Blood transcriptomic signatures distinguish Mycobacterium avium complex pulmonary disease from colonization: a multicenter cohort study Authors: Wen-Yuan Lin, Hung-Ling Huang, Chia-Ni Hsiung, Ya-Yun Huang, Meng-Rui Lee, Meng-Hsuan Cheng, Inn-Wen Chong, Jann-Yuan Wang Published: 2026 ABSTRACT ABSTRACTPulmonary infections caused by nontuberculous mycobacteria (NTM), particularly Mycobacterium avium complex (MAC), are increasingly recognized as an important clinical entity, yet distinguishing active pulmonary disease from asymptomatic colonization remains challenging because current diagnosis relies on composite criteria. In this study, we aimed to identify blood transcriptomic signatures that discriminate MAC pulmonary disease (MAC-PD) from pulmonary colonization (MAC-PC) and to evaluate their potential as candidate biomarkers. MAC-positive patients from two medical centres in Taiwan were enrolled as training and external validation cohorts, and peripheral blood transcriptomes were profiled. Candidate genes were identified using least absolute shrinkage and selection operator regression and recursive feature elimination across multiple random seeds, followed by filtering based on statistical robustness and biological relevance. Machine-learning models were then trained and externally validated. Among 120 patients (training cohort: 46 MAC-PD and 28 MAC-PC; validation cohort: 25 MAC-PD and 21 MAC-PC), seven enriched gene ontology terms were prioritized. Three models demonstrated robust performance in the validation cohort, with areas under the receiver operating characteristic curve of 0.78, 0.78, and 0.75, and accuracies of 0.76, 0.74, and 0.74, respectively. These models shared a five-gene core signature consisting of IGKV1D-39, IGKV6-21, OVCH1, PLAU, and DMD, highlighting convergent biological signals related to immune responses and tissue remodelling. Overall, blood transcriptomic profiling shows promise in differentiating MAC-PD from MAC-PC in our cohorts, and the identified five-gene core signature represents a biologically coherent, minimally invasive candidate biomarker panel warranting further prospective validation.KEYWORDS: Mycobacterium avium complex, nontuberculous mycobacteria, pulmonary disease, transcriptome, gene expression profiling, biomarkers Introduction Pulmonary disease (PD) caused by nontuberculous mycobacteria (NTM) is an emerging global health threat. Its incidence has increased by 3–8% annually in many regions, and in several high-income countries the prevalence now surpasses that of tuberculosis [1–3]. Despite prolonged multidrug therapy, outcomes remain poor, with five-year all-cause mortality rates reaching 25–40% [4–6]. The distribution of NTM species varies geographically, but Mycobacterium avium complex (MAC) consistently accounts for 40–85% of pulmonary NTM infections worldwide [7–9]. Because of its high prevalence, chronic course, and limited treatment success, MAC-PD represents the most clinically significant form of NTM-PD. Despite advances in diagnostic microbiology, distinguishing MAC-PD from asymptomatic pulmonary colonization (MAC-PC) remains a persistent clinical dilemma [10–12]. MAC-PC refers to the presence of MAC in respiratory specimens without fulfilling the full ATS/IDSA diagnostic criteria for active disease [13]. Colonization is far more prevalent than true infection. Population-based data indicate that only approximately a quarter of individuals with MAC-positive respiratory cultures ultimately meet diagnostic criteria for active pulmonary disease [14]. Clinical differentiation between the two states is inherently difficult. Presenting symptoms such as chronic cough and sputum production are nonspecific, and radiographic patterns overlap considerably with underlying structural lung disease. Moreover, standardized microbiological criteria do not fully capture differences in bacterial burden or host response [14,15]. Misclassification, however, is not without clinical consequence, as mortality risk is significantly higher among patients meeting disease criteria than among those with colonization alone (hazard ratio 1.23–1.37) [7,8,14]. This diagnostic uncertainty complicates clinical decision-making, delaying appropriate therapy in some cases while leading to overtreatment in others. Recent immunologic studies have begun to elucidate host factors that underlie this clinical heterogeneity. Comparisons between patients with NTM-PD and NTM-PC have revealed that active disease is associated with impaired T cell–mediated immunity and decreased soluble form T-cell immunoglobulin mucin domain-3 (sTIM-3) [16,17]. Lower peripheral CD4+ T cell counts have also been linked to disease status, supporting the notion that dysregulated adaptive immunity contributes to MAC pathogenesis [18]. Beyond cell-surface or soluble markers, limited transcriptomic evidence further suggests enhanced proinflammatory and immune signalling among individuals who progressed to NTM-PD [19]. Collectively, these studies indicate that host immune imbalance plays a central role in distinguishing PD from PC, yet comprehensive blood transcriptomic profiling directly contrasting these two clinical states has not been systematically explored. In this study, we performed comprehensive transcriptomic profiling of patients with MAC infection across two independent cohorts. By integrating differential gene expression analysis, functional enrichment, and data-driven feature selection, we sought to identify host gene signatures that can discriminate MAC-PD from MAC-PC and further delineate molecular pathways associated with disease status. Materials and methods Ethics statement This study was prospectively conducted in accordance with the principles stated in the Declaration of Helsinki (https://www.wma.net/policies-post/wma-declaration-of-helsinki/). Approval was granted by the Institutional Review Boards of all participating hospitals (Approval numbers: 202301079RINC and KMUHIRB-F(I)−20240298). Written informed consent was obtained from all participants prior to enrolment. All experimental procedures involving human blood samples were performed in compliance with institutional laboratory biosafety regulations and standard operating procedures. No additional hazardous chemicals, live pathogens, or investigational products requiring special safety warnings were used in this study. Study cohort and clinical classification We recruited participants who were culture-positive for MAC in respiratory specimens from two medical centres in Taiwan between 2023 and 2025: Kaohsiung Medical University Hospital (KMUH; Cohort 1, training cohort) and National Taiwan University Hospital (NTUH; Cohort 2, validation cohort). MAC-PD or MAC-PC was determined according to the ATS/IDSA diagnostic criteria [13,20]. All participants received at least three independent respiratory samples for mycobacterial cultures. The final disease status was determined by consensus among three experienced pulmonologists (M.R.L., J.Y.W., and H.L.H.) following a comprehensive review of all clinical, microbiological, and radiographic data collected within six months after enrolment. In cases of initial disagreement, a structured discussion was held among the three pulmonologists until consensus was reached. The key determinants included the number and results of sputum cultures, acid-fast smear grading, and CT findings such as nodular lesions, bronchiectasis, and cavitation, all evaluated in the context of the full ATS/IDSA diagnostic criteria. While we acknowledge that differentiating between MAC-PD and MAC-PC is challenging, certain clinical scenarios create ambiguity. For instance, in patients with long-standing structural lung changes (such as bronchiectasis), they may satisfy both microbiological and radiological criteria; however, the absence of disease progression over time often leads to a diagnosis of MAC-PC rather than MAC-PD. Furthermore, the lack of clinical symptoms remains a definitive factor in excluding a diagnosis of MAC-PD. RNA-seq data generation Peripheral whole blood samples were collected at the clinical visit corresponding to MAC culture confirmation for both cohorts. Blood was collected into 10 mL sodium heparin tubes (BD Vacutainer, Cat. No. 367874). All samples were processed within 24 h of collection at room temperature. Briefly, whole blood was centrifuged to separate plasma, followed by density gradient centrifugation using Ficoll to isolate peripheral blood mononuclear cells (PBMCs). PBMCs were washed and the resulting cell pellet was immediately resuspended in TRIzol reagent (Thermo Fisher Scientific) and stored at −80°C until RNA extraction. Total RNA was extracted using the Qiagen RNeasy Mini Kit, and RNA integrity was assessed with the Qsep S100 system (BiOptic Inc., Taiwan), and detailed sample information is provided in Supplementary Table S1. RNA-seq libraries were constructed using the Illumina TruSeq Stranded mRNA Library Prep Kit, and sequencing was performed on an Illumina NovaSeq 6000 platform to generate paired-end 150 bp reads. RNA-seq data processing and differential expression analysis Raw FASTQ files were processed using Trimmomatic (v0.39) for adapter trimming and quality filtering [21]. Ribosomal RNA reads were removed using SortMeRNA (v4.3.4) [22]. Clean reads were aligned to the GRCh38 human reference genome using STAR aligner (v2.7.10a) with default parameters [23]. Gene-level expression was quantified using featureCounts (v2.0.1) [24]. Per-sample sequencing and alignment statistics are summarized in Supplementary Table S1. Differential expression analysis was performed using the edgeR package (v4.1.2) [25]. Genes with p-value < 0.05 and |log2 fold-change| > 1 were defined as differentially expressed genes (DEGs). Dimensionality reduction was then performed using PCA, t-SNE, and UMAP based on the top 2,000 most variable genes from log₂-transformed, Trimmed Mean of M-values (TMM)-normalized counts. To assess whether transcriptomic differences between groups were confounded by differences in immune cell composition, cell type deconvolution was performed on Cohort 1 using CIBERSORTx with the LM22 reference signature matrix [26]. Functional enrichment and driver term extraction Gene Ontology (GO) enrichment analysis was conducted using g:Profiler (web server: https://biit.cs.ut.ee/gprofiler) with default parameters [27]. The list of DEGs was used as input. The analysis was performed against the default Homo sapiens background gene set, and significance was determined using the g:SCS multiple testing correction method. GO terms from the biological process (BP), molecular function (MF), and cellular component (CC) categories were included. Terms with an adjusted p-value < 0.05 were considered significantly enriched. To support downstream feature selection, enriched GO terms were further curated to identify driver terms that represent key biological pathways or functions relevant. Genes associated with these driver terms were prioritized as candidate features for machine learning–based model development. Feature selection using two-stage strategy A two-stage feature selection strategy inspired by the maximum relevance and minimum redundancy (mRMR) framework was implemented [28]. In this approach, relevance and redundancy were evaluated separately rather than jointly to identify informative and non-overlapping gene candidates. This modification balances feature diversity and biological interpretability, as downstream classifiers such as random forest and SVM can inherently adjust for redundant features. In the first stage, candidate genes were identified using two complementary feature selectors (LASSO and RFE) each executed with four random seeds. LASSO was performed using the glmnet package in R to determine the optimal penalty parameter (λ). Genes with nonzero coefficients at the optimal λ were retained as candidate features. For RFE, the caret package was employed, using a random forest classifier as the base model. Features were recursively ranked according to their contribution to classification performance, and subsets of features were evaluated iteratively. Genes appearing in at least one run were collected as the initial candidate pool. In the second stage, candidate genes were filtered by assessing relevance and functional redundancy independently. For maximum relevance, genes showing log2 fold-change > 3 or false discovery rate (FDR) < 0.05 were retained. The log₂FC > 3 criterion was applied to retain genes with biologically meaningful expression differences, as more modest fold-changes in whole-blood transcriptomics are particularly susceptible to confounding by inter-individual variability in cellular composition. The FDR < 0.05 threshold reflects standard practice in differential expression analysis. For minimum redundancy, genes were required to possess unique GO annotations without functional overlap with other selected genes, thereby minimize functional redundancy among selected features in terms of biological function rather than expression patterns [29,30]. Additionally, genes consistently selected across all runs by either LASSO or RFE were also included. The final gene set was then used for model training. Model training and evaluation The systematically combinations of the final gene set in subsets of up to seven genes (per the N/10 rule [31,32]) were used to train machine learning models for classifying disease versus colonization status. Two algorithms were employed: random forest (RF) and support vector machine (SVM) [33,34]. Model training and evaluation were implemented using the caret package in R, and all models were trained using the default hyperparameter settings. Model performance was primarily evaluated by the area under the receiver operating characteristic curve (AUC), with additional metrics including accuracy, sensitivity, specificity, precision, and F1 score. For clinical interpretability, optimal classification thresholds were determined by maximizing Youden’s J statistic [35]. Confusion matrices were further generated to visualize classification outcomes. External validation was performed using an independent patient cohort from NTUH (Cohort 2). Models trained on Cohort 1 were applied to this dataset without retraining, and predictive performance was assessed using the same evaluation metrics. Results Baseline characteristics of study participants A total of 120 MAC-positive patients were enrolled from two tertiary hospitals, including 74 patients from Cohort 1 and 46 patients from Cohort 2 (Figure 1). All RNA samples achieved transcriptome alignment rates exceeding 50% (range: [e.g. 54.2% to 76.0%]). RNA integrity (RQN) and purity (260/280 ratios) for each sample are summarized in Supplementary Table S1. Among all participants, 73 (60.8%) were classified as MAC-PD and 47 (39.2%) as MAC-PC (Table 1). Analysis workflow. Table 1.Clinical characteristics of 120 participants. All(n = 120)MAC-PD(n = 73)MAC-PC(n = 47)pCohort 1(n = 74)Cohort 2(n = 46)pAge68.2 ± 10.766.9 ± 10.770.3 ± 10.40.08768.1 ± 11.968.3 ± 8.40.927 < 453 (2.5)2 (2.7)1 (2.1) 3 (4.1)0 (0.0) 45–6539 (32.5)26 (35.6)13 (27.7) 23 (31.1)16 (34.8) > 6578 (65.0)45 (61.6)33 (70.2) 48 (64.9)30 (65.2) Body mass index20.8 ± 3.720.1 ± 3.621.9 ± 3.80.01221.2 ± 4.220.3 ± 2.80.213 < 1827 (22.5)21 (28.8)6 (12.8) 17 (23.0)10 (21.7) 18–2578 (65.0)45 (61.6)33 (70.2) 44 (59.5)34 (73.9) > 2515 (12.5)7 (9.6)8 (17.0) 13 (17.6)2 (4.3) Male52 (43.3)24 (32.9)28 (59.6)0.00435 (47.3)17 (37.0)0.266Cohort 0.251 Cohort 174 (61.7)48 (65.8)26 (55.3) Cohort 246 (38.3)25 (34.2)21 (44.7) MAC-PD 48 (64.9)25 (54.3)0.251MAC-PC 26 (35.1)21 (45.7) AFS positive53 (44.2)42 (57.5)11 (23.4)<0.00127 (36.5)26 (56.5)0.032Active or ever smoker31 (25.8)15 (20.5)16 (34.0)0.09921 (28.4)10 (21.7)0.360Underlying conditions Hypertension29 (24.2)15 (20.5)14 (29.8)0.24917 (23.0)12 (26.1)0.779 Diabetes mellitus14 (11.7)10 (13.7)4 (8.5)0.3887 (9.5)7 (15.2)0.377 Malignancy17 (14.2)8 (11.0)9 (19.1)0.20910 (13.5)7 (15.2)0.855 Asthma15 (12.5)6 (8.2)9 (19.1)0.07711 (14.9)4 (8.7)0.289 COPD15 (12.5)7 (9.6)8 (17.0)0.23014 (18.9)1 (2.2)0.006 Autoimmune disease11 (9.2)7 (9.6)4 (8.5)0.8413 (4.1)8 (17.4)0.017 Prior tuberculosis15 (12.5)12 (16.4)3 (6.4)0.10411 (14.9)4 (8.7)0.289Chest CT findings Bronchiectasis92 (76.7)59 (80.8)33 (70.2)0.18053 (71.6)39 (84.8)0.190 Nodules75 (62.5)50 (68.5)25 (53.2)0.09139 (52.7)36 (78.3)0.010 Cavitation23 (19.2)20 (27.4)3 (6.4)0.00411 (14.9)12 (26.1)0.155 Consolidation13 (10.8)11 (15.1)2 (4.3)0.0639 (12.2)4 (8.7)0.511 Pleural effusion4 (3.3)2 (2.7)2 (4.3)0.6524 (5.4)0 (0.0)0.103AFS, acid-fast smear; COPD, chronic obstructive pulmonary disease; CT, computed tomography; MAC, Mycobacterium avium complex; p, p value; PC, pulmonary colonization; PD, pulmonary disease.Data are either mean ± standard deviation or number (percentage). p values were calculated by Student’s t-test for continuous variables and chi-square test for categorical variables. Clinical characteristics of 120 participants. AFS, acid-fast smear; COPD, chronic obstructive pulmonary disease; CT, computed tomography; MAC, Mycobacterium avium complex; p, p value; PC, pulmonary colonization; PD, pulmonary disease. Data are either mean ± standard deviation or number (percentage). p values were calculated by Student’s t-test for continuous variables and chi-square test for categorical variables. Patients with MAC-PD had significantly lower BMI compared to those with MAC-PC (20.1 ± 3.6 vs. 21.9 ± 3.8 kg/m², p = 0.012) and were more likely to be female (67.1% vs. 40.4%, p = 0.004). The disease group also demonstrated a higher rate of AFS positivity (57.5% vs. 23.4%, p < 0.001) and more frequent cavitation on computed tomography (CT) of chest (27.4% vs. 6.4%, p = 0.004). No significant difference was observed in active or ever smoking status between the two groups. Comparing the two cohorts, Cohort 2 showed a higher proportion of AFS-positive cases (56.5% vs. 36.5%, p = 0.032) and a higher prevalence of pulmonary nodules (78.3% vs. 52.7%, p = 0.011). Regarding underlying comorbidities, Cohort 1 had a higher prevalence of chronic obstructive pulmonary disease (COPD) (18.9% vs. 2.2%, p = 0.006), while Cohort 2 had a higher prevalence of autoimmune diseases (17.4% vs. 4.1%, p = 0.017). The workflow illustrates key steps including patient enrolment, transcriptomic profiling, feature selection, machine learning model development, and external validation. Transcriptomic signatures distinguish MAC pulmonary disease from colonization Peripheral blood transcriptomes from 48 MAC-PD and 26 MAC-PC patients in Cohort 1 were compared. A total of 147 DEGs were detected, including 29 genes meeting FDR < 0.05 (Supplementary Figure S1A, Supplementary Table S2). Functional enrichment analysis of the 147 DEGs highlighted immune-related processes. Seven driver Gene Ontology (GO) terms were significantly enriched: antigen binding (MF), adaptive immune response (BP), immunoglobulin complex, extracellular region, cell periphery, specific granule, and tertiary granule (CC) (Supplementary Figure S1B and S1C, Supplementary Table S3). Unsupervised sample clustering heatmap and dimensionality reduction PCA, t-SNE, UMAP results are provided in Supplementary Figure S2. To evaluate whether these transcriptomic differences were confounded by differences in immune cell composition, cell type deconvolution was performed using CIBERSORTx [26]. No statistically significant differences in estimated immune cell fractions were observed between MAC-PD and MAC-PC patients across all 22 cell types (all Benjamini-Hochberg-corrected p > 0.7; Supplementary Figure S3 and Supplementary Table S3), suggesting that the identified transcriptomic signatures are not primarily driven by differences in cellular composition. Feature selection identifies robust transcriptomic biomarkers From the 83 genes annotated under the seven driver GO terms (Supplementary Table S4), 31 were retained after feature-selection analyses aimed at distinguishing MAC-PD from MAC-PC. Refinement based on pre-specified statistical and functional criteria yielded a final panel of 18 genes (Table 2), which was used in subsequent modelling. Table 2.Genes selected by LASSO and RFE across multiple random seeds.Gene IDGene nameLASSO_nRFE_n|log2FC|FDRGO termCriterion fulfilledaENSG00000166444DENND2B441.5930.009cell periphery1,2ENSG00000166035LIPC431.3880.155extracellular region1ENSG00000242371IGKV1-39041.5690.155antigen binding2ENSG00000251546IGKV1D-39041.7420.155antigen binding2ENSG00000169439SDC2302.1750.007cell periphery4ENSG00000187950OVCH1401.3890.435extracellular region1ENSG00000115414FN1421.1800.036extracellular region1,4ENSG00000122861PLAU011.6520.717tertiary granule5ENSG00000198947DMD023.153<0.001cell periphery3,4ENSG00000275743TRBV14211.2310.284adaptive immune response5ENSG00000211653IGLV1-40421.0210.404antigen binding1ENSG00000211959IGHV4-39041.1640.469antigen binding2ENSG00000211611IGKV6-21401.4910.155immunoglobulin complex1ENSG00000143869GDF7011.936<0.001extracellular region4ENSG00000165810BTNL9012.769<0.001cell periphery4ENSG00000211955IGHV3-33401.0370.462antigen binding1ENSG00000144724PTPRG401.4130.119extracellular region1ENSG00000102174PHEX011.5880.002cell periphery4a, five selected criteria: 1, consistently selected by LASSO; 2, consistently selected by RFE; 3, |log2 fold-change| > 3; 4, FDR < 0.05; 5, possessed unique GO annotations without overlapping functional annotations with other candidates Genes selected by LASSO and RFE across multiple random seeds. a, five selected criteria: 1, consistently selected by LASSO; 2, consistently selected by RFE; 3, |log2 fold-change| > 3; 4, FDR < 0.05; 5, possessed unique GO annotations without overlapping functional annotations with other candidates Model development and cross-cohort validation Models were constructed using combinations of the 18 selected genes to classify MAC-PD versus MAC-PC and evaluated by internal resampling and external validation. Thirteen candidate models achieved AUCs above 0.70 in external validation, demonstrating consistent classification of MAC disease versus colonization. All of these models were generated using SVM classifiers, whereas none of the RF models reached this performance threshold (Figure 2A–C, Table 3, Supplementary Table S5). Performance evaluation of machine learning models (a) Scatter plot of model performance showing training versus validation area under the receiver-operating characteristics curve (AUC). Red points indicate models with validation AUC > 0.7. (b) Combined plot of models with validation AUC > 0.7. The x-axis represents individual models; the left y-axis shows training and validation AUCs (bar plots), and the right y-axis shows the difference between training and validation AUC (line plot). A horizontal dashed line indicates a difference of 0.2 between training and validation AUC. (c) Receiver operating characteristic (ROC) curves of the top six models (validation AUC > 0.7 and the difference between training and validation AUC < 0.2). (d) Confusion matrices of the three models with the highest validation AUCs. Table 3.Top-performing models.Set nameModelTrain AUCValidation AUCdAUCAccuracyPrecisionSensitivitySpecificityF1 ScoreGene setf23516SVM0.910.780.130.760.750.840.670.79IGKV1-39, IGKV1D-39, OVCH1, PLAU, DMD, IGKV6-21f9957SVM0.920.780.150.740.810.680.810.74IGKV1D-39, OVCH1, PLAU, DMD, IGKV6-21f27077SVM0.940.750.190.740.760.760.710.76IGKV1D-39, OVCH1, PLAU, DMD, IGKV6-21, BTNL9f2412SVM0.920.740.180.70.870.520.90.65IGKV1-39, OVCH1, PLAU, IGKV6-21f52466SVM0.910.740.170.720.930.520.950.67IGKV1-39, IGKV1D-39, OVCH1, PLAU, DMD, IGKV6-21, BTNL9f3969SVM0.870.70.170.6710.410.57TRBV14, GDF7, BTNL9, PTPRGf27080SVM0.950.750.210.740.740.80.670.77IGKV1D-39, OVCH1, PLAU, DMD, IGKV6-21, PHEXf52469SVM0.960.740.220.720.880.560.90.68IGKV1-39, IGKV1D-39, OVCH1, PLAU, DMD, IGKV6-21, PHEXf8984SVM0.930.720.20.70.820.560.860.67IGKV1-39, OVCH1, PLAU, IGKV6-21, BTNL9f2785SVM0.920.720.210.720.730.760.670.75IGKV1D-39, OVCH1, DMD, IGKV6-21f2421SVM0.90.710.20.70.870.520.90.65IGKV1-39, OVCH1, DMD, IGKV6-21f25075SVM0.920.710.210.720.830.60.860.7IGKV1-39, OVCH1, PLAU, DMD, IGKV6-21, BTNL9f55253SVM0.950.70.250.70.760.640.760.7IGKV1-39, OVCH1, PLAU, DMD, IGKV6-21, BTNL9, PHEXdAUC, difference between Train AUC and Validation AUC; SVM, support vector machine; bold font, highly conserved five-gene core signature; above the dashed line indicates dAUC < 0.2. Bold text indicates the five-gene core signature. Top-performing models. dAUC, difference between Train AUC and Validation AUC; SVM, support vector machine; bold font, highly conserved five-gene core signature; above the dashed line indicates dAUC < 0.2. Bold text indicates the five-gene core signature. Six of these models showed training–validation AUC gaps below 0.20, indicating stable generalizability. Across these six models, validation accuracies ranged from 0.67 to 0.76, precision from 0.75 to 1.00, sensitivity from 0.40 to 0.84, and specificity from 0.67 to 1.00. This variation reflects the expected trade-off between sensitivity and specificity across models. Three models achieved the highest AUCs of 0.78, 0.78, 0.75, with confusion matrices indicating balanced discrimination between disease and colonization (Figure 2D). Their corresponding accuracies were 0.76, 0.74, and 0.74; precision values were 0.75, 0.81, and 0.76; F1 scores were 0.79, 0.74, and 0.76. Detailed metrics are summarized in Table 3. Comparison of clinical, transcriptomic, and combined models To evaluate whether incorporating clinical variables improves classification, we compared three models using the same modelling framework: a clinical-only model using five clinical features (age, BMI, sex, AFS positivity, and cavitation), the transcriptomic five-gene core signature, and a combined model integrating both. The clinical-only model achieved a median validation AUC of 0.58, whereas the five-gene transcriptomic model achieved 0.77. The combined model yielded a median AUC of 0.66, indicating that the addition of clinical variables did not improve discriminative performance (Supplementary Table S6). Gene composition of top-performing models Examination of the top three models (AUC = 0.78, 0.78, and 0.75) revealed a highly conserved five-gene core signature: IGKV1D-39, IGKV6-21, OVCH1, PLAU, and DMD (Table 3). The first-ranked model additionally included IGKV1-39, while the third incorporated BTNL9. Among these five genes, IGKV1D-39, IGKV6-21, OVCH1, and PLAU were upregulated in the disease group (log₂FC = 1.74, 1.49, 1.39, 1.65, respectively), whereas DMD was downregulated (log₂FC = −3.15) (Figure 3 and Supplementary Figure S4). Across all 13 top-performing models, OVCH1 and IGKV6-21 appeared most frequently (92.3%), followed by PLAU (76.9%), DMD (76.9%), IGKV1-39 (61.5%), and IGKV1D-39 (53.8%). (Table 3). Seven genes from the initial 18-gene set were absent from all top-performing models, while those containing PHEX consistently exhibited large training–validation AUC gaps (>0.20) (Table 3). Therefore, the subsequent analysis focused on the remaining ten genes that were recurrently selected across stable models. These genes delineate two interconnected biological axes, dysregulated adaptive immunity and impaired tissue remodelling, that collectively distinguish MAC-PD from MAC-PC (Figure 3), as further elaborated in the following Discussion section. Summary of key molecular signatures associated with MAC-PD. The schematic illustrates two key functional alterations in MAC-PD: Dysregulated adaptive immunity (left) and Impaired tissue remodeling and repair (right). Gene names are followed by their log₂ fold-change values. Genes belonging to the five-gene core signature (IGKV1D-39, IGKV6-21, OVCH1, PLAU, DMD) are shown in bold. Upward red arrows indicate genes with increased expression, while downward blue arrows indicate decreased expression. Discussion In this study, we systematically explored peripheral blood transcriptomic profiles to distinguish MAC-PD from MAC-PC. Using a two-stage feature selection strategy combining LASSO regression and RFE, followed by statistical and functional filtering, we identified a biologically informed panel of 18 candidate genes. The top three models achieved validation AUCs of 0.78, 0.78 and 0.75 and shared a recurrent five-gene core signature observed in our cohorts comprising IGKV1D-39, IGKV6-21, OVCH1, PLAU, and DMD. Although diagnostic criteria for MAC-PD have been defined, the clinical differentiation remains challenging [13,20]. In practice, the diagnosis often requires consensus among experienced experts [13], and even then, subjective interpretation and interobserver bias are unavoidable [36]. Recent efforts have investigated various biological dimensions to aid this distinction. Breathomic profiling using an electronic nose (AUC = 0.68 to 0.74) captures metabolic alterations associated with airway infection [37], whereas serum sTIM-3 measurement (AUC = 0.68) reflects host immune activation [17]. Our PBMC transcriptome–based model achieved slightly higher discriminative performance (AUC up to 0.78) by characterizing transcriptional responses. Notably, a clinical-variable-only model constructed from features significantly different between MAC-PD and MAC-PC (age, BMI, sex, AFS positivity, and cavitation) achieved substantially lower predictive performance (median AUC 0.58), and a combined model (median AUC 0.66) did not outperform the transcriptomic model alone, suggesting that the five clinical features do not collectively provide sufficient discriminative information, and their addition may introduce noise that dilutes the transcriptomic signal. Each of these approaches captures a distinct but complementary dimension of the host-pathogen interaction. How these biological signals might be most effectively combined, and whether their integration would yield clinically meaningful gains over any single modality, remains an open question that future studies will need to address. Applications of blood transcriptomics in infectious diseases and opportunities for MAC diagnosis Blood-based transcriptomic profiling has emerged as a powerful tool for investigating host immune responses to infections. In bacterial and viral diseases, it enables systemic and minimally invasive characterization of host–pathogen interactions, revealing dynamic changes across innate and adaptive immune pathways. Transcriptomic signatures have been successfully applied to distinguish active infection from latent or resolved states, predict disease severity, and monitor treatment responses in conditions such as influenza, sepsis, and COVID-19 [38,39]. These advances underscore the potential of peripheral blood mRNA sequencing for identifying diagnostic and prognostic biomarkers across infectious diseases. In tuberculosis (TB), extensive transcriptomic studies have identified consistent perturbations in interferon signalling, neutrophil activation, and antigen-presentation pathways, reflecting robust innate immune activation [38,40–42]. However, adaptive immune receptor transcripts, including immunoglobulin and T cell receptor variable region genes, are less frequently detected, likely due to their lower abundance and greater individual variability. These findings have established a conceptual framework for understanding the systemic immune landscape of mycobacterial infections and provide a benchmark for comparison with other NTM diseases. In contrast, blood-based transcriptomic biomarkers for distinguishing MAC-PD from MAC-PC remain largely unexplored. Conventional diagnostic criteria rely on microbiologic and radiologic findings, which capture only selected aspects of disease manifestation and may miss early or low-burden cases [13]. Peripheral blood transcriptomics thus offers an opportunity to complement current diagnostic approaches by integrating systemic immune responses and potentially identifying molecular signatures that differentiate true infection from colonization. Dysregulated adaptive immunity and tissue remodelling distinguish MAC-PD from MAC-PC The molecular signatures identified in MAC-PD consistently highlight a dual pathophysiological axis: robust adaptive immune activation coupled with dysregulated tissue remodelling and impaired repair. On the immune side, the enrichment of immunoglobulin variable genes (IGKV1D-39, IGKV1-39, IGKV6-21) reflects B cell–mediated humoral activation, likely driven by chronic antigenic stimulation inherent to persistent MAC infection [43–46]. B cells not only produce antibodies but also coordinate T cell responses, present antigens, and secrete cytokines [47,48], suggesting that ongoing humoral activity may directly influence cellular immunity. TRBV14 upregulation indicates heightened antigen-specific T cell engagement, while concurrent downregulation of BTNL9 and PTPRG, molecules that normally restrain excessive T cell activation [49,50], points to impaired immune regulatory control. This pattern of heightened adaptive activation in the absence of adequate regulatory signals implies a systemic imbalance that may sustain inflammation and contribute to tissue damage. These findings suggest that MAC disease is not merely characterized by immune activation, but by a failure to fine-tune adaptive responses, which may differentiate active disease from colonization and reveal potential targets for immunomodulatory interventions. Parallel to immune dysregulation, genes associated with tissue remodelling and repair reveal that structural compromise accompanies persistent inflammation. PLAU, encoding urokinase-type plasminogen activator, mediates extracellular matrix degradation and fibrinolytic activity, processes crucial for both protective remodelling and fibrosis [51–53]. Dysregulation of plasminogen activation has been causally linked to excessive matrix deposition and structural lung injury [54,55], indicating that chronic infection may tilt the balance toward maladaptive remodelling. OVCH1, a serine protease implicated in matrix turnover [56], reinforces the concept that extracellular matrix dynamics are closely intertwined with immune activation. Consistent downregulation of DMD, encoding dystrophin, suggests cytoskeletal instability and structural vulnerability, potentially paralleling the lung damage observed in MAC-PD. Reduced expression of GDF7, a TGF-β superfamily member involved in anti-inflammatory and tissue repair signalling [57], further indicates impaired reparative capacity. Collectively, these genes define a composite signature linking immune activation to structural damage and compromised repair, highlighting a mechanistic axis by which chronic MAC infection drives persistent pathology. This integrated view suggests that active MAC-PD disease arises from the intersection of uncontrolled adaptive immune activation and maladaptive tissue remodelling, with insufficient regulatory and reparative signalling amplifying tissue injury. The molecular interplay observed here not only differentiates disease from colonization but also provides a framework for understanding heterogeneity in MAC pathogenesis and for identifying novel therapeutic targets that could modulate immune responses or restore tissue integrity (Figure 3). Limitations and future directions Several limitations of this study should be acknowledged. First, the cross-sectional design means that the identified signature reflects disease state at the time of blood sampling and should not be interpreted as a predictive, prognostic, or theranostic tool. Whether this signature can anticipate disease onset prior to clinical manifestation, change in response to antimycobacterial therapy, or correlate with long-term outcomes remains unknown. Future prospective studies with longitudinal sampling designs, ideally beginning at the first positive culture, are needed to address these questions. Furthermore, the interval between blood collection and culture-positivity was not uniform. To effectively differentiate MAC-PD from MAC-PC in phenotypically similar patients, transcriptomic profiling should ideally be performed at the first or second positive culture. Sampling later in the disease course may coincide with clinically evident progression, thereby diminishing the added value of transcriptomic analysis. However, this time window was likely constrained in this study, as participants were usually followed in clinic at monthly intervals, and recruitment and blood sampling were conducted immediately upon confirmation of culture-positivity. Second, for many of the signature genes identified, including IGKV1D-39, OVCH1, PHEX, GDF7, and DMD, functional validation in lung tissue or in vitro infection models was not performed. In addition, immune repertoire genes are characterized by high polymorphism and variability in expression, which raises the possibility of technical noise or sampling bias. Confirming reproducibility across different populations and sample preparation protocols will therefore be an important next step. Third, remodelling- and protease-related genes may capture aspects of structural lung damage, including bronchiectasis and fibrosis, rather than reflecting active infection alone. Because patients classified as MAC-PD under ATS/IDSA criteria are by definition more likely to exhibit advanced structural lung abnormalities, the transcriptomic contributions of active infection and pre-existing structural damage cannot be rigorously disentangled in our current design. A future study incorporating a diseased control group with structural lung disease, or paired quantitative imaging data, would help clarify this distinction. Fourth, information on corticosteroid and immunosuppressant use was not systematically captured in our dataset. Given that autoimmune disease prevalence was relatively low in Cohort 1 (4.1%) and higher in Cohort 2 (17.4%), as shown in Table 1, the potential confounding effect of immunosuppressive therapy on whole-blood transcriptomic profiles cannot be fully excluded. Future validation studies should prospectively document immunosuppressive medication use as a covariate to better characterize its influence on the identified signature. Lastly, although our study represents the largest cohort reported to date, the sample size remains relatively modest, which may limit the statistical power and generalizability of our findings. Conclusion Taken together, our integrative approach utilizing an independent validation cohort identified compact gene signatures capable of discriminate MAC disease from colonization. Importantly, these signatures comprised small panels enriched in immune- and matrix-related genes, supporting their potential as candidates for future clinical investigation. Our findings may shed light on future studies investigating molecular mechanisms underlying MAC disease progression and host susceptibility. Supplementary Material Supplementary Table S6.docx Figure S4.tif Supplementary Table S2.xlsx Supplementary Table S4.xlsx Figure S2.tif Figure S3.tif Supplementary Table S1.xlsx Supplementary Table S5.xlsx Supplementary Table S3.docx Figure S1.tif Acknowledgements We would like to thank all participants for their involvement in this study. We are also grateful to the Data Science Statistical Cooperation Center of Academia Sinica (AS-CFII-111-215) for providing statistical support. In addition, we thank the staff of the Eighth Core Laboratory, Department of Medical Research, National Taiwan University Hospital, for their technical assistance throughout the study. Part of the results of this study was previously presented as a poster at the TSPCCM 2025 Congress. Funding Statement This work was supported by the Taiwan Ministry of Science and Technology under Grant NSTC 114-2314-B-002-265-MY3, NSTC 114-2314-B-002-282, NSTC 113-2628-B-002-017-MY3, 112-2314-B-002-127; and Kaohsiung Medical University Hospital under Grant KMU-TB114001, KMUH-DK(C)114006, KMUH-11406T. The funders had no role in the study design, data analysis, and manuscript writing. Disclosure statement No potential conflict of interest was reported by the author(s). Declaration of generative AI and AI-assisted technologies in the manuscript preparation process During the preparation of this work the authors used ChatGPT (OpenAI) for English language editing and grammar correction. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. Data sharing statement Data were available upon reasonable request from the corresponding author. Supplemental Material Supplemental data for this article can be accessed online at https://doi.org/10.1080/22221751.2026.2678657. References Associated Data Supplementary Materials Supplementary Table S6.docx Figure S4.tif Supplementary Table S2.xlsx Supplementary Table S4.xlsx Figure S2.tif Figure S3.tif Supplementary Table S1.xlsx Supplementary Table S5.xlsx Supplementary Table S3.docx Figure S1.tif