Machine-Learning-Derived, Mechanistically Informed Transcriptomic Signature to Diagnose Active Tuberculosis and Guide...

Area of Science:

• Infectious Diseases
• Genomics
• Biomarker Discovery

Background:

• Differentiating active tuberculosis (TB) from latent TB infection (LTBI) is a significant diagnostic challenge.
• Current biomarkers offer limited insight into TB pathogenesis for guiding treatment decisions.
• A need exists for novel diagnostic tools and therapeutic targets in TB management.

Purpose of the Study:

• To develop a novel two-mode biomarker signature for differentiating active TB from LTBI.
• To identify host pathways involved in TB pathogenesis for potential therapeutic interventions.
• To create a transcriptomic diagnostic tool aligned with World Health Organization (WHO) performance targets.

Main Methods:

• Multicohort transcriptomic analysis incorporating stringent machine learning pipelines.
• Feature selection using ANOVA, Boruta-XGBoost, and LASSO regression to identify key genes.
• Development of an ensemble stacking classifier (Random Forest and XGBoost) for diagnostic performance evaluation.

Main Results:

• A four-gene signature (TAP2, SORT1, WARS, ANKRD22) was identified, significantly upregulated in active TB.
• The signature achieved high diagnostic accuracy (ROC-AUC = 0.991) in stratifying infection phases, validated in an independent cohort.
• Genes mapped to core dysregulated host pathways: antigen presentation, lipid trafficking, interferon response, and inflammasome signaling.

Conclusions:

• The developed biomarker signature provides a highly specific, non-sputum diagnostic for active TB.
• The signature offers a mechanistic map of host pathways, identifying potential targets for therapeutic intervention.
• This transcriptomic discovery facilitates the clinical implementation of new strategies against TB.