PTBD: a machine learning-based non-invasive diagnostic model for pulmonary tuberculosis using large-scale blood transcriptomes

Abstract

Background

Accurate and rapid diagnosis is essential for controlling pulmonary tuberculosis (PTB) by enabling timely intervention and reducing disease transmission. Existing diagnostic methods for PTB are limited by sensitivity, specificity, or practicality.

Results

Here, we present a non-invasive blood transcriptome-based diagnostic model that integrates top-scoring pair with machine learning, identifying novel and robust transcriptomic biomarkers for diverse PTB patients. Using 2,792 peripheral blood transcriptome samples, the proposed model (PTBD) effectively distinguishes PTB from healthy individuals, latent tuberculosis infection, pneumonia, lung cancer, and pulmonary nodules, reflecting its robustness against the complex and heterogeneous negative sample background typical of real-world clinical settings, and achieving AUCs of 0.869 in the test set and 0.909 in an independent external validation set. Its performance is consistent across different geographic regions, age groups, and special conditions, including Bacillus Calmette-Guérin vaccination, HIV infection, diabetes, and drug resistance, meeting WHO requirements for community-based triage, children, and PTB with HIV infection. Furthermore, PTBD also enables diagnosis of extrapulmonary tuberculosis and prediction of treatment outcomes, with feature scores serving as molecular biomarkers reflecting disease progression and prognosis.

Conclusions

This study provides a broadly applicable tool for early PTB diagnosis, facilitating timely intervention and potentially reducing global PTB burden. Moreover, PTBD uncovers novel transcriptomic biomarkers, represented by five diagnostic gene-pair expression patterns that embody the molecular hallmarks of PTB.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12915-026-02585-9.

Background

Tuberculosis (TB) is an infectious disease caused by Mycobacterium tuberculosis (MTB), with over 10 million new infections reported annually. It remains one of the leading causes of infectious disease-related deaths worldwide, with fatalities nearly double those caused by HIV. Pulmonary tuberculosis (PTB), which accounts for approximately 80% of adult TB cases, represents a major global health challenge [1–3]. PTB symptoms typically include cough, expectoration, and chest pain; if left untreated, it can lead to severe lung damage and death [4]. The treatment of PTB is further complicated by MTB variability and the emergence of drug resistance. Moreover, the COVID-19 pandemic has exacerbated the global burden of PTB [5].

Accurate and rapid diagnosis is essential for controlling PTB. However, current diagnostic methods have notable limitations. The tuberculin skin test (TST) lacks specificity [6]; while chest X-rays and CT scans are insufficiently accurate and may pose risks to the body [6]. Sputum smear microscopy has low sensitivity, missing a significant proportion of cases [7]. Although MTB culture is considered the gold standard, it is slow and may take several weeks to yield results [8]. Emerging methods like interferon-gamma release assays (IGRA) and Xpert MTB/RIF assays offer promise but also have limitations. The IGRA cannot distinguish between PTB and latent tuberculosis infection (LTBI) and may produce false results in immunocompromised individuals [9]. The Xpert MTB/RIF assay improves diagnostic capabilities but has reduced accuracy in HIV-infected individuals and children, and its application is limited by challenges in obtaining induced sputum from children and from adults after symptom improvement [10, 11]. Despite available treatments, misdiagnoses and delays in PTB detection remain prevalent, contributing to high morbidity and mortality. Thus, there is an urgent need for novel, rapid, accurate, and non-invasive PTB diagnostic methods [12]. Blood transcriptome biomarkers offer a potential solution, as blood is easier to collect and more stable than sputum [13, 14].

Studies have used peripheral blood or peripheral blood mononuclear cell (PBMC) transcriptome data to identify PTB diagnostic genes [15–20]. Host peripheral blood consists of various immune cell types, including neutrophils, monocytes, CD4 T cells, and CD8 T cells. Distinct transcriptional alterations within these immune cell subpopulations may occur during MTB infection and disease progression, reflecting the dynamic host immune response and enabling the capture of potential PTB progression. For example, Berry et al. [21] reported that the blood transcriptional signature driving PTB is predominantly derived from neutrophil-associated pathways. These findings highlight the importance of considering immune cell heterogeneity in TB transcriptomic studies and suggest that characterizing key immune cell types and their molecular features may enhance the biological interpretability of diagnostic models. However, these studies have several notable limitations. Most relied on limited types of controls that do not reflect real-world clinical settings, and focused exclusively on adults. Many were conducted in a single country, limiting generalizability. Furthermore, few accounted for critical factors such as Bacillus Calmette-Guérin (BCG) vaccination, PTB with HIV coinfection, PTB with diabetes, or drug resistance. Existing models often did not compare performance against other studies or the World Health Organization’s Target Product Profiles (WHO-TPPs) [22], included large numbers of feature genes that complicate practical use, and depended on absolute gene expression values that require batch effect adjustment for clinical implementation.

To overcome these challenges, we developed a diagnostic model for PTB that extracts key gene pairs from peripheral blood transcriptome data using gene expression rank-order relationships combined with machine learning. This model reflects real-world conditions [23], performs robustly across different age groups and countries, and is insensitive to special conditions such as BCG vaccination, HIV infection, diabetes, or drug resistance. It outperforms previous models and meets parts of the WHO-TPPs performance criteria, while requiring only a small set of genes. By leveraging rank-order gene expression rather than absolute values, the model avoids the need for batch effect correction. Notably, it is also applicable for diagnosing extrapulmonary tuberculosis (EPTB) and predicting PTB treatment efficacy. We further provide an online web server that enables rapid, accurate, and non-invasive identification of PTB patients, supporting early intervention, improved treatment outcomes, reduced transmission, and alleviation of the global TB burden [24].

Results

A novel transcriptome-driven non-invasive model for improving PTB diagnosis

To construct a robust diagnostic model for PTB, we collected and curated 33 transcriptome datasets comprising 2,792 samples across diverse populations, geographic regions, and clinical conditions (Fig. 1A, Additional File 2: Table S1). The framework for model construction is illustrated in Fig. 1B. The discovery dataset comprised 15 datasets with a total of 855 samples from 10 countries, including 4 high-burden TB countries [1]. Among these, 415 were confirmed PTB cases, and 440 served as negative controls. Following a 7:3 ratio, the various types of data in the discovery data are further divided into a training set (596 samples) and a testing set (259 samples) (Fig. 2A-B).

Fig. 1.

Overview of the data and the study design. A Dataset inclusion summary. This study utilized a total of 33 datasets comprising 2,792 samples. The discovery data consisted of 15 datasets with 855 samples, while the external validation data included 10 datasets with 414 samples. Some datasets were used for multiple analyses. Additional data types including extrapulmonary tuberculosis, disease progression, and prognosis were grouped as “Other”. The presence of each category in a dataset is indicated by ‘Yes’ or ‘No’. A stacked bar chart visualizes the sample sizes within each dataset. B Overall study design. The process includes model construction, model validation, core feature analysis, extension of biomarker applications to prognosis and disease progression, and development of an online web server. BCG: Bacillus Calmette-Guérin; PTB: pulmonary tuberculosis; LTBI: latent tuberculosis infection; HC: healthy control; PNA: pneumonia; LC: lung cancer; PN: pulmonary nodules; EPTB: extrapulmonary tuberculosis; TSP: Top-Scoring Pair; CTRL: control

Fig. 2.

Feature selection and model selection. A Distribution of samples in the training set. B Distribution of samples in the test set. C Feature selection results, with the x-axis representing the number of gene pairs and the y-axis representing the AUC. D Performance of feature importance rankings by the mRMR method across different algorithms, with the x-axis showing algorithms and the y-axis showing the number of features used. AUC values exceeding 0.920 are highlighted in yellow. E Barcode plots illustrate the relative expression levels of top 5 gene pair in the training set. F Performance comparison of six models constructed using the final feature set. MIC: maximal information coefficient; Chi2: chi-squared test; mRMR: minimum redundancy maximum relevance; RF: random forest; RD: reversal degree; Avg: average weight; LR: logistic regression; SVM: support vector machine

In the training set, gene pair scores were calculated using the Top-Scoring Pair (TSP) method. Applying a threshold of 0.40 identified 454 gene pairs (Additional File 2: Table S2). These gene pairs were then used to encode the training set, and their importance scores were calculated using six feature importance scoring methods (Additional File 2: Table S3). To ensure practicality, only the top ten features from each ranking method were retained for incremental feature selection (IFS). The results showed that the Minimum Redundancy Maximum Relevance (mRMR) method consistently achieved the highest AUC across all models (Fig. 2C). Notably, even with only five gene pairs, the AUC exceeded 0.925, with minimal further improvement upon inclusion of additional features (Fig. 2D). Similar trends were also observed for the area under the precision-recall curve (AUC-PR) (Additional file 1: Fig. S1).

Based on the mRMR ranking, the top five gene pairs, each exhibiting consistent expression inversion between PTB and control groups, were selected as the final features (Fig. 2E). Using these features, multiple models were constructed, including eXtreme Gradient Boosting (XGBoost), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), soft voting, and stacking. Among them, the LR-based model achieved the highest combined AUC and AUC-PR scores (Fig. 2F: LR: 1.8539; SVM: 1.8537; soft voting: 1.8528; stacking: 1.8516; XGBoost: 1.8447; RF: 1.8398), and was thus selected as the optimal model, referred to as PTBD. The detailed parameters and configuration of the PTBD (LR-based) model are provided in Additional File 2: Table S4.

Comprehensive validation of PTBD across diverse populations, clinical conditions, and geographic settings

We first validated the model on the test set, which comprised 259 samples (Fig. 2B). The model performed well across comparisons between PTB and different control categories, with AUC values exceeding 0.850 in most cases (Fig. 3A). When analyzed by country, the average AUC was 0.906 in high-burden countries (Fig. 3B) and 0.857 in other countries (Fig. 3C).

Fig. 3.

Validation of model performance on the test and external validation sets. A Model performance across different categories in the test set. B–C Model performance across different countries in the test set. D Distribution of samples in the external validation set. E–H Model performance in the external validation set, evaluated by different categories and countries

External validation was conducted using 414 samples from 8 countries, including 3 high-burden TB countries (Fig. 3D). Among these, 131 were PTB cases and the remainder were controls. In category-specific analyses, the AUC for distinguishing PTB from pneumonia (PNA) was lower at 0.786, while all other comparisons yielded AUC values above 0.847. Notably, the AUC for distinguishing PTB from all controls was 0.909 (Fig. 3E). Country-specific validation showed that all high-burden countries, except China (AUC = 0.775), had AUC values above 0.943 (Fig. 3F). However, when focusing solely on PTB and control samples from China, the AUC was similarly high at 0.923 (Fig. 3G). The average AUC across other countries remained high at 0.891 (Fig. 3H). These results demonstrate that the model can accurately identify PTB across diverse populations and geographic regions, reflecting the complexity of real-world settings.

Further validation was performed on additional datasets (Fig. 4A). Due to limited pediatric data, only one dataset (GSE39940) was available, which included 70 positive samples (PTB and EPTB) and 54 healthy controls from Malawi and South Africa. Despite the inclusion of EPTB cases, the AUC for this dataset reached 0.775 (Fig. 4B). The model’s robustness was also confirmed under specific conditions, including BCG vaccination, PTB with HIV infection, PTB with diabetes, and drug-resistant PTB, with high AUC values observed across multiple datasets (Fig. 4C–F). Collectively, these findings highlight the model’s resilience to variations in age, comorbidities, and clinical settings, supporting its broad applicability.

Fig. 4.

Model performance on additional datasets, comparison with WHO-TPP criteria, and benchmarking against previous methods. A Distribution of samples in additional datasets. B Model performance by age group with children data. C–F Model performance under specific clinical conditions. G Comparison with WHO-TPP performance criteria. H–J Performance comparison with previous studies (based on external validation data excluding lung cancer cases)

The Youden method was applied to optimize sensitivity (83%) and specificity (86%) in the external validation set, although these values did not meet the community-based triage criteria. When sensitivity was set to 93%, the optimal specificity was 70% at a threshold of 0.469, meeting the WHO-TPP requirements. We further evaluated country-specific performance, where sensitivity and specificity similarly met the target criteria in most countries (Fig. 4G: 6/7). Re-evaluation of the model at this threshold for PTB with HIV infection and for children showed a sensitivity of 90% and specificity of 81% for PTB with HIV, and a sensitivity of 71% and specificity of 73% for children. These findings confirm that the model’s performance meets WHO-TPP targets under specific conditions.

To further assess the relative effectiveness of our model, we compared its performance against previously published methods. Due to the absence of previously reported feature genes in the lung cancer dataset, we evaluated our model using the external validation set excluding LC samples. The results demonstrated that our model outperformed prior methods both in distinguishing PTB from all controls collectively and from specific control categories individually (Fig. 4H-J).

Given the satisfactory performance of PTBD across diverse populations, clinical conditions, and geographic settings, we developed an easy-to-use web platform to facilitate its practical application [24]. The interface allows users to upload whole-blood gene expression data and obtain diagnostic predictions intuitively, without requiring bioinformatics expertise. The results are automatically visualized with classification outcomes, accompanied by downloadable tables for further analysis. A comprehensive tutorial is provided to guide users through the process, ensuring a smooth and reproducible workflow that meets the needs of both clinicians and researchers engaged in tuberculosis diagnosis and biomarker studies.

Cell-type-specific and transcriptome-based analyses revealing robust PTB feature genes in PTBD

We combined two datasets (GSE19443 and GSE42832) containing purified cell subpopulation data, resulting in four cell types: CD4 T cells, CD8 T cells, monocytes, and neutrophils. Each cell type included 26 samples, comprising 12 positive and 14 negative controls (Fig. 5A). Our model demonstrated good ability to distinguish PTB from controls based on the expression profiles of monocytes and neutrophils, whereas it showed limited discriminatory power when applied to CD4 and CD8 T cells (Fig. 5B). These findings suggest that monocytes and neutrophils are the main contributors to the peripheral blood features associated with PTB. This observation is consistent with previous studies, including Berry et al., who proposed that neutrophils predominantly drive the blood transcriptomic signatures of tuberculosis [21], as well as other research emphasizing the role of monocytes in the immune features of TB in blood [25–27].

Fig. 5.

Feature analysis. A Distribution of the purified cell subpopulation data. B The model's performance in distinguishing between positive samples and negative controls across different cell subpopulations. C Differentially expressed genes (DEGs) were identified between PTB and non-PTB samples. The upper panel shows volcano plots of DEGs across multiple datasets, where red dots represent dysregulated genes (adjusted p-value < 0.05 and |logFC|> 1), and arrows indicate the feature genes used in our model. The right panel summarizes the differential expression status of the three feature genes across different datasets, with white asterisks denoting datasets in which a gene was significantly differentially expressed and colors indicating the direction of dysregulation in PTB. The lower panel presents box plots showing the expression differences of the feature genes between PTB and non-PTB groups. Significance levels above each box indicate the statistical difference between each group and the PTB group, determined by two-sided Wilcoxon rank-sum tests: ****p < 0.0001; ***p < 0.001; **p < 0.01; *p < 0.05

To further explore the gene-level contributions, differential expression analysis was conducted on 11 datasets containing both PTB and control samples (Fig. 5C). Among the ten feature genes examined, only three—AIM2, SERPING1, and ST6GAL1—were consistently differentially expressed. Specifically, ST6GAL1 showed differential expression in only one dataset, whereas AIM2 and SERPING1 were differentially expressed in nine out of the eleven datasets. These findings strongly support AIM2 and SERPING1 as robust differentially expressed genes, both exhibiting elevated expression in PTB patients. Functionally, AIM2 plays a role in inflammasome assembly and pyroptosis, processes that may drive tissue damage and disease progression in PTB [28]. SERPING1 encodes the C1 inhibitor (C1-INH), which regulates the classical complement pathway by limiting its activation [29]. Interestingly, only two out of the five gene pairs contained differentially expressed genes, suggesting that the gene pair-based approach may uncover discriminative features that single-gene analyses might overlook.

Extended applications of PTBD in EPTB diagnosis, PTB progression monitoring, and treatment prognosis

We hypothesized that the features used in our model might also serve as indicators for PTB progression and could be applied to EPTB diagnosis. To evaluate this, we analyzed two EPTB datasets, GSE83456 and GSE144127. In GSE83456, which included 108 samples, the model achieved an AUC of 0.844 when distinguishing EPTB cases (n = 47) from healthy controls (n = 61). In GSE144127, comprising 100 EPTB cases and 122 negative controls, the AUC was 0.755 (Fig. 6A). The relatively lower performance in GSE144127 was likely due to the heterogeneity of the negative controls, which included healthy individuals as well as cases of chest infection, lower respiratory tract infection (LRTI), upper respiratory tract infection (URTI), LTBI, PNA, LC, and sarcoidosis. Despite this, the results suggest that the model can effectively contribute to the diagnosis of EPTB.

Fig. 6.

Model performance in EPTB diagnosis, PTB progression, and prognosis. A The performance of the model on EPTB. B-C Progression data (GSE94438): As the time from diagnosis approaches, the AUC increases, and the RGPscore changes from negative to positive, with differences compared to healthy controls becoming progressively significant, indicating the gradual emergence of features. For the RGPscore bar chart, the column height is the median RGPscore, and the line in the middle is the quartile line. D-E Non-progression data (GSE94438): At any stage, the model cannot distinguish between positive and negative samples, and there is no significant difference in RGPscore compared to healthy controls, indicating that non-progress patients did not exhibit our features. F-G Cured data (GSE40553): As treatment progresses, the AUC gradually decreases and eventually approaches around 0.5. Additionally, the RGPscore changes from positive to negative and the significance compared to LTBI gradually diminishes, indicating that our features gradually disappear with the completion of treatment. H-I Uncured data (GSE89403): No matter how long the treatment lasts, our model can accurately distinguish between positive and negative samples. Additionally, the RGPscore consistently exhibits significant differences compared to healthy controls, indicating the persistence of our features in patients who have treatment failure. J Cured and uncured data (GSE89403): Using uncured data as positive samples and cured data as negative samples, validate the performance of the model in predicting curative effects. Significance markers above each bar denote statistical differences between the corresponding group and the HC/LTBI group, determined by two-sided Wilcoxon rank-sum tests: ****p < 0.0001; ***p < 0.001; ** p < 0.01; * p < 0.05

For PTB progression, we analyzed the only available dataset with relevant follow-up data, GSE94438, which included samples from South Africa, Gambia, Ethiopia, and Uganda. The participants ranged in age from 8 to 60 years, and the dataset comprised 236 samples, including 89 progression cases, 127 non-progression cases, and 20 healthy controls. We stratified the progression samples into three time windows before diagnosis: 1–180 days (n = 49), 181–360 days (n = 19), and 361–540 days (n = 21). Using progression samples as positive cases and both non-progression and healthy controls as negatives, the model demonstrated increasing AUC values as the time of diagnosis approached, indicating that disease-specific transcriptomic features gradually emerged (Fig. 6B). Consistently, the RGPscore shifted from negative to positive with increasing proximity to diagnosis, and the difference between progression cases and controls became more pronounced (Fig. 6C).

We also divided the non-progression data, representing follow-up samples from individuals who did not develop PTB, into three time windows: 1–180 days (n = 55), 181–360 days (n = 31), and 361–540 days (n = 41) prior to the end of follow-up. Using non-progression samples as positive cases and healthy controls as negatives, the model showed no discriminatory ability in any time window (Fig. 6D), indicating that the features identified by our model were absent in non-progression individuals. Similarly, the RGPscore showed no significant differences compared with healthy controls at any time point (Wilcoxon p > 0.05) (Fig. 6E). These findings suggest that the features identified by our model are specific to PTB and reflect disease progression.

For PTB prognosis, we analyzed four datasets (GSE193978, GSE19435, GSE31348, and GSE40553) comprising follow-up data from PTB patients undergoing standard anti-tuberculosis treatment until cure. Notably, GSE40553 included the largest cohort, with 204 samples from England and South Africa. This dataset contained serial samples from 37 PTB patients at diagnosis, after 2 weeks and 2 months of treatment, at treatment completion, and post-treatment, as well as 38 LTBI cases as controls. The model’s AUC progressively declined as treatment advanced, with no discriminatory power between positive and negative samples after treatment completion, indicating that the transcriptomic features identified by our model diminished as PTB was cured (Fig. 6F). The RGPscore also shifted from positive to negative as treatment progressed, with the difference relative to LTBI controls decreasing over time and becoming non-significant at treatment completion (Wilcoxon p > 0.05) (Fig. 6G). Similar patterns were observed in the other datasets (Additional file 1: Fig. S2).

We further evaluated uncured PTB cases using GSE89403, which included 135 samples. The uncured cohort comprised 28 samples from 8 PTB patients who failed to respond to treatment, collected at diagnosis, 1 week, 4 weeks, and 24 weeks of therapy. This dataset also included 20 healthy controls and 87 cured samples. Using uncured data as positives and healthy controls as negatives, the model achieved an AUC exceeding 0.912, demonstrating that the disease-associated features persisted in patients with poor treatment response (Fig. 6H). The RGPscore was significantly higher in uncured cases compared to healthy controls (Wilcoxon p < 0.001) (Fig. 6I). Using uncured data as positives and cured data as negatives, the AUC was 0.861 (Fig. 6J), indicating that the model may have potential for predicting PTB treatment outcomes.

In summary, these results demonstrate that the features identified by our model are specific to PTB and dynamically reflect disease progression and treatment response. The model shows promise not only for diagnosing EPTB but also for monitoring PTB progression and prognosis. Its ability to distinguish between progression and non-progression, as well as between cured and uncured cases, highlights its potential utility for clinical decision-making and treatment evaluation in diverse settings.

Discussion

The global burden of PTB remains substantial, highlighting the urgent need for rapid, accurate, and non-invasive diagnostic approaches. In this study, we developed a PTB diagnostic model based on five gene pairs identified from peripheral blood transcriptome data using machine learning. The model was rigorously validated across a testing set and multiple independent datasets, demonstrating robust diagnostic performance under real-world conditions. It is applicable to diverse age groups and geographic regions and remains unaffected by confounding factors such as BCG vaccination, HIV co-infection, diabetes, or drug resistance. Notably, with a universal threshold, the model meets WHO performance requirements for community-based triage, children, and patients co-infected with HIV. Beyond its excellent diagnostic performance, the model also shows potential for predicting PTB treatment efficacy and diagnosing EPTB. More importantly, the RGPscore effectively reflects PTB progression and prognosis. Future longitudinal follow-up studies are warranted to further validate its effectiveness in monitoring treatment response and disease progression. Such studies will help determine whether dynamic changes in the RGPscore can serve as a reliable biomarker for treatment efficacy and relapse risk. Moreover, integrating the model into prospective clinical cohorts may facilitate personalized disease monitoring and guide therapeutic decision-making in PTB management.

Compared with existing studies, our model addresses several important limitations of previous work. Many earlier studies relied on limited control groups and did not account for the complexity of real-world conditions, often excluding patients with other lung diseases [30, 31]. Moreover, diagnostic models developed from homogeneous populations typically lack generalizability across different settings [12, 16, 32]. In contrast, our model demonstrates broad applicability, independent of age, region, and specific clinical circumstances. Importantly, it requires only ten genes, greatly enhancing its practicality compared to models using larger panels of feature genes [15, 21]. A key strength is that our model is based on gene expression rank rather than absolute expression levels, making it resistant to batch effects and preprocessing variability and thus suitable for implementation across different platforms and settings [18]. When compared with traditional diagnostic methods, our approach also offers several advantages. It is more convenient and accurate than TST and IGRA. Unlike sputum microscopy and culture, our blood-based method is easier to implement and enables faster results. Additionally, it avoids the risks associated with chest X-rays and CT scans. In comparison to Xpert MTB/RIF, our model not only accurately diagnoses PTB but also performs well in children and HIV co-infected patients, populations in which conventional methods often fall short.

From a practical perspective, the model only requires relative expression measurements of ten genes. With the increasing availability of portable nucleic acid quantification devices, this approach could be implemented for point-of-care testing, including in resource-limited settings [33]. Such devices can utilize stored energy from intermittent sources, ensuring operational stability even in areas with unreliable electricity supply [34, 35].

The biological relevance of the identified feature genes further supports the utility of the model. Two upregulated genes, SERPING1 and AIM2, are associated with PTB progression [28, 29], and have been proposed as biomarkers in previous studies [30, 36]. The elevated expression of these genes likely contributed to the identification of the SERPING1–SUCLG2 and AIM2–YIPF6 gene pairs. Downregulation of ACBD3 may impair type I interferon responses, reducing inflammation and weakening host defense [37], while ST6GAL1 downregulation could promote macrophage apoptosis and disease progression [38]. Although not all genes in the feature pairs were significantly differentially expressed, subtle expression changes may have contributed to the formation of pairs such as ACBD3–GALNT3 and AKR1A1–ST6GAL1. The functional roles of the NETO2–ZNF235 pair remain unclear and warrant further investigation.

Despite these strengths, our study has limitations. The sample size for certain conditions that closely mimic PTB was relatively small, and future work should expand these datasets. Furthermore, our current model addresses only binary classification; future studies could explore multi-class classification to enhance diagnostic resolution. In conclusion, we have developed and validated a peripheral blood-based diagnostic model for PTB, supported by an online web server [24], which offers a promising tool for reducing the global burden of PTB.

Conclusions

In conclusion, this study establishes a robust and broadly applicable non-invasive diagnostic model for PTB, which demonstrates consistent performance across diverse patient populations and clinical conditions. By integrating top-scoring pair with machine learning, PTBD not only achieves reliable discrimination of PTB against complex real-world backgrounds but also extends its utility to extrapulmonary tuberculosis and treatment outcome prediction. Importantly, PTBD uncovers novel transcriptomic biomarker signatures, exemplified by five diagnostic gene-pair expression patterns that capture key molecular features underlying PTB pathogenesis. These advances provide a powerful tool to facilitate early diagnosis, enable timely intervention, and ultimately contribute to reducing the global burden of tuberculosis.

Methods

Benchmark dataset construction

To integrate as much data as possible, both microarray data and sequencing data were incorporated into the analysis. Specifically, the study requires data of the following types:

Discovery and external validation data: To establish a model that aligns with the real world, the PTB samples were used as positive samples. Alongside latent tuberculosis infection and healthy control, several common lung diseases such as pneumonia, lung cancer, and pulmonary nodules were also included as negative samples. To ensure the final model can be applied across different countries, samples were sourced from multiple countries. Additionally, the above data only considers adults due to limited pediatric samples.

Other data: To validate the model’s applicability across age groups and real-world scenarios (e.g., BCG vaccination, PTB with HIV, PTB with diabetes, drug resistance PTB), data from children and individuals with these conditions were included.

Purified cell subpopulation data: Data on purified cell subpopulations were collected to explore which cells contribute most to the identified features.

EPTB data and PTB progression, prognosis data: Data on PTB progression, prognosis, and EPTB were included to assess the model’s utility for the prediction of PTB progression, prognosis, and EPTB.

According to the purposes and requirements, we performed keyword searches on two public databases: Gene Expression Omnibus (GEO) and European Nucleotide Archive (ENA) [39, 40], and finally collected a total of 354 datasets. The search strategy focused on identifying peripheral blood mRNA expression profiles related to pulmonary tuberculosis and other disease cohorts used for comparison. For PTB datasets, the following keywords were applied to maximize dataset coverage: “tuberculosis” OR “pulmonary tuberculosis” AND “Homo sapiens”. All human blood transcriptome datasets related to PTB were downloaded without excluding any potentially eligible dataset. For comparison cohorts, such as lung cancer, disease-specific queries were used to avoid the inclusion of unrelated samples. For instance: “lung cancer” AND “Homo sapiens”. To ensure the clinical relevance and consistency of included samples, we also reviewed the diagnostic criteria reported in each dataset. PTB was typically diagnosed based on sputum culture and/or bronchoalveolar lavage (BAL) fluid culture and/or sputum microscopy and/or suggestive clinical and imaging findings. EPTB required positive results supported by sputum culture and suggestive clinical and imaging findings, while latent LTBI was defined using TST and/or IGRA results. These criteria guided the subsequent exclusion of datasets lacking reliable diagnostic information.

Later on, irrelevant data, non-peripheral blood mRNA transcriptome data, and data without clear PTB and LTBI diagnostic methods were excluded. This exclusion process resulted in the retention of 31 datasets, which were then allocated by purpose. Additionally, two peripheral blood mRNA transcriptome datasets for lung cancer (LC) were also collected due to insufficient LC data (Fig. 1A) [41, 42]. The dataset used in this study can be accessed through http://i.uestc.edu.cn/PTBD-dataset.

Data processing

The final data set included microarray and sequencing data from various platforms. Due to differences in processing methods across datasets, all data (except GSE62525, for which raw data was unavailable) were reprocessed through a consistent bioinformatics pipeline to avoid any possible biases.

For microarray data, raw data were retrieved from GEO and processed with background correction and normalization. The ‘affy’ R package (v1.76.0) was used for Affymetrix data, and the ‘limma’ R package (v3.54.2) was used for Illumina and Agilent data. Probe IDs were mapped to Entrez Gene IDs and then to Gene Symbols using the GPL and NCBI “gene_info” file. This process ensures consistency in Gene Symbol across various datasets. In cases of multiple Entrez Gene IDs mapping to the same Gene Symbol, the one with the highest mean expression was retained.

For RNA-seq data, raw FASTQ files were assessed with FastQC (v0.11.9) and MultiQC (v1.13). Low-quality reads and adapter sequences were filtered using Fastp (v0.22.0). Later, the clean reads were aligned to the human reference genome (GRCh38) using STAR (v2.7.10b), and gene counts were quantified with HTSeq (v2.0.2). The expression values were normalized to Transcripts Per Million (TPM) and then transformed using the log2(TPM + 1). The Ensembl Gene IDs were subsequently converted to the Gene Symbols using the GTF file. Similarly, the gene with the highest mean expression was retained in case of multiple Ensembl Gene IDs mapping to the same Gene Symbol.

Furthermore, to ensure the presence of the features identified in the discovery data within the subsequent validation and analysis datasets, the intersection of genes across all datasets was calculated, which ultimately resulted in 11,912 genes.

Model construction

Feature extraction

The discovery data included multiple disease types, with each type split into a 7:3 ratio for training and testing sets. In the training set, PTB samples were treated as positive, and samples from other disease types as negative controls. Gene pair features were extracted using the TSP method [43–45]. The TSP procedure is as follows:

• (i)Assuming the columns of the gene expression profile represent samples (X₁, …, X_N) and rows represent genes (G₁, …, G_P), where the expression value of the i-th gene G_i in the n-th sample X_n is denoted as X_i,n. For binary classification, set the vector of class labels for the samples as (Y₁, …, Y_N), where Y_n belongs to C = {C₁, C₂};
• (ii)Within each sample, rank all genes based on their expression levels, replacing the expression values of each gene with its rank. This process results in a new matrix R, where the rank of the i-th gene G_i in the n-th sample X_n is denoted as R_i,n;
• (iii)
  Pair the genes to form gene pairs (G_i, G_j) where i, j belong to {1, …, P} and i ≠ j. For each gene pair (G_i, G_j), calculate the frequency P_ij of R_i < R_j in each of the two classes of samples;

  $$\begin{array}{c}{P}_{\mathit{ij}}\left(C_m\right)=Prob\left(R_i,<,R_j,|Y=,C_m\right),m=\left\{1,,,2\right\}\end{array}$$ 1
• (iv)
  Here Δij represents the "score" for the gene pair (G_i, G_j), and select the gene pair(s) with the highest score as the top-scoring pair(s);

  $$\begin{array}{c}{\mathrm{\Delta }}_{\mathit{ij}}=\left|P_{\mathit{ij}},\left(C_1\right),-,P_{\mathit{ij}},\left(C_2\right)\right|\end{array}$$ 2
• (v)
  If the top-scoring pair is (G_i, G_j) and assuming P_ij(C₁) > P_ij(C₂), the new sample X_new is classified according to the following rule:

  $$\begin{array}{c}{Y}_{\mathit{new}}=\left\{\begin{array}{c}{C}_1,if{R}_{i,new}<R_{j,new}\\ C_2,Otherwise.\end{array}\right)\end{array}$$ 3

If P_ij(C₂) > P_ij(C₁), the classification rule is reversed. If there are multiple top-scoring pairs, the final class is determined based on the majority voting principle using the above rules.

TSP is used for feature extraction rather than classification in this study. The process stops at Step iv, when gene pair scores are obtained. The scores are sorted in descending order, and gene pairs are selected as features based on a threshold. If a gene appears in multiple pairs, only the pair with the highest score is retained. This method uses relative gene expression levels, enabling direct integration of microarray and RNA-seq data without concerns about technical noise. It is unaffected by platform or batch effects and eliminates the need for expression level adjustments, improving both convenience and accuracy.

Feature selection

Based on the extracted features, the training set underwent feature encoding. The column names were formatted as “Genei | Genej”, with the row names representing the sample names. The matrix values were encoded as 1 for Genei < Genej, 0 for Genei = Genej, and −1 for Genei > Genej. After feature encoding, feature importance ranking was conducted using six methods: Maximal Information Coefficient (MIC) [46], Chi-squared Test (Chi2) [47], Minimum Redundancy Maximum Relevance (mRMR) [48], Random Forest (RF) [49], Reversal Degree (RD) [50], and Average Weight (Avg) [51]. RD measures the geometric mean of the average rank difference of each gene pair between sample groups, with higher value indicating greater importance. Avg calculates the average rank of each gene pair across the five previous ranking methods, with lower value denoting higher importance.

Subsequently, an incremental feature selection (IFS) method was applied to refine the feature set [52–54]. Features from each ranking method were incrementally added to algorithms, including eXtreme Gradient Boosting (XGBoost), Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM) for ten-fold cross-validation and grid search. The XGBoost-based model was established using the Python package xgboost (v1.7.3), and the SVM, LR, and RF-based models were built using the Python package scikit-learn (v1.3.0). The final features were selected by balancing the number of features with the receiver operating characteristic (ROC) curve’s area under the curve (AUC).

Model selection and construction

The feature selection process inherently accounts for the feature, algorithm, and parameter optimization, identifying the optimal model for the final feature set. Nonetheless, to further improve the performance and stability of the machine learning models, ensemble algorithms such as Soft Voting and Stacking were employed. Soft Voting combines the prediction results from multiple base models (XGBoost, LR, RF, SVM) through a weighted average to obtain the final prediction result. Conversely, Stacking feeds the prediction of these same base models (XGBoost, LR, RF, SVM) into a meta-model, specifically LR, to integrate the results and yield the final prediction.

Finally, multiple algorithms (XGBoost, LR, RF, SVM, Soft Voting, Stacking) are applied to the training set. The optimal model is selected according to the AUC from ten-fold cross-validation, and the entire training dataset is used for constructing the final model.

Model validation

Validation on testing and external validation sets

The model’s performance was evaluated across categories and countries using both the testing and external validation sets. Specifically, the categorical validation involved comparing all PTB samples with all control samples (CTRL) and individual comparisons with LTBI, healthy control (HC), pneumonia (PNA), LC, and pulmonary nodules (PN) samples. The geographical validation compared PTB samples from each country with all control samples in the dataset. AUC was used to assess model performance.

Comprehensive validation

In addition to categorical and geographical validation, the model’s suitability across age groups and robustness under special conditions were also assessed. This included analysis of children, individuals with BCG vaccination, PTB cases with HIV infection, PTB cases with diabetes, and drug-resistant cases.

According to WHO-TPPs, community-based triage should have sensitivity > 90% and specificity > 70%. For PTB in HIV-positive patients, sensitivity should be ≥ 80%, and for pediatric cases, it should be > 66%. To evaluate the model’s performance against these TPPs, two methods were used: (1) the Youden method, which maximizes both sensitivity and specificity, and (2) setting sensitivity to at least 90% to determine the corresponding specificity.

Model comparison

Many studies on PTB diagnosis can be categorized into scoring-based and model-based classification. In comparing our model with existing ones, we aimed for faithful reproduction, but the lack of published data or model parameters in many studies, as well as the large number of features (which may not be available in the external validation set), could skew the comparison. Thus, we focused on comparing our model with two top-performing scoring-based methods: the “TB score” by Sweeney et al. [18]. and the “RISK6geo score” by Penn-Nicholson et al. [36] (Additional File 2: Table S5). Notably, the “TB score” has demonstrated superiority over several multi-feature diagnostic models. Finally, adhering to the methodologies described in their original publications, we reproduced their scoring methods and compared their performance with ours on the external validation set.

Feature analysis

Cell subpopulation analysis

To investigate which immune cell types contribute most to the identified peripheral blood gene-pair features, we applied our identified features to transcriptome data of purified cell subpopulations, including CD4 T cells, CD8 T cells, monocytes, and neutrophils. For each cell type, we calculated the AUC to assess the ability of these identified features to distinguish PTB from non-PTB samples. A high AUC in a given cell type indicates that the corresponding features are primarily attributable to that cell type, providing biological interpretability for the model.

Differential gene expression analysis

To identify truly differentially expressed genes, we performed differential analysis on the training set using the ‘limma’ R package (v3.54.2). Genes were considered differentially expressed if they had a Benjamini-Hochberg (BH) adjusted p-value < 0.05 and an absolute logarithmic fold change > 1. The frequency of each gene being identified as differentially expressed was used to determine the final set of differentially expressed genes.

Extended applications

We expected the identified features to be PTB-specific, allowing the model to also diagnose EPTB and predict PTB progression and prognosis. To test this, we collected relevant data and assessed model performance using AUC. Additionally, to better validate the correlation between our features and disease progression and prognosis, we introduced the Reverse Gene Pair Score (RGPscore):

• (i)Identify the less expressed gene in each gene pair in the control group;
• (ii)For each gene pair, calculate the difference by subtracting the expression level of the less-expressed gene in the control group from that of its paired gene. The final RGPscore is the mean of these computed differences;
• (iii)
  Interpret RGPscore values: In control samples, the RGPscore is typically negative, reflecting the subtraction of lower expressed genes from the higher expressed genes. Conversely, in PTB samples, the relationships between gene expressions are inverted, leading to a positive RGPscore. This polarity aids in analyzing whether the identified features can reflect different stages of disease progression or prognosis. The formula for the RGPscore used in this study is as follows:

  $$\begin{array}{c}RGPscore=\frac{\begin{array}{c}\left(S,E,R,P,I,N,G,1,-,S,U,C,L,G,2\right)+\left(G,A,L,N,T,3,-,A,C,B,D,3\right)+\\ \left(A,K,R,1,A,1,-,S,T,6,G,A,L,1\right)+\left(N,E,T,O,2,-,Z,N,F,235\right)+\left(A,I,M,2,-,Y,I,P,F,6\right)\end{array}}{5}\end{array}$$ 4

Statistical analysis

The between-group comparisons of the RGPscore were performed using the Wilcoxon rank-sum test, with a p-value less than 0.05 indicating statistical significance. All statistical analyses were conducted using R (v4.2.3).

Abbreviations

PTB

Pulmonary tuberculosis

TB

Tuberculosis

MTB

Mycobacterium tuberculosis

TST

Tuberculin skin test

IGRA

Interferon-gamma release assays

LTBI

Latent tuberculosis infection

BCG

Bacillus Calmette-Guérin

EPTB

Extrapulmonary tuberculosis

LC

Lung cancer

MIC

Maximal information coefficient

Chi2

Chi-squared Test

mRMR

Minimum redundancy maximum relevance

RF

Random forest

RD

Reversal degree

Avg

Average weight

IFS

Incremental feature selection

XGBoost

EXtreme Gradient Boosting

LR

Logistic regression

SVM

Support vector machine

TSP

Top-scoring pair

ROC

Receiver operating characteristic

AUC

Area under the curve

PNA

Pneumonia

PN

Pulmonary nodules

DEG

Differentially expressed genes

BAL

Bronchoalveolar lavage

Authors’ contributions

Conceptualization: J.H. and H.L. Data curation: C.W. and X.X.. Methodology: C.W., Z.H. and Y.Z.. Validation: Y.G. and M.D.. Writing—original draft: C.W.. Writing—review and editing: H.L. and J.H.. All authors read and approved the final manuscript.

Funding

This work was supported by grant from the National Natural Science Foundation of China (62501110, 62371112), Medico-Engineering Cooperation Funds from University of Electronic Science and Technology of China (ZYGX2022YGRH004), and Incubation Program for Innovative Science and Technology in UESTC (Y03023206100209).

Data availability

The datasets generated or analyzed during the current study are available in the published materials referenced in this manuscript, and can be accessed through http://i.uestc.edu.cn/PTBD-dataset.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.