BATF2 Expression Versus a Six-Gene Interferon-Associated Blood Signature for Distinguishing Active Pulmonary Tuberculosis from Latent Infection: A Two-Cohort Transcriptomic Analysis

Background: Host blood transcriptional signatures may help distinguish active pulmonary tuberculosis (PTB) from latent tuberculosis infection (LTBI), but their diagnostic performance and generalizability require further evaluation. This study investigated the discriminatory performance of BATF2 and compared it with a six-gene interferon-associated signature. Methods: Publicly available Illumina microarray datasets GSE37250 and GSE19491 were reanalyzed. The development cohort comprised 94 HIV-negative South African participants from GSE37250 (46 PTB and 48 LTBI), while the external evaluation cohort included 51 South African participants from GSE19491 (20 PTB and 31 LTBI). BATF2 expression was evaluated individually and compared with a six-gene panel comprising BATF2, GBP5, GBP1, GBP2, STAT1, and IFI35. Diagnostic discrimination was assessed using receiver operating characteristic (ROC) analysis. A ridge-regularized six-gene logistic regression model was evaluated using stratified 10-fold cross-validation, with penalty selection nested within the training folds. The trained model was subsequently applied to GSE19491 without refitting its coefficients. Results: BATF2 expression was higher in PTB than in LTBI in both cohorts. In GSE37250, BATF2 achieved an apparent area under the ROC curve (AUC) of 0.9515 (95% confidence interval (CI): 0.9007–1.0000), compared with 0.9751 for the apparent six-gene logistic regression model. A new, seeded nested 10-fold R cross-validation yielded AUCs of 0.9466 for BATF2 and 0.9461 for the six-gene ridge model (difference −0.0005; 95% conditional paired-bootstrap CI −0.0140 to 0.0136). In GSE19491, BATF2 achieved an AUC of 0.9774 (95% CI: 0.9475–1.0000), while the six-gene model achieved an AUC of 0.9839 (95% CI: 0.9576–1.0000). Differences in expression distributions between datasets limited the interpretation of absolute diagnostic thresholds. Conclusions: BATF2 alone demonstrated high discrimination between PTB and LTBI, without a demonstrable improvement in cross-validated AUC from the addition of five interferon-associated genes. These findings support further investigation of BATF2 as a potentially simpler host-response biomarker. However, standardized assays, prespecified diagnostic thresholds, and prospective validation in clinically representative populations are required before clinical application.