Concise whole blood transcriptional signatures for incipient tuberculosis: a systematic review and patient-level pooled meta-analysis

Abstract

Background: Multiple blood transcriptional signatures have been proposed for identification of active and incipient tuberculosis. We aimed to compare the performance of systematically identified candidate signatures for incipient tuberculosis and to benchmark these against WHO targets.

Methods: We did a systematic review and individual participant data meta-analysis. We searched Medline and Embase for candidate whole blood mRNA signatures discovered with the primary objective of diagnosis of active or incipient tuberculosis, compared with controls who were healthy or had latent tuberculosis infection. We tested the performance of eligible signatures in whole blood transcriptomic datasets, in which sampling before tuberculosis diagnosis was done and time to disease was available. Culture-confirmed and clinically or radiologically diagnosed pulmonary or extrapulmonary tuberculosis cases were included. Non-progressor (individuals who remained tuberculosis-free during follow-up) samples with less than 6 months of follow-up from the date of sample collection were excluded, as were participants with prevalent tuberculosis and those who received preventive therapy. Scores were calculated for candidate signatures for each participant in the pooled dataset. Receiver operating characteristic curves, sensitivities, and specificities were examined using prespecified intervals to tuberculosis (<3 months, <6 months, <1 year, and <2 years) from sample collection. This study is registered with PROSPERO, number CRD42019135618.

Results: We tested 17 candidate mRNA signatures in a pooled dataset from four eligible studies comprising 1126 samples. This dataset included 183 samples from 127 incipient tuberculosis cases in South Africa, Ethiopia, The Gambia, and the UK. Eight signatures (comprising 1-25 transcripts) that predominantly reflect interferon and tumour necrosis factor-inducible gene expression, had equivalent diagnostic accuracy for incipient tuberculosis over a 2-year period with areas under the receiver operating characteristic curves ranging from 0·70 (95% CI 0·64-0·76) to 0·77 (0·71-0·82). The sensitivity of all eight signatures declined with increasing disease-free time interval. Using a threshold derived from two SDs above the mean of uninfected controls to prioritise specificity and positive-predictive value, the eight signatures achieved sensitivities of 24·7-39·9% over 24 months and of 47·1-81·0% over 3 months, with corresponding specificities of more than 90%. Based on pre-test probability of 2%, the eight signatures achieved positive-predictive values ranging from 6·8-9·4% over 24 months and 11·2-14·4% over 3 months. When using biomarker thresholds maximising sensitivity and specificity with equal weighting to both, no signature met the minimum WHO target product profile parameters for incipient tuberculosis biomarkers over a 2-year period.

Interpretation: Blood transcriptional biomarkers reflect short-term risk of tuberculosis and only exceed WHO benchmarks if applied to 3-6-month intervals. Serial testing among carefully selected target groups might be required for optimal implementation of these biomarkers.

Funding: Wellcome Trust and National Institute for Health Research.

Figure 1

Genes comprising the eight best performing blood transcriptomic signatures for incipient tuberculosis (A) Matrix showing constituent genes for each signature. (B) Network diagram showing statistically enriched (p<0·05) upstream regulators of the 40 genes, identified by Ingenuity Pathway Analysis. Coloured nodes represent the predicted upstream regulators, grouped by function (red=cytokine, blue=transcription factor, green=other). Black nodes represent the transcriptional biomarkers downstream of these regulators. STAT1, represented by a blue node as a predicted upstream regulator of a number of genes, is also gene target for other upstream regulators. The identity of each node is indicated using Human Genome Organisation nomenclature. The size of the nodes is proportional to the number of downstream biomarkers associated with each regulator and the thickness of the edges is proportional to the –log₁₀ p value for enrichment of each of the upstream regulators.

Figure 2

Scatterplots showing scores of eight best performing transcription signatures for incipient tuberculosis, stratified by interval to disease Dashed horizontal lines indicate thresholds set as standardised scores of two for each signature. Number of samples included for each signature, at each timepoint, indicated in the appendix 1 (p 19). Repeated measures analysis of variance with linear trend method showed p<0·0001 for association of categorical interval to disease with decreasing scores for each of the eight signatures. Scatterplots showing scores of these signatures plotted against days to tuberculosis are shown in the appendix 1 (p 18).

Figure 3

Receiver operating characteristic curves showing diagnostic accuracy of eight best performing transcriptional signatures for incipient tuberculosis Receiver operating characteristic curves shown stratified by months from sample collection to disease. Area under the curve estimates and 95% CIs are shown in the appendix 1 (p 15). Number of samples included for each signature, at each timepoint, indicated in the appendix 1 (p 19).

Figure 4

Diagnostic accuracy of eight best performing transcriptional signatures for incipient tuberculosis shown in receiver operating characteristic space, stratified by months to disease Dashed lines represent positive-predictive values of 5%, 10%, and 15%, based on 2% pre-test probability. Grey shading indicates 95% CIs for each signature. Cutoffs derived from two standard scores above the mean of control population. The number of samples included for each signature, at each timepoint, is indicated in the appendix 1 (p 19). Point estimates and 95% CIs are also shown in the appendix 1 (p 20).