Machine learning-based feature selection suggests circulating proteins as best at predicting and diagnosing...

Mycobacterium tuberculosis (Mtb), bacteria causing tuberculosis (TB), is a leading cause of morbidity and mortality worldwide. Even though billions of individuals have evidence of past or present Mtb infection and millions develop TB yearly, most individuals exposed to Mtb do not progress to TB (active disease) and successfully control (and perhaps eliminate) the infection. Factors determining the likelihood of TB progression of Mtb-exposed individuals remains poorly understood, however. Mtb-exposed non-human primates (NHPs) such as rhesus macaques (RMs) also exhibit variable likelihood of developing active disease. Previous analysis of Mtb-exposed RMs suggested that blood proteins such as C-reactive protein (CRP), albumin to globulin (A/G), and kynurenine-to-tryptophan ratios and blood leukocytes (e.g., neutrophil frequency) were best correlated with TB diagnosis. We extended this previous work by using several machine learning (ML) techniques to further quantify the relative contribution of 20 features (and their derivatives) to predict development or diagnosis of active disease in Mtb-exposed RMs. We found that independent of the specific ML technique (random forest, linear support vector machine, or logistic regression), CRP and A/G ratio values and lung CFU were most informative at discriminating between animals with active disease or no disease; furthermore, when excluding endpoint measurements (measurements done at necropsy), CRP.peak and A/G.bottom dramatically outperformed all other features at determining progression to active disease. Top 5 (when including all features) or top 2 (when excluding endpoint measurements) features together had nearly as high discriminatory power (AUC ≥0.98) as all 20+ features. Importantly, in our analysis, frequency of blood leukocytes (e.g., percent of neutrophils or lymphocytes) or their ratios (e.g., neutrophil to lymphocyte ratio) had relatively poorer predictive or diagnostic value (AUC<0.8). Our results thus suggest that a combination of CRP and A/G ratio have a high power at predicting and/or diagnosing TB in Mtb-exposed RMs. Future studies will need to investigate if temporal changes in blood protein concentrations and leukocyte frequencies may further improve our ability to predict and/or diagnose TB in NHPs.