Four-Gene Pan-African Blood Signature Predicts Progression to Tuberculosis

Abstract

Rationale: Contacts of patients with tuberculosis (TB) constitute an important target population for preventive measures because they are at high risk of infection with Mycobacterium tuberculosis and progression to disease.

Objectives: We investigated biosignatures with predictive ability for incident TB.

Methods: In a case–control study nested within the Grand Challenges 6-74 longitudinal HIV-negative African cohort of exposed household contacts, we employed RNA sequencing, PCR, and the pair ratio algorithm in a training/test set approach. Overall, 79 progressors who developed TB between 3 and 24 months after diagnosis of index case and 328 matched nonprogressors who remained healthy during 24 months of follow-up were investigated.

Measurements and Main Results: A four-transcript signature derived from samples in a South African and Gambian training set predicted progression up to two years before onset of disease in blinded test set samples from South Africa, the Gambia, and Ethiopia with little population-associated variability, and it was also validated in an external cohort of South African adolescents with latent M. tuberculosis infection. By contrast, published diagnostic or prognostic TB signatures were predicted in samples from some but not all three countries, indicating site-specific variability. Post hoc meta-analysis identified a single gene pair, C1QC/TRAV27 (complement C1q C-chain / T-cell receptor-α variable gene 27) that would consistently predict TB progression in household contacts from multiple African sites but not in infected adolescents without known recent exposure events.

Conclusions: Collectively, we developed a simple whole blood–based PCR test to predict TB in recently exposed household contacts from diverse African populations. This test has potential for implementation in national TB contact investigation programs.

At a Glance Commentary

Scientific Knowledge on the Subject

Reliable predictors of incident tuberculosis that can guide preventive treatment are urgently needed.

What This Study Adds to the Field

This study describes a small gene expression signature that predicts incident tuberculosis in household contacts across different African sites.

Tuberculosis (TB), caused by infection with Mycobacterium tuberculosis (Mtb) (1, 2), is the leading cause of death caused by a single pathogen globally (3). Before development of symptomatic disease, latent Mtb infection can be detected by measuring immunological sensitization using the tuberculin skin test (TST) and/or IFN-γ release assays (IGRAs) (4). Most infected individuals have effective defense mechanisms to control Mtb (5); only 5 to 10% will progress to TB during their lifetime. Despite this, each year over 10 million new cases of TB caused by either Mtb reactivation or direct transmission are diagnosed, and almost 2 million people die of the disease (3). Although recent Mtb exposure and TST- or IGRA-confirmed conversion are associated with higher risk of TB progression (6), the positive predictive values of these tests are low (1.5% and 2.7%, respectively) (7), falling short of current World Health Organization–supported guidelines for incipient TB (8). Thus, the number of TST+ or IGRA+ individuals requiring treatment to prevent progression to a single incident case of TB is prohibitively high (9).

Factors associated with elevated risk of progression to TB include age, sex, HIV (10, 11), and especially being in recent contact with a patient with active pulmonary TB (12, 13). A biomarker that identifies household contacts (HHCs) who will progress to TB would provide an opportunity to arrest disease progression through targeted prophylactic intervention (14, 15). Such prognostic biomarkers would have the most impact in point-of-care tests for resource-limited settings, such as those in Sub-Saharan Africa. Test performance should not be adversely affected by diversity of ethnic background (16) and circulating Mtb lineage (17), as seen in Africa. A TB risk test must be practical for field application and therefore based on accessible biological samples routinely used in clinical settings, such as peripheral blood (18).

Transcriptional profiling of blood cells has emerged as a powerful platform for discovery of potential TB biomarkers discriminating patients with TB from healthy uninfected and/or individuals with latent Mtb infection (19–24). We previously defined a 16-gene blood transcriptional correlate of risk (COR) signature that predicts risk of progression to TB in HIV-negative South African adolescents with Mtb infection and HHCs from South Africa and the Gambia (25). However, given that this COR signature was developed using a single cohort of South African adolescents with latent Mtb infection, the predictive accuracy for HHCs in diverse African populations may be suboptimal (25). It would also be desirable to reduce the number of transcripts in the signature to facilitate implementation of a point-of-care test. These tests pave the way for simple identification of individuals at highest risk for progression. Some of the results of this study were previously reported in the form of abstracts (26, 27).

Methods

Study Design and Participants

All clinical sites adhered to the Declaration of Helsinki principles and good clinical practice guidelines. Ethical approval was obtained from institutional review boards (see Table E1 in the online supplement). The HHC study included participants from four African sites (South Africa, the Gambia, Ethiopia, and Uganda) as part of the Bill and Melinda Gates Foundation Grand Challenges 6-74 (GC6-74) program (Figure 1 and Table E2). Samples were collected at enrollment/baseline and at 6 and 18 months, with the exception of South Africa, where samples at 6 months were not available. The ACS (Adolescent Cohort Study) was described previously (25, 28) and included IGRA+ and/or TST+ South African adolescents aged 12–18 years with Mtb infection occurring at unspecified times. Adult participants, or legal guardians of participants aged 10–17 years old, provided written or thumbprinted informed consent to participate after careful explanation of the study and potential risks.

Figure 1.

Consolidated Standards of Reporting Trials diagram describing the inclusion and exclusion of participants from the different African cohorts in the Grand Challenges 6-74 (GC6-74, GC6) household contact (HHC) study for training predictive transcriptomic biomarker for tuberculosis (TB) progression: Stellenbosch University in South Africa (SUN), Armauer Hansen Research Institute in Ethiopia (AHRI), Makerere University in Uganda (MAK), Medical Research Council in the Gambia (MRC), and the external validation natural history study of South African adolescents (Adolescent Cohort Study [ACS]). “QC Excluded” refers to samples excluded because they did not meet the minimum quality control requirement for RNA sequencing of an RNA yield greater than or equal to 200 ng and an RNA integrity number greater than or equal to 7. RISK4 = four-gene correlate-of-risk signature.

Sample Processing and RNA Sequencing

PAXgene (PreAnalytiX) blood RNA samples were collected from all participants. Progressors were defined as individuals who developed TB 3–24 months after household contact. Nonprogressor samples were matched to the prediagnosis time points of each progressor by site, sex, age, and recruitment year (see the online supplement). RNA sequencing (RNA-seq) was performed by the Beijing Genomics Institute (Shenzhen, China); additional details regarding processing and quality control are provided in the online supplement. FASTQ files have been deposited in the Gene Expression Omnibus database under accession number GSE94438.

Identification of Predictive Signatures

The hypothesis of the study was that gene expression signatures exist that can predict incident TB in recent HHCs of patients with active TB. Candidate site-specific signatures of risk for TB disease progression and final, simplified qRT-PCR–based candidate signatures were developed using the pair ratios algorithm (see the online supplement), which was previously described (29) and is a variation of the pairwise approach used to discover the ACS COR signature (25). To summarize, the step-by-step procedure for computing the four-gene COR signature (RISK4) scores using sample qRT-PCR measurements was as follows:

1. Measure the cycle thresholds for the four primer-probe sets (TaqMan assays; Applied Biosystems) listed in Table E3.

2. For each of the four pairs of primer-probe sets, compute the difference in raw cycle thresholds, which produces the log-transformed ratio of expression.

3. Compare the measured ratio to ratios in the lookup table for the given pair of transcripts listed in Tables E4–E7. Find the minimal ratio in column 1 of the table that is greater than or equal to the measured ratio.

4. Assign the corresponding score in the second column of the lookup table to the ratio. If the measured ratio is larger than all ratios in column 1 of the lookup table, then assign a score of 1 to the ratio.

5. Compute the average of the scores generated from the set of pairs. If any assay failed on the sample, compute the average score of all ratios, not including the failed assays. The resulting average is the final score for that sample.

Adaptation of Published Diagnostic Signatures to qRT-PCR

The previously published signatures reported by Maertzdorf and colleagues (30) and Sweeney and colleagues (31) were adapted to the qRT-PCR platform, where we refer to them as DIAG4 and DIAG3, respectively. Primer-probe sets were selected for each gene in the respective signatures, and overall scores were computed for each sample as the difference in the mean of the upregulated and downregulated transcripts (Tables E8 and E9). All statistical analyses were performed in R (version 3.1.0) using the R package pROC (32). Details are provided in the online supplement.

Results

We enrolled 4,466 HIV-negative healthy HHCs of 1,098 index TB cases between 2006 and 2010 into the GC6-74 cohorts across four African sites (Figure 1 and Table E2). Samples from Uganda were not available in sufficient quantities for this analysis (Figure 1). TB incidence in HIV-negative healthy HHCs was highest in South Africa and lowest in Ethiopia (Table 1), as defined by TB case classifications A–K in Table E10. Incident cases (progressors) were defined as those who developed TB between 3 and 24 months after index case diagnosis. “Co-incident” cases (i.e., diagnosed with TB within 3 mo of contact with the index case; see Methods above) were not included in the analysis. Prior TB was an exclusion criterion (see the online supplement); thus, progressors likely had their first TB episode during follow-up. The median age of progressors was comparable across the four African sites (Kruskal-Wallis P = 0.92) (Table 1). Median times to progression were 7 months in both South Africa and Uganda, 10.5 months in the Gambia, and 10 months in Ethiopia (Tables 1 and E11A). Progressors, as defined by clinical symptoms, chest and other radiographs consistent with TB, and response to chemotherapy, without microbiological confirmation, comprised 25% (4 of 12) of progressors in Ethiopia, 2% (1 of 43) in South Africa, and 6% (3 of 34) in the Gambia (TB classification K in Tables E10 and E11A).

Table 1.

Baseline Demographic Characteristics of Progressors and Matched Nonprogressor Control Subjects in the Four African Household Contact Cohorts

	South Africa	The Gambia	Ethiopia	Uganda
HIV-negative HHC, n	1,197	1,948	818	499
Progressors, n	43	34	12	11
Incidence, %	3.6	1.7	1.5	2.2
Age, yr, median (IQR)
Progressors	25 (18–41)	22.5 (20–30.75)	23 (19.75–27)	23 (18–36)
Nonprogressors	24 (18–38)	24 (18–30.25)	25 (20–35)	27 (19–38.75)
Male sex, %
Progressors	41.9	44.1	33.3	54.5
Nonprogressors	40.7	44.1	35.4	54.5
Time to TB, mo, median (IQR)
Progressors	7 (5–17)	10.5 (7–18.75)	10 (6.5–15)	7 (5–11)

Definition of abbreviations: HHC = household contacts; IQR = interquartile range; TB = tuberculosis.

RISK4, a Four-Gene COR Signature Predicts TB Progression in HHCs

We divided South African and Gambian HHC cohorts into training and test sets, whereas the entire Ethiopian cohort was assigned to the test set owing to its small sample size (Figure 1 and Tables E11A and E11B). We used the South African and Gambian training sets to construct site-specific signatures of TB risk using RNA-seq transcriptomes and the pair ratio approach, which uses ratios of transcripts regulated in opposite directions during TB progression (see online supplement text and Tables E12 and E13). Leave-one-out cross-validation analysis (applied to all samples from specific individuals) indicated strong potential for predicting TB progression in both cohorts (South Africa [Figure 2A], area under the receiver operating characteristic curve [AUC], 0.86 [95% confidence interval {CI}, 0.79–0.94], P = 8.4 × 10⁻¹⁰; the Gambia [Figure 2B], AUC, 0.77 [95% CI, 0.66–0.88], P = 2.5 × 10⁻¹⁰). By applying the algorithm to the South African and Gambian cohorts, we generated two distinct, site-specific risk signatures (Figures 2C and 2D). When measured by qRT-PCR using primer-probe sets that corresponded to the exons, predictive accuracy was maintained (Figure E1). The Gambia signature failed to validate in samples from South Africa (Figure 2A) (AUC, 0.59 [95% CI, 0.46–0.73]; P = 0.061), whereas the South Africa signature weakly validated in Gambian samples (Figure 2B) (AUC, 0.66 [95% CI, 0.54–0.76]; P = 8.8 × 10⁻³), suggesting site-specific progression signatures in South Africa and the Gambia.

Figure 2.

Site-specific feature selection and translation to RT-PCR. (A) Receiver operating characteristic curve for leave-one-out cross-validation (LOOCV) of South Africa (blue; area under the receiver operating characteristic curve [AUC], 0.86 [95% confidence interval (CI), 0.79–0.94]; P = 8.4 × 10⁻¹⁰) versus the Gambia–trained prospective signature (red; AUC, 0.59 [95% CI, 0.46–0.73]; P  =  0.06) in the South African training set (samples listed in Tables E11A and E11B). (B) Receiver operating characteristic curves for LOOCV of the Gambia (blue; AUC, 0.77 [95% CI, 0.66–0.88]; P = 2.5 × 10⁻⁵) versus South Africa prospective signature (red; AUC, 0.66 [95% CI, 0.54–0.77]; P = 8.8 × 10⁻³) in the Gambia training set containing 26 progressor and 76 nonprogressor samples. (C and D) Heat maps showing the expression of each splice junction in (C) the South Africa and (D) the Gambia signatures in nonprogressors (left columns), progressors 1–2 years before diagnosis (middle columns), and progressors 0–1 year before diagnosis (right columns). For each group of samples, the central column is the mean fold expression change versus nonprogressors, whereas left/right columns in each group correspond to mean ± SEM. Each row corresponds to a splice junction, and genes with multiple rows are represented by multiple splice junctions in the signature.

The poor cross-prediction of South African and Gambian signatures motivated explicit development of a multicohort signature using a training set that combined samples from both sites. We pooled PCR-based transcript pairs that comprised all of the South Africa (38 transcripts), and the Gambia (35 transcripts) signatures (Figures 2C and 2D and Tables E12 and E13) to identify top transcript pairs that were significantly predictive of TB progression in both cohorts, and we successively added the next best pair to the ensemble and reassessed the predictive power at each stage until the gain in predictive performance reached a plateau (see online supplement text and Table E14). This resulted in the RISK4 signature comprising four unique genes: GAS6 (growth arrest–specific 6) and SEPT4 (septin 4), which were upregulated, and CD1C (cluster of differentiation 1C) and BLK (B lymphocyte kinase), which were downregulated, in progressors versus matched control subjects (Figure 3A).

Figure 3.

Validation of a multicohort four-gene signature (RISK4) derived from the South African and Gambian training sets. (A) Expression ratio of gene pairs in the RISK4 signature in the South African (top) and Gambian (bottom) training sets: nonprogressors (left columns), progressors 1–2 years before diagnosis (middle columns), and progressors 0–1 (right columns) year before diagnosis. In each group, the central column is the mean fold expression over nonprogressors, whereas left/right columns in each group correspond to mean ± SEM. (B) Receiver operating characteristic (ROC) curves for blind predictions of RISK4 on test set samples of all sites (black; area under the receiver operating characteristic curve [AUC], 0.67 [95% confidence interval (CI), 0.57–0.77]; P = 2.6 × 10⁻⁴), South Africa (red; AUC, 0.72 [95% CI, 0.53–0.92]; P = 6.3 × 10⁻³), the Gambia (blue; AUC, 0.72 [95% CI, 0.55–0.88]; P = 5.4 × 10⁻³), and Ethiopia (green; AUC, 0.67 [95% CI, 0.5–0.83]; P = 0.02). (C) Performance of RISK4 signature in test set samples obtained within 1 year of diagnosis (red; AUC, 0.66 [95% CI, 0.55–0.78], P = 2 × 10⁻³; 30 progressor samples, 201 nonprogressor samples) or 1–2 years before diagnosis (blue; AUC, 0.69 [95% CI, 0.51–0.86]; P = 0.02; 12 progressor samples, 201 nonprogressor samples). (D) ROC curve of RISK4 on all baseline test set samples (AUC, 0.69 [95% CI, 0.52–0.86]; P = 5 × 10⁻³). (E) ROC curve blind prediction of RISK4 in South African adolescents with latent Mycobacterium tuberculosis (M.tb) infection (AUC, 0.69 [95% CI, 0.62–0.76]; P = 3.4 × 10⁻⁷). BLK = B lymphocyte kinase; CD1C = cluster of differentiation 1C; GAS6 = growth arrest–specific 6; SEPT4 = septin 4; TB = tuberculosis.

Having developed a multisite PCR-based signature of risk, we validated it by blind prediction of TB progression on the multicohort test sets from South Africa, the Gambia, and Ethiopia (Figure 1). The RISK4 signature significantly predicted progression in the entire combined test set (AUC, 0.67 [95% CI, 0.57–0.77]; P = 2.6 × 10⁻⁴) (Figure 3B) and on each individual site (South Africa, the Gambia, and Ethiopia with AUCs of 0.66–0.72; P < 0.03) (Figure 3B). Surprisingly, performance of the signature on combined test set samples within one year of TB diagnosis (AUC, 0.66 [95% CI, 0.55–0.78]; P = 1.9 × 10⁻³) (Figure 3C) was comparable to samples collected more than one year before diagnosis (AUC, 0.69 [95% CI, 0.51–0.86]; P = 0.015). Because deploying a risk signature in a screen-and-treat strategy in TB HHCs would most likely entail testing early after exposure, we assessed the predictive performance of RISK4 on samples from HHCs collected within two months of diagnosis of the index case. RISK4 also was validated in this setting (Figure 3D) (AUC, 0.69 [95% CI, 0.52–0.86]; P = 4.8 × 10⁻³). Finally, to test the robustness of RISK4, we performed blinded predictions on samples from an external cohort of IGRA+/TST+ South African adolescents (the ACS cohort), in whom the time of TB exposure was unknown (25). RISK4 also significantly predicted risk of TB progression in this cohort (Figure 3E) (AUC, 0.69 [95% CI, 0.62–0.76]; P = 3.4 × 10⁻⁷).

Comparison of RISK4 with Published Diagnostic TB Signatures

To benchmark the predictive performance of the RISK4 signature, we compared it with qRT-PCR–based versions of three published transcriptional signatures for TB diagnosis: “DIAG3,” the three-gene diagnostic signature reported by Sweeney and colleagues (31); “DIAG4,” the four-gene diagnostic signature reported by Maertzdorf and colleagues (30); and our own previously reported 16-gene COR signature for TB progression (“ACS COR”; Zak and colleagues [25]). Because HHC training set samples were used to discover the RISK4 signature, we compared the performance of PCR-adapted published signatures with that of RISK4 in the HHC test set only. The three signatures predicted TB progression in the combined test set with accuracy comparable to that of RISK4 (Figure 4A) (AUCs, 0.64–0.68; P < 3 × 10⁻³). When we compared the predictive accuracy of RISK4 with each of the three PCR-adapted signatures, the AUCs were not statistically different (Table E15). However, unlike RISK4 (Figure 3B), the three other signatures did not validate on all sites when evaluated individually (Figures 4B–4D), suggesting that RISK4 represents a more generalizable prognostic signature.

Figure 4.

Comparison of PCR-adapted signatures: RISK4, ACS 16-gene correlate of risk (ACS COR), and tuberculosis diagnostic signatures (DIAG3 and DIAG4). (A) Receiver operating characteristic curves for blind predictions of RISK4 (black; area under the receiver operating characteristic curve [AUC], 0.67 [95% confidence interval (CI), 0.57–0.77]; P = 3 × 10⁻⁴), DIAG3 (red; AUC, 0.68 [95% CI, 0.59–0.78]; P = 8 × 10⁻⁵), DIAG4 (blue; AUC, 0.64 [95% CI, 0.53–0.74]; P = 3 × 10⁻³), and ACS COR (green; AUC, 0.66 [95% CI, 0.55–0.76]; P = 6 × 10⁻⁴) in all test set samples. (B–D) Blind prediction of PCR-adapted signatures: (B) DIAG3 (South Africa AUC, 0.66 [95% CI, 0.47–0.84]; the Gambia AUC, 0.61 [95% CI, 0.45–0.77]; and Ethiopia AUC, 0.78 [95% CI, 0.64–0.92]), (C) DIAG4 (South Africa AUC, 0.77 [95% CI, 0.62–0.91]; the Gambia AUC, 0.52 [95% CI, 0.33–0.71]; and Ethiopia AUC, 0.64 [95% CI, 0.46–0.83]), and (D) ACS COR (South Africa AUC, 0.82 [95% CI, 0.71–0.92]; the Gambia AUC, 0.56 [95% CI, 0.37–0.75]; and Ethiopia AUC, 0.6 [95% CI, 0.41–0.79]). South Africa, the Gambia, and Ethiopia AUCs are depicted in red, blue, and green, respectively. ACS = Adolescent Study Cohort; RISK4 = four-gene correlate-of-risk signature.

After unblinding the South African, Gambian, and Ethiopian test sets, we interrogated whether the RISK4 signature could be reduced to a single pair of transcripts without a loss of predictive accuracy. We applied each of the four ratios in the RISK4 signature to each of the test set cohorts individually, and we compared the performance with that of the entire RISK4 signature (Table E16). The ratio between the SEPT4 and BLK primers reproduced the performance of the RISK4 signature on all three test set cohorts, demonstrating feasibility of a highly simplified two-gene host RNA-based signature for identifying HHCs at greatest risk of progressing to active TB.

Meta-Analysis Identifies Gene Pairs That Predict TB Progression across Africa

Overall, predictions for TB progression were the least accurate for the Ethiopian cohort, which was not used to develop the initial RISK4 signature (Figures 1, 3, and 4). To determine whether accuracy could be further improved for a signature performing well at all sites, we performed a meta-analysis of RNA-seq profiles for the combined training and test datasets from all of our three cohorts. This post hoc analysis was performed after unblinding of the test set and was focused on identifying better predictive gene pairs, given that the single transcript pair SEPT4/BLK performed equivalently to the RISK4 signature (Table E16).

We combined RNA-seq data from all training and test cohorts, thus merging the three independent cohorts from South Africa, the Gambia, and Ethiopia. Pairs of upregulated and downregulated transcripts were formed from all transcripts that individually discriminated progressors from control subjects in at least one cohort (Tables E17 and E18; Wilcoxon false discovery rate <0.05 in at least one of the three cohorts). Each pair was then analyzed on each of the three sites. We identified nine transcript pairs that discriminated progressors from control subjects with AUC greater than 0.75 on all three sites (Table E19). The optimal pair consisted of C1QC (complement C1q C-chain) (upregulated) and TRAV27 (T-cell receptor-α variable gene 27) (downregulated) and achieved AUC greater than 0.76 on all three sites. We performed logistic regression analysis to determine whether the remaining eight pairs (Table E20 and Supplementary Methods in the online supplement) captured information about TB progression that was redundant or complementary to the signals detected by C1QC/TRAV27. The ratio between ANKRD22 (ankyrin repeat domain 22; upregulated with TB progression) and OSBPL10 (oxysterol binding protein like 10; downregulated with progression) led to significantly increased discrimination between progressors and control subjects when it was combined with the C1QC/TRAV27 ratio in HHC cohorts (Figures 5A–5C), increasing the AUC on all three HHC cohorts individually to greater than 0.79 (Table E21). Thus, the ratios C1QC/TRAV27 and ANRKD22/OSBPL10 capture distinct aspects of TB progression signals in HHC that are shared across three distinct African sites.

Figure 5.

Gene pairs to predict tuberculosis progression in African cohorts. Ratios of C1QC/TRAV27 and ANKRD22/OBSPL10 plotted on samples from (A) South Africa, (B) the Gambia, and (C) Ethiopia, together with an optimal discriminant (dashed line; optimizes sum of sensitivity and specificity) separating progressors (orange) from nonprogressors (blue). In each cohort, the two pairs provide complementary information. P values correspond to chi-square complementation analysis in Table E16. (D) Receiver operating characteristic curves showing the ability of the Grand Challenges 6-74 (GC6)-trained C1QC/TRAV27 (solid line; area under the receiver operating characteristic curve [AUC], 0.57 [95% confidence interval (CI), 0.49–0.64]; P = 0.042), ANKRD22/OBSPL10 (dashed line; AUC, 0.75 [95% CI, 0.68–0.81]; P = 2.86 × 10⁻¹¹), and a linear combination of C1QC/TRAV27 and ANKRD22/OBSPL10 (dotted line; AUC, 0.69 [95% CI, 0.61–0.76]; P = 4.3 × 10⁻⁷) models to predict tuberculosis disease progression in the South African Adolescent Cohort Study (ACS) cohort. (E) Receiver operating characteristic curves showing the ability of the GC6-trained C1QC/TRAV27 (solid line; AUC, 0.77 [95% CI, 0.71–0.83]; P = 4.5 × 10⁻¹⁶), ANKRD22/OBSPL10 (dashed line; AUC, 0.76 [95% CI, 0.7–0.82]; P = 3.7 × 10⁻¹⁵), and a linear combination of C1QC/TRAV27 and ANKRD22/OBSPL10 (dotted line; AUC, 0.79 [95% CI, 0.74–0.85]; P = 6.1 × 10⁻¹⁹) models to predict TB disease progression in the GC6 household contact (HHC) cohort. (F and G) Logarithmic ratios of expression (mean ± 95% CI) for (F) ANKRD22/OBSPL10 and (G) C1QC/TRAV27 are plotted as a function of time to diagnosis for both GC6-74 (blue) and ACS (red) progressor samples. Comparison of C1QC/TRAV27 expression at 19–24 months before diagnosis between the GC6-74 HHC and ACS cohorts was statistically significantly different (P = 3 × 10⁻³) using the Mann-Whitney U test. ANKRD22 = ankyrin repeat domain 22; C1QC = complement C1q C-chain; OBSPL10 = oxysterol binding protein like 10; TB = tuberculosis; TRAV27 = T-cell receptor-α variable gene 27.

To determine whether the C1QC/TRAV27 and ANKRD22/OSBPL10 signatures captured universal aspects of TB progression rather than HHC-associated biology, we evaluated them using data from the cohort of IGRA+ TST+ South African adolescents (25). The ANKRD22/OBSPL10 ratio strongly predicted TB progression among the adolescents with Mtb infection (Figure 5D) (AUC, 0.75 [95% CI, 0.68–0.81]; P = 2.86 × 10⁻¹¹), but the C1QC/TRAV27 ratio was poorly predictive in the adolescent cohort (Figure 5D) (AUC, 0.57 [95% CI, 0.49–0.64]; P = 0.042). In contrast to the HHCs, combining the two ratios did not lead to improved discrimination of progressors and control subjects in the adolescent cohort (AUC, 0.69 [95% CI, 0.61–0.76]) (Figures 5D and E2A). To further understand the disparity in the predictive performance for the HHC cohorts and the adolescents with Mtb infection, we evaluated the longitudinal behavior of the transcript ratios for progressor samples in the HHC and adolescent cohorts (Figures 5F and 5G). The ANKRD22/OSBPL10 pair exhibited similar behavior in the HHC and ACS cohorts, with a steady upregulation during progression and no significant difference between GC6-74 and adolescent participants in any 6-month time window preceding TB diagnosis (Figure 5F). In contrast, the C1QC/TRAV27 ratio was significantly higher in HHC progressors than in adolescents with Mtb infection 19–24 months before TB diagnosis (P = 3 × 10⁻³) (Figure 5G). Importantly, samples from HHC progressors were collected mostly at enrollment, immediately after exposure to the respective TB index cases, thus possibly representing a signature of recent Mtb exposure.

Discussion

We identified and validated a simple PCR-based transcriptomic signature, “RISK4,” to predict the risk of progression to active TB disease in diverse African cohorts of recently exposed HHCs of index TB cases. This four-gene signature predicted risk of progression with similar accuracy in four cohorts from three sub-Saharan African populations with heterogeneous genetic backgrounds, TB epidemiology, and circulating Mtb strains (33). Importantly, RISK4 exhibited consistent predictive performance in all test set cohorts, whereas previously reported signatures (25, 30, 31) exhibited cohort-specific variability in performance. We previously reported that the ACS COR signature was validated in the combined South African and Gambian HHC cohorts (25). In the present analysis, the stochastic partitioning of HHC samples into training and test sets resulted in ACS COR performance measurements different from the previously reported results.

The signatures we report represent significant and translational improvements over currently used biomarkers for predicting risk of TB, such as IGRAs or the TST (14, 15). Recent estimates suggest the TB incidence rates in South Africa and the Gambia to be 0.8% (3) and 0.3% (34), respectively. However, IGRA+ and TST+ prevalence can reach up to 50% in the Gambia and 80% in South Africa (3), and although IGRAs and the TST have a high (approximately 80%) sensitivity for Mtb infection, they have poor positive predictive values of 2.7% and 1.5%, respectively, for TB progression. Therefore, dozens of individuals would require prophylactic treatment to prevent progression to TB in a single individual (35, 36). The World Health Organization recently published guidelines for an incipient TB target product profile to predict TB progression (8) in order to ensure that individuals at high risk of TB progression are not falsely excluded (7, 18) but instead are referred for additional investigation for TB or offered prophylactic treatment (37). At sensitivities of 81%, 71%, 62%, and 50%, the RISK4 signature achieved specificities of 34%, 52%, 63%, and 77%, respectively, in healthy asymptomatic individuals by selection of different thresholds (Table E22). Although RISK4 has a poor positive predictive value of 3%, similar to IGRA or TST, owing to its lower sensitivity at higher specificity thresholds (Table E23), it importantly has lower positivity rates in the target population. To achieve test performance similar to IGRAs (between 70 and 80% sensitivity and the number to harm to prevent one case of approximately 85), the RISK4 threshold would identify between 38 and 54% of HHCs for preventive measures, compared with 78% for IGRA (Table E22).

There are several limitations of this study, including the small sample size (only 100 of the 4,460 HHCs progressed to active TB). Furthermore, although the intended application is for a trans-Africa test, we could sample from only three regions. The poor RISK4 performance in Ethiopia calls for larger multicenter studies, particularly of underrepresented populations throughout Africa and other TB endemic areas. Although we defined two-transcript signatures that have broader application, the sample size was not adequate to have unblinded validation sets or to ensure that transcript pairs were not selected by chance during our post hoc analyses. The test is based on blood samples, which are easily and routinely obtained in laboratory diagnostics. However, the test still requires translation into field-friendly instrumentation to bypass the multistep processing involved in RNA extraction and RT-PCR. Encouragingly, recent advances in point-of-care PCR technologies offer promise for developing rapid diagnostics. We envision platforms where blood from a finger prick can be translated using field-friendly handheld PCR instruments into interpretable scores. Field staff would then triage near-patient contacts into low-risk and higher-risk groups for further assessment and potential treatment for subclinical or active TB disease (8). One advantage of the calculation of RISK4 or the two-transcript scores is the pairs ratio structure, which eliminates the need for housekeepers or other standardization methods. As a proof of principle, a clinical trial (registered with www.clinicaltrials.gov [NCT02735590]) stratifying participants by ACS COR positivity (25) will provide real-world data on the efficacy of a strategy that screens South African adults with the COR signature and provides preventive therapy to those who are COR positive. Evaluation of the costs and benefits of such strategies should be carefully evaluated in future implementation studies of RISK4 (36). One benefit might be a strong motivation for both healthcare personnel and patients alike to initiate and complete preventive treatment in the face of a positive COR test. The potential need for repeat test performance also needs to be evaluated.

We identified several transcript pairs that recapitulated the predictive performance of the RISK4 signature and reflected complementary signals in predicting risk of TB progression. RISK4 comprises GAS6 and SEPT4 (upregulated) and BLK and CD1C (downregulated). Interestingly, CD1C and GAS6 (activating ligand of AXL) are expressed in two distinct dendritic cell (DC) subsets where GAS6 expression defines a newly characterized AXL (AXL receptor tyrosine kinase)⁺SIGLEC6 (sialic acid binding Ig like lectin 6)⁺ DC population (38), suggesting that TB pathogenesis may involve redistribution of circulating DC subsets. The SEPT4 protein has antiapoptotic functions, and its deletion improved wound healing in mice (39), suggesting a possible association with lung healing during TB progression. BLK is a B-cell receptor kinase, and its downregulation is consistent with reduced B-cell proportions in blood during TB (19, 40). The most generalizable pair defined in our meta-analysis showed upregulation of C1QC and downregulation of TRVAV27. Interestingly, complement pathway genes are markedly upregulated after Mtb infection of nonhuman primates (41), consistent with the upregulation of C1QC/TRAV27 at baseline in the HHCs. Complement activation is also observed early during human progression to TB (40), whereas C1q is downregulated early after starting TB treatment (22), suggesting that C1q may be a proxy of early TB pathology. Conversely, downregulation of TRAV27 and several other T-cell genes (Table E18) is likely associated with the overall decrease in peripheral T-cell frequencies and their associated gene expression modules during TB progression, potentially due to migration of T cells to the disease site (19, 21, 40). The simple C1QC/TRAV27 signal may thus be a readout of TB risk after initial exposure to an individual with pulmonary TB, which is more synchronized in an HHC study design, even though prior exposure to Mtb cannot be ruled out in our GC6-74 study and individuals with progression to TB disease within the first three months of the observation period were excluded from the analysis. This may explain why C1QC/TRAV27 signal was less predictive in the natural history cohort of adolescents with Mtb infection, where the time of Mtb exposure was unspecified. Early clinical studies suggest that recent exposure to Mtb, indicated by TST conversion, can correlate with symptoms consistent with febrile disease, such as fever and erythema nodosum (42, 43), which are markers of systemic inflammation. C1QC/TRAV27 may reflect this inflammatory response induced by failed containment of Mtb after recent exposure.

Overall, our study identifies and validates a simple PCR-based test with accessible blood samples that predicts TB in heterogeneous African populations with intermediate to high TB burdens (14, 15). Such a test can potentially be developed into a screening test for risk of progression during TB contact investigation implemented by national public health structures (13, 35, 36). The next steps include assessment of the performance of RISK4 and the two-transcript C1QC/TRAV27 signature in other settings, including non-African populations, and determination of the feasibility of developing a near-patient test for targeted intervention.