Tuberculosis (TB) remains a formidable global health threat, yet rapid and accurate diagnostic biomarkers capturing host immune-metabolic dysregulation remain elusive. Here, we aimed to map the TB immune microenvironment and engineer a reliable, explainable diagnostic signature targeting the macrophage-associated immune-metabolic axis. Utilizing single-cell RNA sequencing (scRNA-seq) as an exploratory discovery tool, we initially dissected the intercellular communication network in TB. We then deployed an exhaustive consensus machine learning framework, comprising 113 algorithm combinations, across multiple transcriptomic cohorts to pinpoint core diagnostic features. The optimal model was interpreted using SHapley Additive exPlanations (SHAP) and prospectively evaluated by enzyme-linked immunosorbent assay (ELISA) in an independent pilot clinical cohort. Furthermore, network pharmacology and molecular docking were leveraged to identify potential small-molecule modulators. scRNA-seq analysis highlighted a myeloid-biased immune reprogramming, wherein activated monocytes act as key inflammatory orchestrators through adhesion and migration signaling. Our large-scale machine learning screening identified an optimal glmBoost + RF ensemble model underpinned by a 6-gene mitochondrial-macrophage signature (IL1B, ATG3, CYBB, MX1, RPS27A, RPS3), achieving consistent diagnostic discrimination (AUC > 0.79) across independent cohorts. Pilot clinical ELISA validation confirmed the systemic elevation of IL1B, ATG3, CYBB, and MX1 proteins in TB patients. Furthermore, computational molecular docking models suggested that the candidate phytochemicals galangin and kaempferol exhibit strong theoretical binding affinities within the catalytic pockets of IL1B and ATG3. We derived and provided preliminary validation for an AI-based, explainable 6-gene signature reflecting monocyte immune-metabolic reprogramming in TB. This signature not only demonstrates translational potential as a triage