Diagnostic performance of host protein signatures as a triage test for active pulmonary TB

Abstract

The current four-symptom screen recommended by the World Health Organization (WHO) is widely used as screen to initiate diagnostic testing for active pulmonary tuberculosis (TB), yet the performance is poor especially when TB prevalence is low. In contrast, more sensitive molecular tests are less suitable for placement at primary care level in low-resource settings. In order to meet the WHO End TB targets, new diagnostic approaches are urgently needed to find the missing undiagnosed cases. Proteomics-derived blood host biomarkers have been explored because protein detection technologies are suitable for the point-of-care setting and could meet cost targets. This study aimed to find a biomarker signature that fulfills WHO's target product profile (TPP) for a TB screening. Twelve blood-based protein biomarkers from three sample populations (Vietnam, Peru, and South Africa) were analyzed individually and in combinations via advanced statistical methods and machine learning algorithms. The combination of I-309, SYWC and kallistatin showed the most promising results to discern active TB throughout the data sets meeting the TPP for a triage test in adults from two countries (Peru and South Africa). The top-performing individual markers identified at the global level (I-309 and SYWC) were also among the best-performing markers at country level in South Africa and Vietnam. This analysis clearly shows that a host protein biomarker assay is feasible in adults for certain geographical regions based on one or two biomarkers with a performance that meets minimal WHO TPP criteria.

Keywords: Mycobacterium tuberculosis; biomarker; diagnostics; host marker; machine learning.

Fig 1

Stratification of the population into country, TB and HIV status.

Fig 2

Receiver operator characteristic (ROC) curve of different machine learning algorithms with fivefold cross-validation for best performing (A) single biomarker, (B) combination of two, and (C) three different biomarkers. The dotted light gray line indicates the minimal TPP target with 90% sensitivity and 70% specificity. With the signature consisting of I-309, SYWC, and kallistatin the TPP could be reached by the random forest algorithm. AUC, area under the curve; SVM, support vector machine.

Fig 3

The ROC plot of all 12 biomarkers combined for (A) South Africa alone, (B) Peru alone, (C) Vietnam alone, (D) the whole data set, and (E) the whole data set excluding Vietnam. The dotted light gray line indicates the TPP with 90% sensitivity and 70% specificity. AUC, area under the curve; SVM, support vector machine.