Novel RT‒qPCR-based blood signatures for the diagnosis of pulmonary tuberculosis Authors: Thainá Horta, Camilly Hipólito, Luana E. Araújo, Camila Pimentel, Dimitri Sokolowskei, Silvania Cerqueira, Flávia Marília Fonseca Oliveira, Carlos Oliveira, Jéssica Petrilli, Adriano Queiroz, Sérgio Arruda Published: 2026 ABSTRACT Whole-blood gene expression has the potential to become an affordable alternative for the diagnosis of pulmonary tuberculosis (TB). Through an immune-based gene expression profile via the NanoString platform, we previously identified 12 genes from whole blood that distinguished active TB from other pulmonary diseases. Here, we performed a three-phase diagnostic test evaluation to determine the scientific validity and analytical and clinical performance of these 12 previously identified genes in accurately diagnosing active TB. We further employed binary classification machine learning models to assess and evaluate transcriptional signatures. Samples from 175 adults were included in the study. During phase one, we assessed the expression of 12 genes in 10 TB patients and 9 healthy controls. Elevated expression of FCGR1A, GBP5, PDCD1LG2, and IRAK3 was observed in TB patients compared with healthy controls (p value < 0.05). The analytical performance of these four genes was evaluated in a cohort of 100 symptomatic respiratory volunteers screened for TB, comprising 35 tuberculosis patients and 65 non-tuberculosis controls. The expression of FCGR1A, PDCD1LG2, and GBP5 consistently distinguished patients from both groups, with an area under the ROC curve (AUC) ≥ 0.75. These genes were combined into three transcriptional signatures (FCGB-2TB, FPG-3TB, and FCPD-2TB), which demonstrated good performance in a pilot clinical validation study phase (N = 56; sensitivity > 77%, specificity > 94%). The present analysis allowed the discovery of RT‒qPCR-based signatures for TB diagnosis. These signatures could serve as affordable alternatives for TB diagnosis at primary care centers. Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-60489-8. Introduction Tuberculosis (TB) is an infectious disease caused by Mycobacterium tuberculosis (Mtb) and ranks among the top 10 causes of death worldwide attributable to infectious agents1. Despite advances in diagnostics, all sputum-dependent TB diagnostic tests face several limitations. Direct sputum smear microscopy has low sensitivity, ranging between 60% and 80%2,3, and the results of bacteriological culture take approximately 45 days to release. Xpert MTB/RIF has enhanced diagnostic capabilities; however, the requirement of specific equipment for testing significantly increases costs, halting its widespread implementation for primary care centers, especially in poor and developing countries where funding is limited4–6. The diagnosis of TB in countries where the disease is endemic faces some challenges. For example, in Brazil, where the prevalence of TB cases is 0.05%7, the interferon gamma release assay (IGRA) test has not been widely used for latent tuberculosis infection screening in high-risk groups. The purified protein derivative test (PPD) is highly compromised by Bacillus Calmette-Guérin (BCG) vaccination (largely applied in this country)8. Thus, early detection and treatment of new Mtb infections are very limited, increasing the number of people with active TB who need a sputum-based diagnostic test9. Additionally, in Brazil, Xpert MTB/RIF is not available in many small countryside cities located far from major urban areas. TB-suspected individuals might need to either travel to larger cities for diagnosis or be subjected to sputum smear or culture tests10,11. Similarly, many other poor and developing countries face similar diagnostic challenges, prompting the need for new suitable tests that can be readily available for use at primary care centers. Given the current limitations of available tests, the World Health Organization (WHO) advocates the development of new tests with high sensitivity that do not rely on sputum samples, such as blood, while being cost-effective12. Several studies have explored the host response to Mtb infection through whole-genome transcriptional profiling, identifying genes capable of discriminating TB patients from either healthy individuals or those with other pulmonary diseases (OPDs)6,13–15,22,23. However, these studies only achieved reasonable diagnostic parameter improvements by using several gene markers, thus making it cost-ineffective, laborious, and technically challenging to adopt in funding-capped or infrastructure-limited primary care centers. Previous analysis with the NanoString platform allowed us to identify 23 whole-blood markers that significantly differentiated TB patients from OPD patients14. Among them, we identified the genes FCGR1A, PDCD1LG2, and GBP5, which have previously shown the ability to distinguish TB and OPD patients with sensitivities and specificities higher than 90% and 100%, respectively13,14. Here, we selected 12 genes that presented the best performance, according to a previous study, and used Real-time reverse-transcription quantitative polymerase chain reaction (RT-qPCR) to determine the ability of these markers to diagnose TB. We identified two- and three-gene transcriptional signatures that were used to diagnose TB with 90% accuracy and sensitivity and specificity above 77%. Methods Study design and population We evaluated the diagnostic performance of whole-blood genes in three phases. For phase one, 19 RNA samples that were stored at −80 °C and collected from 10 patients with active TB and nine healthy controls (HCs) were included (Supplementary Table 2). Patients with TB were recruited from 2017 to 2020. The healthy controls were individuals with no respiratory symptoms, which was confirmed by a negative result for active and latent tuberculosis screening. Additionally, individuals in this group were household contacts of patients with TB. Household contacts were defined as those who lived in the same residence or reported at least 100 h of contact with bacilliferous patients. For phases two and three, respectively, technical and clinical validation studies were performed using a subset of prospectively recruited patients from a symptomatic respiratory disease cohort. Patients with comorbidities known to significantly affect gene expression (such as human immunodeficiency virus-positive individuals, those with a previous cancer diagnosis, and those with immunosuppressive drug usage) were excluded. We enrolled patients under investigation for pulmonary tuberculosis between December 2021 and October 2024 at the Octávio Mangabeira Specialized Hospital in Bahia, Brazil (Supplementary Tables 3 and 4). According to the diagnosis, those patients were then categorized into pulmonary active tuberculosis (TB) or non-TB symptomatic respiratory (SR) groups. Patients had a confirmatory diagnosis of TB if they tested positive on either the GeneXpert MTB/RIF PCR test or if at least one Mtb culture was positive. All the subjects were between 18 and 65 years old. Written informed consent was obtained from all participants enrolled in this study. A total of 204 participants under investigation for pulmonary tuberculosis were enrolled between December 2021 and August 2024 at the Octávio Mangabeira Specialized Hospital in Bahia, Brazil (Fig. 1). According to the TB diagnosis, those patients were categorized into pulmonary active TB (n = 53) and non-TB (n = 103) groups on the basis of positive or negative test results, respectively, for the GeneXpert MTB/RIF PCR test or at least a positive Mtb culture. Sample collection and RNA isolation For the samples from phase one, RNA was isolated and purified via the PAXgene Blood RNA Kit (Qiagen) according to the manufacturer’s instructions. For each donor in phases two and three, we collected three mL of peripheral blood in a Tempus Blood RNA tube (Thermo Fisher). After two hours at room temperature, the samples were frozen at −20 °C until RNA extraction. RNA was isolated and purified via the Tempus Spin RNA Isolation Kit following the manufacturer’s instructions, including on-column DNase treatment to eliminate genomic DNA contamination. RNA concentration and purity were assessed using a Nanodrop spectrophotometer. All samples presented A260/A280 ratios within the acceptable range for high-quality RNA. Gene expression analysis Twelve whole-blood gene markers that we previously found to differentiate active TB from asthma were selected14(Supplementary Table 1) and their specificity has been previously validated. A total of 90 to 300 ng of RNA from each sample was used for real-time PCR amplification, depending on RNA availability, while ensuring consistent reaction conditions across samples. cDNA was generated via reverse transcription and quantified via real-time PCR via the Superscript III Platinum SYBR Green One-Step qPCR w/ROX Kit following the manufacturer’s specifications. Real-time PCR reactions were performed in a final volume of 25 µL, and the concentrations of Mg²⁺ and dNTPs were used according to the manufacturer’s instructions. Reverse transcription was carried out using a one-step RT-qPCR approach with gene-specific priming, as provided by the kit. An Applied Biosystems 7500 Real-Time PCR system was used for detection. Amplification specificity was confirmed by melting curve analysis. No-template controls were included and showed no amplification. The expression of all the tested genes was normalized to that of Glyceraldehyde 3-phosphate dehydrogenase (GAPDH), which was selected as the reference gene based on its widespread use and reported stability in whole-blood gene expression studies16. To assess the diagnostic parameters of phase two, gene expression was calculated and represented as the 2−ΔΔCTfold change. A pool of RNA from five healthy control samples was prepared, and its threshold cycle (CT) value was used to assess the relative expression. Thus, the data are presented as the fold change in gene expression normalized to that of GAPDH and relative to the healthy control pool. Additionally, for each sample, we calculated the reactivity index (RI) to determine the positive and negative results on the basis of the fold change values and the calculated cutoff. It also contributed to data normalization and standardization. The RI was calculated according to the formula RI = 2− ΔΔCT/cutoff, where the cutoff value was defined on the basis of the values of sensitivity and specificity of each gene. All samples with values equal to or greater than 1.00 (RI > 1.00) were considered positive, whereas those with values less than 1 (RI < 1.00) were considered negative. For the phase three analysis, we used the ΔCT values to assess the diagnostic accuracy of the transcriptional markers. Statistical analysis The ability of genes to differentiate study groups was evaluated for sensitivity, specificity, and AUC through receiver operating characteristic (ROC) curve analysis. The Youden method was employed to determine the optimal cutoff point for the ROC curve. The Shapiro‒Wilk test was used for data assessment. Upon rejecting the null hypothesis, the Mann‒Whitney test was applied, and when data normality was confirmed, the t-test was used. The significance level was set at p < 0.05 with a 95% confidence interval (CI). The assay performance was calculated via a dichotomous approach, and parameters such as accuracy, likelihood ratio, odds ratio, and kappa index were evaluated. Signature evaluation During phase two analysis and via a regression model, transcriptional signatures were identified in the training set and tested in the test set via random forest (Liaw and Wiener, 2002), and graphs were generated via the ggplot2 package. The data were encoded and analyzed via scatterplot computer graphic software (Prism, version 8.0; GraphPad, San Diego, CA) and R Studio. Machine learning (ML) model classifiers were trained using previously labeled phase two transcriptional signatures. First, the datasets were randomly partitioned into training and testing datasets at an 80:20 ratio. The following ML classification algorithms were harnessed: random forest, xgboost, logistic regression, and support vector machine. Classifier model training uses cross-validation k-fold with k = 5 to mitigate overfitting and improve the model’s performance. Each classifier’s overall performance was assessed via a receiver operating characteristic (ROC) curve. The R programming language (v4.4.1) and the packages “caret”, “pROC” and “tidyverse” were used for ML analysis and data visualization. The code used to train the ML models is available on GitHub (https://github.com/dimitrisokolowskei/signaturesML). Results Study population characteristics Samples from 175 participants were included in the study: 19 for Phase one (nine HCs and 10 with TB) and 156 from the symptomatic respiratory cohort, 100 patients for Phases two and 56 for Phase three (Fig. 1). The laboratory results of the studied population included in phase one are depicted in Supplementary Table 2. The sociodemographic, clinical, and laboratory characteristics of patients from symptomatic respiratory cohorts are shown in Supplementary Table 3. The median age of the TB patients was 37 ± 12.63 years, and that of the SR patients was 43 ± 12.32 years (p = 0.0187). Most of the SR patients presented symptoms common in patients with active tuberculosis, such as cough (87.5%), chills (64%), and dyspnea (69%); however, the number of participants with chill, fever, and weight loss was greater in the TB group than in the SR group (p < 0.05) (Supplementary Table 3). Among the 156 patients included in the symptomatic respiratory cohort, 134 resided in the same city where they were diagnosed, 14 lived at distances ranging from 27 to 100 km, and 8 resided more than 100 km from the locality where the diagnostic test was performed (Supplementary Fig. S1). Fig. 1Overview of phases two and three of the study. schematic representation of participant recruitment and analysis for Phases two and three. created with BioRender.com. Overview of phases two and three of the study. schematic representation of participant recruitment and analysis for Phases two and three. created with BioRender.com. Scientific validation of the gene targets In the first phase, the ability of 12 genes (FCGR1A, GBP5, IRAK3, PDCD1LG2, MAPK14, IFITM1, CD274, CD59, ICAM1, PML, CR1, and C1QA) to discriminate 10 TB and nine HC participants was assessed (Fig. 2A). Based on the gene expression levels, only the genes FCGR1A (p value < 0.001), GBP5 (p value < 0.005), PDCD1LG2 (p value < 0.05) and IRAK (p value < 0.05) were capable of distinguishing TB patients from HCs (Fig. 1A). ROC curve analysis revealed that the FCGR1A gene presented the best parameters (area under the curve (AUC) = 0.93, sensitivity = 89.9% and specificity = 90%), followed by GBP5, with AUC = 0.83, sensitivity = 66.7% and specificity = 90% (Fig. 2B). Fig. 2Evaluation of the expression of 12 genes in healthy individuals and tuberculosis patients. (A) Gene expression levels. (B) ROC curve analysis. Each point in the scatter plot represents the expression level of the target gene in each patient sample. The cycle threshold (Ct) values from real-time PCRs were normalized to the internal reference gene (Housekeeping) GAPDH. Statistical differences between groups were assessed by t tests. AUC, sensitivity, and specificity values were calculated. Significance was considered at p < 0.05. ***p < 0.001; **p < 0.005 *p < 0.05; ns: not significant. The established confidence interval was 95%. Evaluation of the expression of 12 genes in healthy individuals and tuberculosis patients. (A) Gene expression levels. (B) ROC curve analysis. Each point in the scatter plot represents the expression level of the target gene in each patient sample. The cycle threshold (Ct) values from real-time PCRs were normalized to the internal reference gene (Housekeeping) GAPDH. Statistical differences between groups were assessed by t tests. AUC, sensitivity, and specificity values were calculated. Significance was considered at p < 0.05. ***p < 0.001; **p < 0.005 *p < 0.05; ns: not significant. The established confidence interval was 95%. Analytical performance of the FCGR1A, GBP5, PDCD1LG2 and IRAK3 genes Genes with better performance in previous analyses were tested for their ability to distinguish TB (N = 35) and SR (N = 65) patients recruited from the symptomatic respiratory cohort (Fig. 3). The genes FCGR1A, PDCD1LG2, and GBP5 were differentially expressed among patients in the two groups (p values < 0.0001), whereas no difference was detected in the expression of the IRAK3 gene (p = 0.4682). ROC curve analysis of FCGR1A revealed an AUC of 0.87 (95% CI 0.80–0.95), with an optimal sensitivity of 82.9% (95% CI 67.3–91.9), specificity of 78.5% (95% CI 67.0–86.7), and accuracy of 80% (95% CI 71.1–86.7) (Fig. 3; Table 1). The PDCD1LG2 and GBP5 genes also discriminated the groups with AUCs of 0.80 (95% CI 0.70–0.90) and 0.80 (95% CI 0.71–0.90), respectively (Fig. 3). IRAK3 failed to discriminate the TB and SR groups, with an AUC of 0.54 (95% CI 0.41–0.67), a sensitivity of 57.1% (95% CI 40.9–72.0), and a specificity of 63.1% (95% CI 50.9–73.5) (Fig. 3; Table 1). The likelihood ratio, diagnostic odds ratio, and index kappa of these genes are depicted in Table 1 and confirm the better performance of FCGR1A, PDCD1LG2, and GBP5. Fig. 3Differences in gene expression levels between the TB and SR groups and assessment of the discriminatory capacity of the genes FCGR1A, PDCD1LG2, GBP5, AND IRAK3. Each point in the scatter plot represents the expression level of the target gene in each patient sample after the reactivity index was calculated. The cycle threshold (Ct) values from real-time PCRs were normalized to the internal reference gene (Housekeeping) GAPDH. Statistical differences between groups were assessed by the Mann‒Whitney test. The dotted line at 1 represents the cutoff point. ROC curve analyses were conducted to individually assess the ability of the genes to distinguish TB patients from SR patients. The values of the AUC, sensitivity and specificity were calculated. Significance was considered at p < 0.05. ****p < 0.0001; ns: not significant. The established confidence interval was 95%. Differences in gene expression levels between the TB and SR groups and assessment of the discriminatory capacity of the genes FCGR1A, PDCD1LG2, GBP5, AND IRAK3. Each point in the scatter plot represents the expression level of the target gene in each patient sample after the reactivity index was calculated. The cycle threshold (Ct) values from real-time PCRs were normalized to the internal reference gene (Housekeeping) GAPDH. Statistical differences between groups were assessed by the Mann‒Whitney test. The dotted line at 1 represents the cutoff point. ROC curve analyses were conducted to individually assess the ability of the genes to distinguish TB patients from SR patients. The values of the AUC, sensitivity and specificity were calculated. Significance was considered at p < 0.05. ****p < 0.0001; ns: not significant. The established confidence interval was 95%. Table 1Evaluation of diagnostic performance parameters.FCGR1A(95% CI)GBP5(95% CI)PDCD1LG2(95% CI)IRAK3(95% CI)AUC0.87(0.80 – 0.95)0.80(0.71 – 0.90)0.80(0.70 – 0.90)0.54(0.41 – 0.67)Sensitivity (%)82.9(67.3 – 91.9)71.4(54.9 – 83.7)68.6(52.0 – 81.4)57.1(30.9 – 72.0)Specificity (%)78.5(67.0 – 86.7)84.6(73.9 – 91.4)87.7(77.5 – 93.6)63.1(50.9 – 73.5)Accuracy (%)80.0(71.1 – 86.7)80.0(71.1 – 86.7)81.0(72.2 – 87.5)61.0(51.2 – 70.0)LR +3.85(2.36 – 6.27)4.64(2.53 – 8.52)5.57(2.80 – 11.07)1.55(1.01 – 2.37)LR -0.22(0.10 – 0.46)0.34(0.20 – 0.58)0.36(0.22 – 0.59)0.68(0.44 – 1.06)DOR17.61(8.90 – .39)13.75(5.08 – 37.22)15.55(5.56 – 43.46)2.28(0.99 – 5.26)K Index0.58(0.42 – 0.75)0.56(0.39 – 0.73)0.57(0.40 – 0.75)0.19(-0.01 – 0.39)CI confidence interval, AUC area under curve, LR: likelihood ratio, DOR diagnostic odds ratio, K Index index kappa. Table 1Evaluation of diagnostic performance parameters.FCGR1A(95% CI)GBP5(95% CI)PDCD1LG2(95% CI)IRAK3(95% CI)AUC0.87(0.80 – 0.95)0.80(0.71 – 0.90)0.80(0.70 – 0.90)0.54(0.41 – 0.67)Sensitivity (%)82.9(67.3 – 91.9)71.4(54.9 – 83.7)68.6(52.0 – 81.4)57.1(30.9 – 72.0)Specificity (%)78.5(67.0 – 86.7)84.6(73.9 – 91.4)87.7(77.5 – 93.6)63.1(50.9 – 73.5)Accuracy (%)80.0(71.1 – 86.7)80.0(71.1 – 86.7)81.0(72.2 – 87.5)61.0(51.2 – 70.0)LR +3.85(2.36 – 6.27)4.64(2.53 – 8.52)5.57(2.80 – 11.07)1.55(1.01 – 2.37)LR -0.22(0.10 – 0.46)0.34(0.20 – 0.58)0.36(0.22 – 0.59)0.68(0.44 – 1.06)DOR17.61(8.90 – .39)13.75(5.08 – 37.22)15.55(5.56 – 43.46)2.28(0.99 – 5.26)K Index0.58(0.42 – 0.75)0.56(0.39 – 0.73)0.57(0.40 – 0.75)0.19(-0.01 – 0.39)CI confidence interval, AUC area under curve, LR: likelihood ratio, DOR diagnostic odds ratio, K Index index kappa. Evaluation of diagnostic performance parameters. 0.87 (0.80 – 0.95) 0.80 (0.71 – 0.90) 0.80 (0.70 – 0.90) 0.54 (0.41 – 0.67) 82.9 (67.3 – 91.9) 71.4 (54.9 – 83.7) 68.6 (52.0 – 81.4) 57.1 (30.9 – 72.0) 78.5 (67.0 – 86.7) 84.6 (73.9 – 91.4) 87.7 (77.5 – 93.6) 63.1 (50.9 – 73.5) 80.0 (71.1 – 86.7) 80.0 (71.1 – 86.7) 81.0 (72.2 – 87.5) 61.0 (51.2 – 70.0) 3.85 (2.36 – 6.27) 4.64 (2.53 – 8.52) 5.57 (2.80 – 11.07) 1.55 (1.01 – 2.37) 0.22 (0.10 – 0.46) 0.34 (0.20 – 0.58) 0.36 (0.22 – 0.59) 0.68 (0.44 – 1.06) 17.61 (8.90 – .39) 13.75 (5.08 – 37.22) 15.55 (5.56 – 43.46) 2.28 (0.99 – 5.26) 0.58 (0.42 – 0.75) 0.56 (0.39 – 0.73) 0.57 (0.40 – 0.75) 0.19 (-0.01 – 0.39) CI confidence interval, AUC area under curve, LR: likelihood ratio, DOR diagnostic odds ratio, K Index index kappa. Performance of transcriptional signatures Since the genes FCGR1A, PDCD1LG2, and GBP5 presented the best performances in the previous analysis, we then applied the random forest methodology to categorize individuals into training and testing sets to identify optimal gene combinations capable of discriminating against TB and SR patients. The performance of the nine transcriptional signatures identified is shown in Fig. 4. The two-gene signatures FCGR1A+PDCD1LG2 and IRAK3 + PDCD1LG2 presented an AUC of 0.88 (95% CI 0.7888–0.9643), with both sensitivity and specificity above 80%. IRAK3 + GBP5 presented the poorest performance, with an AUC of 0.73 (95% CI 0.6049–0.8742) and sensitivity below 60%; however, this signature showed the highest specificity among all combinations (94%). All six other signatures also demonstrated high performance, with AUCs, sensitivities and specificities higher than 0.86%, 70%, and 79%, respectively (Fig. 4). Fig. 4Performance of transcriptional signatures for active pulmonary TB diagnosis. Each gene combination was obtained via random forest analyses via R Studio software. The first numerical values below the curve refer to the cutoff point. The values within parentheses represent specificity and sensitivity, respectively. The established confidence interval was 95%. Performance of transcriptional signatures for active pulmonary TB diagnosis. Each gene combination was obtained via random forest analyses via R Studio software. The first numerical values below the curve refer to the cutoff point. The values within parentheses represent specificity and sensitivity, respectively. The established confidence interval was 95%. Clinical validity and accuracy of the transcriptional signature We chose the FCGR1A+PDCD1LG2 (FCPD-2TB), FCGR1A+PDCD1LG2 + GBP5 (FPG-3TB), and FCGR1A+GBP5 (FCGB-2TB) signatures for clinical validation. Participants from phase two (N = 100) were defined as the training set, and another 56 patients recruited later in the same cohort were assigned as the test set. The sociodemographic, clinical, and laboratory characteristics of these 56 patients are shown in Supplementary Tables 4, and no significant differences were found between the TB and SR groups. Machine learning model classifiers were trained via previously labeled phase two transcriptional signatures and applied to diagnose patients from the test set. The FCGB-2TB and FPG-3TB were able to distinguish TB from SR with high accuracy (Table 2), with sensitivities of 88.9% and 83.3% and specificities of 94,7% and 97,4%, respectively. The signature FCPD-2TB had a sensitivity of 77,8% and a specificity of 97,4%. A comparison of the identified signatures with a reference standard diagnostic test revealed that the signatures presented close performance to the ideal test (Fig. 5). Finally, the overall performance of the transcriptional signatures was assessed via confusion matrices and visualized via ROC curves (Supplementary Fig. S2). Based on this analysis, the FCGB-2TB achieved the best performance, with an AUC of 0.91, sensitivity of 100%, and specificity of 71,4%. Table 2Evaluation of the diagnostic performance parameters of the transcriptional signatureFCPD-2TB(95% CI)FCGB-2TB(95% CI)FPG-3TB(95% CI)Sensitivity (%)77.8(54.8 – 91.0)88.9(67.2 – 96.9)83.3(60.8 –94.2)Specificity (%)97.4(86.5 – 99.5)94.7(82.7 – 98.5)97.4(86.5 – 99.5)Accuracy (%)91.1(80.7 – 96.1)92.9(83.3 – 97.2)92.9(83.0 – 97.2)LR +29.6(4.2 – 207.7)16.9(4.3 – 65.7)31.7(4.5 – 221.5)LR -0.23(0.1 – 0.5)0.1(0.0 – 0.4)0.2(0.1 – 0.5)DOR129.5(13.3 – 1260.0)144.0(18.6 – 1114.7)185.0(17.8 – 1923.2)CI confidence interval, LR likelihood ratio, DOR diagnostic odds ratio. Table 2Evaluation of the diagnostic performance parameters of the transcriptional signatureFCPD-2TB(95% CI)FCGB-2TB(95% CI)FPG-3TB(95% CI)Sensitivity (%)77.8(54.8 – 91.0)88.9(67.2 – 96.9)83.3(60.8 –94.2)Specificity (%)97.4(86.5 – 99.5)94.7(82.7 – 98.5)97.4(86.5 – 99.5)Accuracy (%)91.1(80.7 – 96.1)92.9(83.3 – 97.2)92.9(83.0 – 97.2)LR +29.6(4.2 – 207.7)16.9(4.3 – 65.7)31.7(4.5 – 221.5)LR -0.23(0.1 – 0.5)0.1(0.0 – 0.4)0.2(0.1 – 0.5)DOR129.5(13.3 – 1260.0)144.0(18.6 – 1114.7)185.0(17.8 – 1923.2)CI confidence interval, LR likelihood ratio, DOR diagnostic odds ratio. Evaluation of the diagnostic performance parameters of the transcriptional signature 77.8 (54.8 – 91.0) 88.9 (67.2 – 96.9) 83.3 (60.8 –94.2) 97.4 (86.5 – 99.5) 94.7 (82.7 – 98.5) 97.4 (86.5 – 99.5) 91.1 (80.7 – 96.1) 92.9 (83.3 – 97.2) 92.9 (83.0 – 97.2) 29.6 (4.2 – 207.7) 16.9 (4.3 – 65.7) 31.7 (4.5 – 221.5) 0.23 (0.1 – 0.5) 0.1 (0.0 – 0.4) 0.2 (0.1 – 0.5) 129.5 (13.3 – 1260.0) 144.0 (18.6 – 1114.7) 185.0 (17.8 – 1923.2) CI confidence interval, LR likelihood ratio, DOR diagnostic odds ratio. Fig. 5Comparison of the identified signatures with a reference standard diagnostic test. (1) FCPD-2TB. (2) FPG-3TB. (3) FCGB-2TB. (4) Perfect test. Comparison of the identified signatures with a reference standard diagnostic test. (1) FCPD-2TB. (2) FPG-3TB. (3) FCGB-2TB. (4) Perfect test. Discussion Diagnosing tuberculosis is still a major challenge in many regions of the world, prompting the WHO to advocate for the development of new tests that do not rely on sputum samples, with a sensitivity above 65% overall and exceeding 98% for patients with positive sputum smear microscopy results12. These parameters are compared with the current performance of sputum smear microscopy and the rapid molecular test XPERT MTB/RIF. XPERT MTB/RIF represents an advance in the diagnosis of TB; however, the high cost of the equipment that is used to perform the test represents a barrier in countries with limited resources1,7,10,12. For example, in Brazil, according to the latest health ministry update, only 248 devices are available to attend to 5,565 municipalities10. Consequently, many patients rely on low-sensitivity and time-consuming approaches for diagnostics, such as sputum smear microscopy, or may need to travel to obtain a definitive diagnosis. In our study, gene expression analysis of whole-blood samples identified the genes FCGR1A, PDCD1LG2, and GBP5 that distinguished TB patients from SR patients with an AUC ≥ 0.80. The random forest analysis identified nine transcriptional signatures with AUCs, sensitivities, and specificities higher than 0.86, 70%, and 79%, respectively. Three of them, namely the FCPD-2TB, FPG-3TB, and FCGB-2TB, showed better performance and were evaluated in a pilot-blind clinical validation, which revealed that specificity and sensitivity ranged from 94.7% to 97.4% and from 77.8% to 88.9%, respectively. Although most of the genes and signatures identified herein achieved the overall minimum sensitivity established by the WHO (> 65%), none of the targets surpassed 98% sensitivity in individuals with positive sputum smear microscopy. A comparative analysis of transcriptional signatures (Supplementary Table 5) indicates that the models identified in our study show improved performance relative to previously reported signatures, including the widely used Sweeney 3-gene signature15. Our signatures achieved higher AUC values (0.86–0.88 vs. 0.83) and substantially higher specificity (94.7–97.4% vs. 76%), indicating better discrimination between TB and symptomatic respiratory conditions. When benchmarked against the WHO Target Product Profile (TPP)12, none of the signatures—including ours—met the criteria for a confirmatory test (specificity ≥ 98%); however, our models approached this threshold more closely than previous studies. For triage purposes, although specificity was adequate, sensitivity remained below the ≥ 90% threshold, resulting in partial fulfillment of TPP criteria across all signatures. The signature’s sensitivity and specificity also make the proposed diagnostic test a possible alternative for community-based triage or referral tests for identifying people suspected of having TB12. This is especially important, as broader implementation of XPERT MTB/RIF is hampered by limited infrastructure and expertise in decentralized health care centers17 in most countries with a high burden of TB. In our study, 22 out of 156 patients needed to travel to another city for TB diagnosis (Supplementary Fig. S1). Some aspects of the proposed test that make it suitable for a triage test include the fact that it is a whole-blood gene expression-based test, so there is no need for smear manipulation, and RNA can be extracted with extraction kits, which are easier to perform and require minimal infrastructure resources. The FCGR1A and GBP5 genes, identified in various studies, have proven capable of distinguishing TB patients from individuals with other pulmonary diseases (OPDs)13,18–20. Our results confirmed the differential expression of FCGR1A gene between TB patients and SR individuals, with an AUC of 0.87 (95% CI 0.80–0.95). Furthermore, the likelihood ratio (LR) values for this target suggest a moderate increase in the likelihood of correctly identifying and classifying individuals with TB (Table 1). The sensitivity, specificity, AUC, and accuracy values of the GBP5 gene found in this study also suggested that this gene is a potential biomarker for active TB, which is consistent with data reported in the literature15,21. We also detected differential expression of the PDCD1LG2 gene between the evaluated groups, demonstrating good discriminatory capacity with an accuracy of 81.0% (95% CI 72.2–87.5). PDCD1LG2 is a ligand of PD-1 involved in immune checkpoint regulation and modulation of T cell responses; however, interestingly, the expression of this gene in individuals with TB has been poorly described. Therefore, our findings identify a promising target for the development of new TB diagnostic tests. A major strength of this study is the identification of transcriptional signatures, including FCGR1A, GBP5, and PDCD1LG2, with robust diagnostic performance (AUC > 0.80). A peripheral blood sample used for RT-qPCR analysis can be collected in blood RNA tubes, which usually can be stored at −20 °C until RNA extraction, eliminating the need for a − 80 °C freezer. Additionally, gene expression analysis can be performed using standard RT-qPCR equipment. These reduced infrastructure requirements and operational advantages support the feasibility of the proposed assay as a triage or referral test in resource-limited settings, particularly where access to XPERT MTB/RIF is restricted. Furthermore, the identification of PDCD1LG2 provides a novel and relevant biomarker for active TB. Our study, however, has limitations. 1- The study population size is somewhat small, and the targets were evaluated in patients from only one hospital; thus, the results presented herein are biased to a restricted population. 2- Individuals with known HIV infection were excluded from the study due to its potential impact on host gene expression. However, HIV infection is known to influence transcriptional responses24, and its exclusion may limit the generalizability of our findings to HIV-positive populations, particularly in regions with high TB/HIV co-infection rates. 3- The symptomatic respiratory group is clinically heterogeneous, which limits a more granular evaluation of the transcriptional signature across specific respiratory conditions such as bacterial pneumonia. Due to incomplete and inconsistently annotated clinical data, a stratified analysis was not feasible in the present study. 4- Due to the low sample size, a hit is expected in the machine learning model’s generalization ability, which could be overcome with an increased number of patients for training and testing. 5- Most TB patients included in the present study had previously been vaccinated with BCG, precluding us from evaluating the impact of these transcriptional signatures in the non-vaccinated population. We understand that these signatures still need to be validated with external datasets or tested in a multi-cohort study to assess the feasibility and validity of these markers in different populations. However, in our study, we used data from the RT-qPCR platform and applied the ΔCt method to assess the gene expression values. This model precluded us from finding publicly available datasets generated with a similar approach to perform the external validation. Although GAPDH showed minimal variability across groups in our cohort, the use of a single reference gene represents a potential limitation. In this context, future studies may benefit from the incorporation of multiple reference genes or the application of geometric normalization strategies to further improve the robustness of gene expression analyses. Conclusions Our study identified and assessed the diagnostic performance of novel transcriptional signatures, comprising the genes FCGR1A, GBP5, and PDCD1LG2, for the diagnosis of TB. The signature FCGB-2TB, composed of the genes FCGR1A and GBP5, demonstrated excellent performance during clinical validation. Specifically, it achieved an AUC of 0.91 (Supplementary Figure S2), with a sensitivity of 88.9% and specificity of 94.7% (Table 2). These findings support the FCGB-2 TB signature as a promising candidate for further multicohort validation studies. The assay is compatible with standard RT-PCR platforms and does not rely on sputum samples, supporting their potential applicability in primary healthcare settings and partially aligning with WHO-prioritized efforts to develop non-sputum-based TB diagnostics. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2