Multidimensional cfRNA response modeling identifies a 5-gene pair signature for high-robust pulmonary tuberculosis diagnosis.

Lack of non-invasive biomarkers hinders pulmonary tuberculosis (PTB) management. We developed a multidimensional machine learning framework to systematically evaluate five cell-free RNA (cfRNA)-derived host response modalities: immune cell infiltration, global transcriptional perturbation (expression- and rank-based), key genes, and gene pairs. While immune-cell and global perturbation models showed moderate efficacy, an optimized 5-gene pair classifier demonstrated robust diagnostic power. This signature achieved an AUC of 0.947 in the validation subset derived from the parent study and maintained 100% sensitivity (95% CI: 0.741-1.000) in HIV-coinfected individuals. Model scores correlated significantly with bacterial load (r = 0.65) and radiological severity. Kappa analysis confirmed substantial diagnostic agreement (kappa = 0.42-0.56) between this signature and chest X-ray systems, with the signature providing high discriminative power in cases with ambiguous radiological findings. This cfRNA-based 5-gene pair signature establishes a robust diagnostic paradigm, offering efficient PTB detection in resource-limited settings and a foundation for future point-of-care diagnostics.