averaging, log2-transformed where needed and quantile-normalised. Merging on shared symbols gave 14,131 genes across 286 samples; dataset-level batch effects were removed with limma removeBatchEffect, with COPD/control status kept in the design. Correction adequacy and a ComBat sensitivity analysis are reported in table available at https://cdn.amegroups.cn/static/public/jtd-2026-0843-1.xlsx. Differential expression analysis Differentially expressed genes (DEGs) between COPD patients and control subjects were called with limma and empirical-Bayes moderation. Genes with an absolute log2 fold change (|log2FC|) >1 and Benjamini-Hochberg-adjusted P<0.05 were retained and shown as volcano and ranked fold-change plots. WGCNA Co-expression modules were built in WGCNA with a soft-thresholding power giving approximate scale-free topology (R2≥0.9). An unsigned network from pairwise Pearson correlations was clustered with dynamic tree cutting (minimum module size 50); modules with eigengene correlation above 0.75 were merged. The module most strongly correlated with COPD status was carried forward. Sodium-overload gene set and candidate gene selection SORGs were defined a priori, and solely, by a GeneCards query for “sodium overload”; after removing duplicate symbols, 2,388 unique genes formed a locked list carried unchanged through all subsequent filtering. GeneCards Relevance Scores were used for annotation and sensitivity analysis only—not as an inclusion criterion—and missing values were not imputed. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment of this list (below) characterised its biological content downstream of selection rather than building it; the locked list, its score tiers and the GeneCards re-mapping are in table available at https://cdn.amegroups.cn/static/public/jtd-2026-0843-2.xlsx. Candidate genes were then defined as the three-way overlap of COPD DEGs, the COPD-associated WGCNA module and the locked SORG list, visualised by a Venn diagram. Definition of sodium overload-associated candidate genes Candidate genes were defined as the three-way overlap of COPD-associated DEGs, the COPD-linked WGCNA module and the SORG list. Intersection results were visualised by Venn diagram. Functional enrichment and protein-protein interaction analyses GO and KEGG enrichment [clusterProfiler with org.Hs.eg.db; Benjamini-Hochberg false discovery rate (FDR) <0.05] was applied separately to the full locked SORG list, to characterise its biological content, and to the seven candidate genes; enriched terms were shown as ranked bar plots (16). Protein-protein interactions among the seven candidates were retrieved from STRING (Homo sapiens); external interactors were not added, and the exported associations among the seven seed genes were visualised, with edge thickness reflecting the combined score. Machine learning-based feature selection and model evaluation The seven candidate genes were evaluated across twelve supervised classifiers (linear and regularised, distance-based, probabilistic, neural-network and tree-ensemble), chosen to test feature behaviour under differing modelling assumptions rather than to select one best classifier. Samples were split 70/30 with class proportions preserved, and hyperparameters were tuned by five-fold cross-validation within the training set. Model-agnostic feature importance analysis Permutation-based feature importance was computed across all trained classifiers; for each model, each gene was permuted in turn and the increase in root-mean-square error (RMSE) recorded. Genes recurring across models were prioritised, while marked inter-algorithm variability was read as context-dependence. Model settings and algorithm-level metrics are in table available at https://cdn.amegroups.cn/static/public/jtd-2026-0843-3.xlsx. Construction of a parsimonious gene signature and nomogram A reduced three-gene panel was selected by combining classification gain, biological coverage and feasibility for downstream validation. A multivariable logistic-regression model on normalised expression (COPD as outcome) was fitted and displayed as a nomogram for individual-level estimation, with bootstrap internal validation, calibration against observed probabilities and decision-curve analysis of net benefit. Immune cell infiltration analysis Immune-cell composition in the bulk data was estimated with CIBERSORT (LM22 reference signature); only samples with a deconvolution P<0.05 were retained for downstream analysis. Differences in immune-cell proportions between COPD and controls were tested non-parametrically, pairwise correlations among subsets described co-infiltration, and Spearman correlations linked the three genes to immune-cell fractions. External validation and mucus-score sensitivity analysis External validation used small-airway samples from GSE11906, with healthy non-smokers and healthy smokers combined as controls; expression and receiver operating characteristic (ROC) performance were assessed for each gene and the combined score. To probe mucus bias, a mucus score (mean Z-score of MUC5AC, SPDEF and AGR2) was computed in the discovery cohort; samples were split at the cohort median, and panel performance was tested within matched mucus-score strata by five-fold cross-validation. A MUC5AC-ablation analysis compared the three-gene model with an ALDH3A1/PHEX two-gene model. Single-cell RNA-seq processing and analysis Bronchial brushing scRNA-seq data (GSE173896; COPD, non-COPD smokers and never-smokers) were processed with Scanpy. Cells with 500–6,000 detected genes and <20% mitochondrial reads were retained. After per-cell normalisation and log-transformation, highly variable gene selection, regression of sequencing depth and mitochondrial content, and principal component analysis (PCA)-based dimensionality reduction were performed sequentially. Batch effects were corrected with Harmony, and cells were clustered by the Leiden algorithm on a k-nearest-neighbour graph, with uniform manifold approximation and projection (UMAP) for visualisation. Cell types were annotated with CellTypist, refined by canonical markers and consolidated into nine classes. Differential expression of the candidate genes across conditions and cell types used Wilcoxon rank-sum tests; for sensitivity, epithelial cells were split into Goblet, Ciliated, Basal, Club, AT1 and AT2 subtypes, and given MUC5AC sparsity the three genes were compared within each subtype rather than by MUC5AC binning. RNA extraction and qPCR Lung tissue samples were collected from patients undergoing pulmonary resection at The Second Affiliated Hospital of Guangxi Medical University. The samples were obtained from macroscopically normal lung tissue located away from the tumor. Total RNA was extracted using a commercial kit. One microgram of RNA per sample was reverse-transcribed and amplified by SYBR-Green qPCR, with glyceraldehyde-3-phosphate dehydrogenase (GAPDH) as the internal reference and reactions run in technical triplicate. Primer sequences and the full clinical and pathological characteristics of the cohort—age, sex, smoking status, pack-years, Global Initiative for Chronic Obstructive Lung Disease (GOLD) stage, body mass index (BMI), histological type, tumour size and tumor-node-metastasis (TNM) stage—are provided in table available at https://cdn.amegroups.cn/static/public/jtd-2026-0843-4.xlsx. Ethics approval and consent to participate The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Medical Ethics Committee of The Second Affiliated Hospital of Guangxi Medical University (No. 2023-KY-0476). Written informed consent was obtained from all participants for the use of their tissue samples and clinical data in this study. All clinical data were de-identified prior to analysis, and patient privacy was strictly protected. Statistical analysis Relative expression was calculated by the 2−ΔΔCt method with GAPDH as reference and summarised as mean ± standard error of the mean (SEM). Between-group comparisons were performed on ΔCt values using two-tailed unpaired t-tests, with P<0.05 considered significant. Analyses were performed in R (v4.5.2). Results Differential expression and co-expression network analysis We first mapped the COPD transcriptional landscape across the two integrated cohorts. Differential expression identified 1,280 genes separating COPD from control tissue (|log2FC| >1, adjusted P<0.05), summarised as ranked fold-change and volcano plots (Figure 1A,1B). Rather than analysing these genes individually, we constructed a weighted co-expression network: a soft-threshold power giving approximate scale-free topology was chosen (Figure 1C), and dynamic tree-cutting with eigengene-based merging resolved several co-expression modules (Figure 1D). The yellow module (86 genes) was the most strongly correlated with disease status (cor =0.814, P=1.7×10−21; Figure 1E), and within which genes with higher module membership also showed higher gene significance for COPD (Figure 1F). This module therefore captures a coherent, disease-linked expression signal and provided the focused gene space for the sodium/ion-homeostasis prioritisation that follows. Differential expression and a COPD-associated co-expression module. (A) Genome-wide ranking of DEGs between COPD and control lung tissue by log2 fold change. (B) Volcano plot of up- and down-regulated genes (|log2FC| >1, adjusted P<0.05). (C) Scale-free topology fit used to choose the WGCNA soft-thresholding power. (D) Gene dendrogram and module assignment before and after eigengene-based merging. (E) Module-trait correlation identifying the yellow module as most associated with COPD status (cor =0.814). (F) Gene significance versus module membership within the COPD-associated module. COPD, chronic obstructive pulmonary disease; DEG, differentially expressed gene; FC, fold change; WGCNA, weighted gene co-expression network analysis. Because the two discovery datasets came from different platforms, we confirmed that batch effects were controlled: dataset-driven separation on PCA was much reduced after limma correction while COPD/control structure was preserved (Figure S1), and the dataset variance component fell from 99.76% to 1.09% (table available at https://cdn.amegroups.cn/static/public/jtd-2026-0843-1.xlsx). The three final genes also remained differentially expressed under an alternative ComBat correction (FDR <0.05; table available at https://cdn.amegroups.cn/static/public/jtd-2026-0843-1.xlsx). Sodium/ion-homeostasis gene set and candidate gene identification Before intersecting the SORG list with the COPD data, we confirmed what the list captures. GO analysis of the 2,388 locked SORGs was dominated by sodium-ion transport, transmembrane transport, homeostasis, membrane depolarisation and channel activity (Figure 2A; full theme-grouped enrichment in Figure S2A), so the set behaves as a sodium/ion-homeostasis annotation space rather than an arbitrary list. KEGG terms were broader, spanning signalling and disease pathways (Figure S2B), and the sodium-related GO enrichment held across GeneCards Relevance Score-defined subsets (Figure S2C). Curation of a sodium/ion-homeostasis gene set and identification of seven COPD candidate genes. (A) Sodium-related GO enrichment of the 2,388 locked SORGs, spanning sodium-ion transport, homeostasis, transmembrane transport, membrane depolarisation and channel activity. (B) Venn diagram of COPD DEGs, the COPD-associated WGCNA module and the SORG list, yielding seven intersecting genes. (C) GO/KEGG enrichment of the seven genes, dominated by detoxification, oxidative-stress and xenobiotic-metabolism terms. (D) STRING PPI network of the seven genes. BP, biological process; CC, cellular component; COPD, chronic obstructive pulmonary disease; DEG, differentially expressed gene; FDR, false discovery rate; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; MF, molecular function; PPI, protein-protein interaction; SORG, sodium-overload-related gene; WGCNA, weighted gene co-expression network analysis. Intersecting COPD DEGs, the COPD-associated module and the locked SORG list left seven genes—CYP1A1, ALDH3A1, SPP1, NQO1, MUC5AC, UCHL1 and PHEX (Figure 2B). This small overlap (the 0.2% region in Figure 2B, or 7 of 2,388 SORGs) follows from requiring all three conditions and does not bound the breadth of sodium-related biology. The seven genes were enriched for detoxification, oxidative-stress and xenobiotic-metabolism terms, including cytochrome P450-mediated xenobiotic metabolism and aldehyde dehydrogenase activity (Figure 2C), and the exported STRING associations among them formed a connected network (Figure 2D). Under stricter GeneCards thresholds the overlap shrank, with ALDH3A1 and MUC5AC entering only at lower stringency or with the full locked list (Figure S2D). Machine-learning benchmarking of the candidate genes The seven candidates were passed to twelve supervised classifiers spanning linear, probabilistic, distance-based, neural-network and ensemble families. On absolute-residual distributions, the tree ensembles (random forest, gradient boosting and XGBoost) sat at lower median error with tighter spread than naive Bayes and the single decision tree (Figure 3A), and Figure 3B shows the proportion of observations with absolute residuals exceeding each cutoff. Permutation-based importance varied by algorithm: ALDH3A1 and PHEX recurred in the most model-specific importance outputs, while the other genes were more model-dependent (Figure 3C). Because no single gene dominated every model, the reduced panel was not chosen on importance ranking alone. Machine-learning benchmarking and model-agnostic feature importance. (A) Distribution of absolute residuals across twelve supervised classifiers. (B) Reverse cumulative distribution of absolute residuals, showing the percentage of observations with absolute residuals exceeding each threshold. (C) Permutation-based feature importance across classifiers; ALDH3A1 and PHEX recurred most consistently, while the other genes were more model-dependent. ADA, adaptive boosting; BAG, bagging; DT, decision tree; GBM, gradient boosting machine; GLM, generalized linear model; KNN, k-nearest neighbors; LASSO, least absolute shrinkage and selection operator; NB, naïve Bayes; NNET, neural network; RF, random forest; SVM, support vector machine; XGB, extreme gradient boosting. A parsimonious three-gene diagnostic model We then reduced the candidates to ALDH3A1, MUC5AC and PHEX, chosen by combining joint classification gain, coverage of the detoxification-oxidative-secretory axis, and the practicality of confirming each gene by single-cell analysis and qPCR, rather than by importance ranking alone. Across the algorithm set, ensemble methods again tended to give the higher areas under the curve (AUCs) for this reduced panel (Figure 4A). For the final model we fitted a parsimonious three-gene logistic regression, which gave moderate discrimination (AUC =0.686; Figure 4B), a balanced confusion matrix (Figure 4C) and usable precision across recall levels (Figure 4D). After bias correction, calibration was acceptable (Figure 4E), and decision-curve analysis showed positive net benefit over treat-all and treat-none strategies across the threshold range examined (Figure 4F). Each gene alone discriminated weakly (ALDH3A1 0.620, MUC5AC 0.666, PHEX 0.615); the combined panel modestly exceeded the best single gene (0.686; Figure 4G), consistent with partly complementary signals. The model was then rendered as a nomogram for individual-level probability estimation (Figure 4H). Train/test performance, calibration indices, decision-curve net benefits and selected hyperparameters for all twelve classifiers are in table available at https://cdn.amegroups.cn/static/public/jtd-2026-0843-3.xlsx. Development and evaluation of a reduced three-gene model for COPD. (A) ROC curves of the three-gene panel across classifiers. (B) ROC curve of the three-gene logistic model (AUC =0.686). (C) Confusion matrix for the three-gene model. (D) Precision-recall curve under class imbalance. (E) Calibration plot of predicted versus observed probabilities. (F) Decision-curve analysis of net benefit across threshold probabilities. (G) ROC curves of individual genes versus the combined panel (ALDH3A1 0.620, MUC5AC 0.666, PHEX 0.615; combined 0.686). (H) Nomogram for individual-level COPD probability from the three genes. **, P<0.01. ADA, adaptive boosting; AUC, area under the curve; BAG, bagging; con, control; COPD, chronic obstructive pulmonary disease; DT, decision tree; Freq, frequency; GBM, gradient boosting machine; KNN, k-nearest neighbors; LASSO, least absolute shrinkage and selection operator; NB, naïve Bayes; NNET, neural network; PR, precision-recall; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine; tre, treatment; XGB, extreme gradient boosting. Immune infiltration and gene-immune associations Immune deconvolution showed heterogeneous immune-cell composition across samples, with several subsets differing in relative abundance between COPD and controls (Figure 5A) and structured co-infiltration among populations (Figure 5B). ALDH3A1, MUC5AC and PHEX each correlated with multiple immune subsets (|r| ≥0.2, P<0.05; Figure 5C), suggesting co-variation between these epithelial-stress genes and the local immune milieu without establishing directionality or mechanism. Immune infiltration and gene–immune associations in COPD. (A) CIBERSORT-estimated immune-cell proportions in control and COPD lung tissue. (B) Spearman correlation matrix of immune-subset co-infiltration. (C) Correlation network between the three genes (ALDH3A1, MUC5AC, PHEX) and immune-cell fractions; edge thickness reflects correlation strength. *, P<0.05; **, P<0.01; ***, P<0.001; ns, not significant. COPD, chronic obstructive pulmonary disease; NK, natural killer. External validation and mucus-score sensitivity We next tested the three-gene panel in the independent GSE11906 small-airway cohort. ALDH3A1, MUC5AC and PHEX were each higher in COPD than in controls, and the combined three-gene score reached an AUC of 0.809 (Figure 6A,6B), supporting external reproducibility of the expression pattern. External validation and mucus-phenotype sensitivity of the three-gene panel. (A) Expression of ALDH3A1, MUC5AC and PHEX in the independent GSE11906 small-airway cohort (20 COPD, 79 controls). (B) ROC of individual genes and the combined three-gene score in GSE11906 (combined AUC =0.809). (C) Five-fold cross-validated AUCs within matched high- and low-mucus strata. (D) ROC of the three-gene model versus the ALDH3A1/PHEX two-gene model after MUC5AC removal (AUC 0.686 → 0.650). (E) ALDH3A1 expression across non-Goblet epithelial subtypes in the single-cell data (subtypes shown at FDR <0.05). AUC, area under the curve; COPD, chronic obstructive pulmonary disease; CV, cross-validated; FC, fold change; FDR, false discovery rate; ROC, receiver operating characteristic; SD, standard deviation. Because MUC5AC is closely tied to mucus hypersecretion, we asked whether the panel was driven by a transcriptomic mucus signal. Using a transcriptomic mucus score (MUC5AC, SPDEF, AGR2), COPD and control samples were split at the cohort-wide median; the panel retained above-chance discrimination in both matched mucus-score strata, including the low-mucus comparison (Figure 6C). Removing MUC5AC lowered the AUC only modestly, from 0.686 to 0.650 (Figure 6D), and ALDH3A1 was higher across several non-Goblet epithelial subtypes (Figure 6E). The panel is therefore not solely attributable to MUC5AC-driven mucus hypersecretion. Single-cell localisation and qPCR validation To anchor the bulk findings at cellular resolution, we re-analysed the bronchial-brushing scRNA-seq dataset GSE173896. After quality control, 36,405 cells were retained and integrated with Harmony, giving well-mixed embeddings (Figure 7A,7B), and annotation resolved major immune, stromal and epithelial populations (Figure 7C,7D). At the single-cell level, ALDH3A1 and MUC5AC were higher in COPD and PHEX was lower, with these shifts localised mainly to epithelial cells (Figure 7E). Reverse transcription quantitative polymerase chain reaction (RT-qPCR) in an independent local cohort of airway epithelial samples (12 controls, 13 COPD) reproduced the single-cell direction for all three genes (Figure 7F). Of note, PHEX was higher in COPD in the bulk discovery and external (GSE11906) cohorts but lower at single-cell level and by qPCR; PHEX should therefore be read as a transcriptomic candidate with cross-platform directional heterogeneity rather than a uniformly directional marker, whereas ALDH3A1 and MUC5AC were consistently higher in COPD across all layers. Single-cell localisation and qPCR of the three genes in COPD. (A) UMAP of integrated scRNA-seq data (36,405 cells) by disease group. (B) UMAP of quality-control metrics after integration. (C) Canonical-marker expression defining major lung cell populations. (D) UMAP coloured by annotated cell type. (E) Single-cell expression of ALDH3A1, MUC5AC and PHEX by disease state. (F) qPCR of the three genes in an independent airway epithelial cohort (12 controls, 13 COPD). *, P<0.05; ***, P<0.001; ****, P<0.0001. COPD, chronic obstructive pulmonary disease; DC, dendritic cells; ILC, innate lymphoid cells; pDC, plasmacytoid dendritic cells; qPCR, quantitative polymerase chain reaction; scRNA-seq, single-cell RNA sequencing; UMAP, uniform manifold approximation and projection. Discussion This study asked whether an ion-homeostasis annotation space could help prioritise epithelial-stress genes in COPD. Intersecting COPD differential expression, a disease-associated co-expression module and a locked SORG set returned seven candidates enriched for detoxification, oxidative-stress and xenobiotic-metabolism pathways (17,18), from which a three-gene panel—ALDH3A1, MUC5AC and PHEX—was distilled (19). The panel showed moderate internal discrimination, reproducible expression changes in an external small-airway cohort, and a direction consistent with single-cell and qPCR data for two of its three members. These results connect sodium/ion-homeostasis annotations to a stressed airway-epithelial state in COPD. Because the three-way overlap was no larger than expected by chance in permutation testing, however, we read it as annotation-guided rather than as evidence that sodium overload causally shapes the COPD transcriptome (20). The small size of the overlap deserves comment, since seven genes are only a fraction of a percent of those considered. This is a property of the filter, not a statement about the reach of sodium-related biology: requiring a gene to be differentially expressed, to sit within a COPD-associated module and to appear in the predefined list deliberately traded breadth for cross-layer support and interpretability. The contribution is therefore less the nomination of ALDH3A1, MUC5AC or PHEX as novel COPD genes than the placement of these genes within a sodium/ion-homeostasis-informed stress framework, and the systematic test of that framework across bulk, single-cell, external and qPCR layers. Consistent with this reading, the three genes are not canonical sodium transporters and did not rank highly for sodium-overload relevance—ALDH3A1 and MUC5AC fell in the low GeneCards tier and PHEX in the medium tier, and under stricter relevance thresholds ALDH3A1 and MUC5AC dropped out of the overlap entirely (21). They are best understood as downstream epithelial-stress readouts that surface within this annotation space (22,23), not as direct mediators of ionic stress. From a modelling standpoint, the three-gene model is an internally evaluated, exploratory tool rather than a clinical test (24). Its moderate AUC of 0.686 is unsurprising for a small panel without clinical covariates, and its value is interpretive: the combined panel only modestly exceeded the best single gene, pointing to partly complementary information rather than a deployable classifier. The twelve-algorithm benchmark was used to probe how the candidate features behaved under different assumptions, not to crown a winner—AdaBoost, gradient boosting and naïve Bayes in fact reached the highest held-out AUCs—and the model we report is a parsimonious logistic regression rendered as a nomogram (25), chosen for transparency over optimised performance (26). Full algorithm-level metrics are provided so that this behaviour can be inspected directly (27). Placed in its immune context, all three genes co-varied with several deconvolution-estimated immune fractions, in keeping with the well-described cross-talk between airway epithelium and immune cells in COPD (28-30). Deconvolution returns relative proportions and cannot assign direction, so whether epithelial signals shape immune composition or the reverse cannot be resolved here (31,32), and we have not implied a mechanism. Single-cell re-analysis and independent qPCR placed the transcriptional changes predominantly in airway epithelial cells (33,34), which argues against the bulk signal being a simple artefact of shifting cell composition (35,36). The three genes did not, however, behave identically across platforms, and this is worth stating plainly. ALDH3A1 and MUC5AC were higher in COPD at every level examined—bulk, external, single-cell and qPCR—whereas PHEX was higher in the bulk and external cohorts but lower at single-cell resolution and by qPCR. We therefore treat PHEX as a candidate with cross-platform directional heterogeneity rather than a stable directional marker; the discrepancy may reflect differences in tissue composition, platform sensitivity or cell-type-specific expression, and it tempers any claim that the panel behaves uniformly. Two aspects of the design lend some weight to these findings. Anchoring candidate selection to an externally defined gene set, in addition to differential expression and co-expression structure, limits the data-driven overfitting that can affect purely supervised pipelines (37). And the convergence of bulk, single-cell and qPCR evidence—at least for ALDH3A1 and MUC5AC—gives cross-scale support that no single data type could provide on its own (38). Several limitations bound these conclusions. The discovery datasets are retrospective and were generated on different platforms and tissue sources; batch correction reduced dataset-driven variance substantially and the final genes survived an alternative ComBat adjustment, but residual platform or tissue heterogeneity cannot be excluded. The sodium-overload label derives from database annotation rather than direct perturbation of sodium handling, and, as noted, the three-way overlap was not statistically enriched, so the term should be read as an annotation rather than a demonstrated mechanism. The qPCR cohort was small and drawn entirely from patients undergoing resection for early-stage primary lung cancer, so airway epithelium was sampled from tumour-distant, macroscopically normal tissue rather than from healthy, non-surgical individuals—an unavoidable feature of tissue-based airway studies—and a field-cancerisation effect on the adjacent epithelium cannot be excluded. More importantly, smoking status was completely confounded with disease status in this cohort: all COPD cases were current smokers and all controls never-smokers, and the groups also differed in sex. The qPCR data therefore cannot, on their own, separate COPD-associated from smoking- or sex-associated expression differences and are best read as supportive transcript-level evidence rather than COPD-specific validation. This is partly mitigated by the other analytic layers, since the external cohort (GSE11906) included healthy smokers among its controls and the single-cell dataset (GSE173896) contained non-COPD smokers; the consistent up-regulation of ALDH3A1 and MUC5AC in COPD is thus supported where smoking is not fully confounded with disease, although confirmation in prospectively recruited cohorts matched for age, sex and smoking remains necessary. The mucus-score analysis used a transcriptomic surrogate built from MUC5AC, SPDEF and AGR2 rather than a clinically adjudicated chronic-bronchitis phenotype (39). Finally, validation was at the transcript level only; no protein measurements were made, a gap that matters most for PHEX, whose direction was already inconsistent across platforms and which will require independent protein confirmation. The translational reading should be correspondingly cautious. The panel is not a stand-alone diagnostic but a molecular lens on epithelial stress, and its compact size makes it suitable for targeted measurement in accessible specimens—induced sputum, nasal or bronchial brushings, or exhaled breath condensate—where epithelial signals are sampled more directly than in peripheral blood (40,41). Before any such use, pre-analytical handling, assay platform, normalisation strategy and clinically meaningful thresholds would need to be standardised, and airway-derived measurements compared against established peripheral markers and clinical phenotypes in prospective cohorts (42). More broadly, finding ion-homeostasis-related transcription embedded within COPD co-expression networks invites a re-examination of ion-metabolic stress as one axis of epithelial dysfunction—a question with clear relevance to ongoing work on CFTR modulators and ENaC inhibitors in non-cystic-fibrosis obstructive lung disease (43,44). Conclusions This study provides an annotation-guided analysis linking sodium/ion-homeostasis-related gene space to COPD-associated epithelial transcriptional change. A three-gene panel (ALDH3A1, MUC5AC, PHEX) was supported by bulk transcriptomics, external small-airway validation, mucus-phenotype sensitivity analysis, single-cell re-analysis and qPCR. Rather than a definitive diagnostic or proof of causal sodium overload, the panel offers a biologically interpretable framework for studying epithelial stress in COPD and for guiding future mechanistic and clinically oriented validation. Supplementary The article’s supplementary files as Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0843/prf