A blood RNA signature for tuberculosis disease risk: a prospective cohort study

Abstract

Background: Identification of blood biomarkers that prospectively predict progression of Mycobacterium tuberculosis infection to tuberculosis disease might lead to interventions that combat the tuberculosis epidemic. We aimed to assess whether global gene expression measured in whole blood of healthy people allowed identification of prospective signatures of risk of active tuberculosis disease.

Methods: In this prospective cohort study, we followed up healthy, South African adolescents aged 12-18 years from the adolescent cohort study (ACS) who were infected with M tuberculosis for 2 years. We collected blood samples from study participants every 6 months and monitored the adolescents for progression to tuberculosis disease. A prospective signature of risk was derived from whole blood RNA sequencing data by comparing participants who developed active tuberculosis disease (progressors) with those who remained healthy (matched controls). After adaptation to multiplex quantitative real-time PCR (qRT-PCR), the signature was used to predict tuberculosis disease in untouched adolescent samples and in samples from independent cohorts of South African and Gambian adult progressors and controls. Participants of the independent cohorts were household contacts of adults with active pulmonary tuberculosis disease.

Findings: Between July 6, 2005, and April 23, 2007, we enrolled 6363 participants from the ACS study and 4466 from independent South African and Gambian cohorts. 46 progressors and 107 matched controls were identified in the ACS cohort. A 16 gene signature of risk was identified. The signature predicted tuberculosis progression with a sensitivity of 66·1% (95% CI 63·2-68·9) and a specificity of 80·6% (79·2-82·0) in the 12 months preceding tuberculosis diagnosis. The risk signature was validated in an untouched group of adolescents (p=0·018 for RNA sequencing and p=0·0095 for qRT-PCR) and in the independent South African and Gambian cohorts (p values <0·0001 by qRT-PCR) with a sensitivity of 53·7% (42·6-64·3) and a specificity of 82·8% (76·7-86) in the 12 months preceding tuberculosis.

Interpretation: The whole blood tuberculosis risk signature prospectively identified people at risk of developing active tuberculosis, opening the possibility for targeted intervention to prevent the disease.

Funding: Bill & Melinda Gates Foundation, the National Institutes of Health, Aeras, the European Union, and the South African Medical Research Council.

Figure 1. The Adolescent Cohort Study (ACS) and the Grand Challenges 6-74 Study (GC6-74) cohorts for the discovery and validation of the tuberculosis risk signature

(A) Inclusion and exclusion of participants from the ACS and assignment of eligible progressors and controls to the training and test sets. QFT: Quantiferon Gold In-Tube. TST: tuberculin skin test. (B) Inclusion and exclusion of adult household contacts of patients with lung tuberculosis from the GC6-74 cohorts, and assignment of eligible progressors and controls. HHC: household contact.

Figure 2. Strategy for discovery and validation of the tuberculosis risk signature

(A) Flow diagram for the discovery and validation of the tuberculosis risk signature. The tuberculosis risk signature was obtained by data mining of a whole blood RNA-Seq dataset generated from the ACS training set. The predictive potential of the risk signature was evaluated by rigorous cross-validation. The tuberculosis risk signature was adapted to qRT-PCR, and then the RNA-Seq and qRT-PCR versions of the signature were employed to predict tuberculosis progression using untouched blinded samples from the ACS test set. The qRT-PCR-based tuberculosis risk signature was then employed to predict tuberculosis progression using untouched blinded samples from the South African and Gambian cohorts of GC6-74. (B) Synchronization of the ACS training set in terms of the clinical outcome. To ensure optimal extraction of a tuberculosis risk signature from the ACS training set, the time scale of the RNA-Seq dataset was re-aligned according to tuberculosis diagnosis instead of study enrolment, allowing gene expression differences to be measured before disease diagnosis. Each progressor within the ACS training set is represented by a horizontal bar. The length of the bar represents the number of days between study enrolment and diagnosis with active tuberculosis. During follow-up, each progressor transitioned from an asymptomatic healthy state (green) to pulmonary disease (red). Left side: alignment of PAXgene sample collection (black points) with respect to study enrolment. Right side: alignment of PAXgene sample collection with respect to diagnosis with active tuberculosis, for use in analysis.

Figure 3. The tuberculosis risk signature and validation by prediction of tuberculosis disease progression in the untouched ACS test set and the independent GC6-74 cohorts

(A) Heatmap depicting relative expression level of genes comprising the tuberculosis risk signature in progressors, compared with controls. Higher expression in progressors relative to controls is indicated by intensity of red colour; the average and standard devations (+ and −) are shown. Individual heatmap rows represent distinct splice junctions of individual genes that comprise the signature. Relative expression in each of four 180-day time windows prior to tuberculosis diagnosis is shown. (B) The tuberculosis risk signature was generated by assessing multiple gene-pair interactions; two representative gene-pair signatures are shown. In each scatterplot, the normalized expression of one gene within the pair is plotted against that of the other gene, for all ACS training set data points. The black dots represent control samples, whereas the red dots represent progressor samples. The dotted black line indicates the optimal linear decision boundary for discriminating progressors from controls. (C) Receiver operating characteristic curves (ROCs) depicting the predictive potential of the tuberculosis risk signature for discriminating progressors from controls. Each ROC curve corresponds to a 180-day interval prior to tuberculosis diagnosis. Prediction performance was assessed by 100 four-to-one training-to-test splits of the ACS training set. (D) ROC curves for blind prediction of tuberculosis disease progression on untouched ACS test set samples using the RNA-Seq-based (dotted line) or qRT-PCR-based (solid line) signature. (E) Blind prediction on the combined GC6-74 cohort (blue), South African cohort (purple) or Gambian cohort (green); (F) Stratification of prediction on the overall GC6-74 cohort by time before tuberculosis diagnosis.