The ACS 16-gene signature (Zak et al., 2016)!

The adolescent cohort study was conducted in South African adolescents who were followed-up for two years to determine the prevalence and incidence of Mtb infection, TB disease and risk factors for developing TB (Mahomed et al. 2013). Healthy adolescents (n=6,363) were recruited at 11 high schools from the Worcester area outside Cape Town. Participants were allocated to two groups with different follow- up methods. Those in the active follow-up group were assessed every 3 months and had blood samples drawn every six months for QFT, PAXgene, plasma and PBMC

collection. Those in the passive follow-up group were assessed and had blood collected at baseline and two years after enrolment. The surveillance team accessed clinical records to identify TB cases diagnosed in the public health sector for both groups. Whole blood was collected in PAXgene tubes from each study participant at the follow-up time points in addition to clinical data. Microbiologically confirmed active TB disease (two consecutive sputum smear tests or one sputum culture positive test) was diagnosed during follow-up in 87 adolescents in total, and 46 qualified as progressors for our biomarker study (Zak et al. 2016). The progressors were matched to 107 healthy Mtb-infected controls by ethnicity, school, gender, age and prior episodes of TB disease at enrolment, and were referred to as non- progressors. The progressors and non-progressors were split into a training set for signature discovery and a test set for blind validation at a ratio of 3:1. All participants included in this analysis were Mtb-infected (QFT and/or TST positive) at baseline/enrolment.

RNA-sequencing was used to measure transcript expression in this cohort by comparing the progressors and non-progressors. Unlike microarrays, RNA- sequencing does not rely on pre-defined probe-based sequences, thereby allowing less biased identification of novel transcripts which were previously unknown (Blankley et al. 2014). Support vector machines (SVMs), a generalised linear classifier, was used to generate the signature in this population (Zak et al. 2016).

This algorithm “maps data points onto a multidimensional space in such a way that it is possible to calculate a hyperplane (a multidimensional plane), which separates the classes of samples” (Maertzdorf et al., 2015). Multiple pair-wise ensemble models (representing two transcripts each) that could predict TB disease risk based on

splice junction counts measured by RNA-seq were developed. A transcript signature consisting of 16 differentially expressed interferon response genes representing 47 splice junctions plus 10 reference splice junctions could predict progression to TB disease in the progressors (Figure 1). The signature, henceforth referred to as the ACS 16-gene signature, was adapted to microfluidic quantitative reverse transcription polymerase chain reaction (qRT-PCR) to reproduce the findings using a different measurement platform, allow more high-throughput and cheaper measurement, and to prepare for translation to near-patient testing. Splice junctions from the RNA-seq data were matched to TaqMan primer-probe sets for this and the SVM models were re-parameterized to the PCR data.

Figure 1: Network visualisation of the transcript pairs in the ACS 16-gene signature (adapted from Zak et al., 2016). The nodes represent splice junctions and the colours indicated genes represented by the splice junction.

ETV7 FCGR1A FCGR1B GBP1 GBP2 GBP5 SCARF1 TAP1SERPING1 STAT1 Refer to table 1

The PCR-based signature model was fitted to the Ct value data in the training set.

The qRT-PCR signature consisted of 47 genes of interest (GOI) primer-probes, representing 247 gene-pairs and 10 reference primer-probes for standardization (to calculate delta Ct values, Table 1). The signature is comprised of the following genes: FCGR1B, FCGR1A, STAT1, GBP1, GBP2, GBP4, GBP5, SERPING1, ETV7, BATF2, SCARF1, APOL1, TAP1, TRAFD1, ANKRD22 and SEPT4. Signature scores are calculated by summing the product of normalised abundances (delta Ct values) of two gene products (a gene-pair) with their coefficients and a coefficient of the gene-pair. Custom scripts, written in R, were applied on the results generated from the microfluidic qRT-PCR to calculate a single signature score per sample, expressed as a percentage. The signature scores are directly correlated with gene expression, thus an increasing signature score translates into a high gene expression. A higher score thus implies higher gene expression. The various transcripts could contribute differently to different sample types and signature scores would remain high in all these samples. The proportion of transcript-pairs voting a case relative to the total number of voting pairs is what determines the signature scores.

In the training set, the signature score in progressors increased closer to TB diagnosis. The signature could predict active TB disease up to a year before TB diagnosis with a sensitivity of 71.2% (ROC AUC, 0.79, 95% CI: 0.76 to 0.82) and 62.9% (ROC AUC, 0.77, 95% CI: 0.75 to 0.79) at 1-180 days and 181-360 days before TB diagnosis, respectively. RT-PCR data from the test set cohort were used to perform blind predictions, which demonstrated that the signature could predict

active TB in the test cohort up to a year before diagnosis with a sensitivity of 66%

and specificity of 81%.

For external validation, the signature was applied to blinded samples from an independent cohort of household contacts of active TB disease from the Grand Challenges 6-74 (GC6-74) cohort. Samples from 73 progressors and 301 non- progressors from South Africa and Gambia were evaluated. The signature could predict active TB in this cohort up to a year before diagnosis with a sensitivity and specificity of 54% and 83%, respectively, with a ROC AUC of 0.72 (95% CI please).

In addition, the signature was reparameterised to publicly available microarray datasets (Anderson et al., 2014; Berry et al., 2010; Bloom et al., 2012, 2013; Kaforou et al., 2013) to determine the diagnostic performance of the ACS 16-gene signature for discriminating active TB from uninfected or Mtb-infected controls or controls with other respiratory diseases. The signature performed well and showed excellent promise as a diagnostic and prognostic biomarker for TB. A more practical version of this signature (ACS 11-gene, described below) consisting of 48 transcripts was tested in this thesis.

1.9.2 The ACS 11-gene signature: a more practical PCR-based signature