Using a consensus machine learning approach to identify host biomarkers associated with treatment response in pulmonary tuberculosis.

Abstract

The lack of reliable biomarkers to monitor treatment response constitutes a major challenge in tuberculosis (TB) management, often leading to under- or overtreatment with the standard six-month regimen. Here, we integrated amino acid and trace element profiles, along with four machine learning-based feature selection methods, to develop an evaluation model. Additionally, we investigated the dynamic expression patterns of genes associated with the identified features during TB treatment to gain functional insights. The consensus feature selection from the four methodologies identified 34 potential biomarkers. Notably, three trace elements-Antimony (Sb), Copper (Cu), and Strontium (Sr)-were consistently highlighted as significant biomarkers across all employed methods. Re-analysis of two public RNA-seq datasets revealed that genes associated with these three trace elements yielded higher Area Under Curve values in distinguishing cured TB from untreated TB. Subsequent gene expression profiling showed dynamic expression patterns for genes linked to these biomarkers, predominantly involving pathways of programmed cell death, signal transduction, and immune response. These findings suggest that the combination of Sb, Cu, and Sr may serve as a novel laboratory criterion for monitoring TB treatment response, providing a practical tool to guide clinical decisions and reduce adverse outcomes.