Prediction of anti-tuberculosis treatment duration based on a 22-gene transcriptomic model

Abstract

Background: The World Health Organization recommends standardised treatment durations for patients with tuberculosis (TB). We identified and validated a host-RNA signature as a biomarker for individualised therapy durations for patients with drug-susceptible (DS)- and multidrug-resistant (MDR)-TB.

Methods: Adult patients with pulmonary TB were prospectively enrolled into five independent cohorts in Germany and Romania. Clinical and microbiological data and whole blood for RNA transcriptomic analysis were collected at pre-defined time points throughout therapy. Treatment outcomes were ascertained by TBnet criteria (6-month culture status/1-year follow-up). A whole-blood RNA therapy-end model was developed in a multistep process involving a machine-learning algorithm to identify hypothetical individual end-of-treatment time points.

Results: 50 patients with DS-TB and 30 patients with MDR-TB were recruited in the German identification cohorts (DS-GIC and MDR-GIC, respectively); 28 patients with DS-TB and 32 patients with MDR-TB in the German validation cohorts (DS-GVC and MDR-GVC, respectively); and 52 patients with MDR-TB in the Romanian validation cohort (MDR-RVC). A 22-gene RNA model (TB22) that defined cure-associated end-of-therapy time points was derived from the DS- and MDR-GIC data. The TB22 model was superior to other published signatures to accurately predict clinical outcomes for patients in the DS-GVC (area under the curve 0.94, 95% CI 0.9-0.98) and suggests that cure may be achieved with shorter treatment durations for TB patients in the MDR-GIC (mean reduction 218.0 days, 34.2%; p<0.001), the MDR-GVC (mean reduction 211.0 days, 32.9%; p<0.001) and the MDR-RVC (mean reduction of 161.0 days, 23.4%; p=0.001).

Conclusion: Biomarker-guided management may substantially shorten the duration of therapy for many patients with MDR-TB.

Trial registration: ClinicalTrials.gov NCT02597621.

FIGURE 1

Multistep development of the therapy-end model for tuberculosis (TB) treatment. Simplified flow chart showing the multistep approach of transcriptomic and clinical data analysis to develop the therapy-end model that identifies the optimal time point to stop anti-TB therapy. a) Development of therapy outcome score (TOS) showing the volcano plot representing differentially expressed genes in healthy controls versus therapy-naïve drug-susceptible (DS)- and multidrug-resistant (MDR)-TB patients from the German identification cohorts (GIC). Genes that were significantly up- or down-regulated (significant two-fold or greater change after Benjamini-Hochberg correction) form the basis for the TOS development. b) Therapy progression score (TPS) development depicting penalising regression coefficient adjustment (y-axis) and the explained deviation as a function of log-λ (x-axis) for variable selection to identify genes that predict the remaining days of therapy that has been conducted in reality in all sample measurements from DS- and MDR-GIC TB patients. Each line represents one gene of interest and the genes shown in the plot were pre-selected by the initial lasso regression step. The initial data selection was carried out on the entire dataset with 44000 gene targets. c) End-of-therapy (EOT) list showing penalising regression coefficient adjustment (y-axis) and the explained deviation as a function of log-λ (x-axis) for variable selection to identify genes that classify between sample measurements in DS-GIC TB patients under therapy versus time points at the end of relapse-free therapy in DS-GIC TB patients. Each line represents one gene of interest and the gene targets shown in the plot were pre-selected by the initial lasso regression to reduce the number of genes of interest. d) Therapy-end model (TB22). Implementing the gene scores (TOS and TPS) and the EOT list into a machine-learning algorithm model (random forest), a final simplified therapy-end model for the calculation of EOT time points was developed via a generalised linear model. The initial therapy-end model evaluation was carried out on data from DS-GIC TB patients. The receiver operating characteristic curve shows the therapy-end model’s classification accuracy in the independent dataset of DS German validation cohort (GVC) TB patients (area under the curve (AUC) 0.937, 95% CI 0.899–0.976). The therapy-end model was further applied to patients with MDR-TB from the GIC, GVC and from the Romanian validation cohort.

FIGURE 2

Receiver operating characteristic curve for the therapy-end model classification in drug-susceptible TB patients from the German validation cohort. Receiver operating characteristic curve analysis showing the performance of the therapy-end model (TB22) and the five published signatures/scores with the highest areas under the curve (AUC) (table 2) for the identification of optimal end-of-therapy time points when compared to clinical therapy-end time points in drug-susceptible German validation cohort patients. Therapy-end model: AUC 0.937 (95% CI 0.899–0.976); Kaforou et al. [22], 27 genes: AUC 0.81 (95% CI 0.74–0.89); Thompson et al. [13], nine genes: AUC 0.81 (95% CI 0.71–0.89); Kaforou et al. [22], 53 genes: AUC 0.79 (95% CI 0.70–0.88), Laux da Costa et al. [23], three genes: AUC 0.79 (95% CI 0.70–0.88).

FIGURE 3

Therapy-end model (TB22) scores for individual end-of-therapy time points over time for the five cohorts of patients with tuberculosis (TB). Scores for end of therapy by the TB22 over the time of anti-TB treatment for the five cohorts of drug-susceptible (DS)-TB and multidrug-resistant (MDR)-TB patients of a) German identification cohort (GIC), b) German validation cohort (GVC) and c) Romanian validation cohort (RVC) following the TB22. 6 months of therapy is the common time point of therapy end (TE) in DS-TB; 20 months of therapy represents the usual time point for TE in MDR-TB. Data are presented as smoothed mean lines based on shown calculation and 95% CI. Cut-off: TB22 threshold (p≥0.5) for relapse-free end of therapy.

FIGURE 4

Therapy-end model (TB22) scores for the different cohorts of patients with tuberculosis (TB) at relevant clinical time points. Violin plots with TB22 scores for drug susceptible (DS)-TB patients from the German identification cohort (GIC) and the German validation cohort (GVC), and for multidrug-resistant (MDR)-TB patients from the GIC, the GVC and the Romanian validation cohort (RVC) at different time points during therapy: a) at therapy start; b) after 14 days of therapy; c) at smear conversion; d) at culture conversion; e) at 6 months of therapy; and f) at individual therapy ends. Cut-off: TB22 threshold (p≥0.5) for relapse-free end of therapy; all values below the threshold indicate an ongoing need for anti-TB therapy. There were no smear or culture conversion data for the MDR-RVC.

FIGURE 5

Comparison of the therapy-end model (TB22) with published RNA signatures and scores to identify end-of-therapy (TE) time points in drug-susceptible (DS) and multidrug-resistant (MDR)-tuberculosis (TB) patients from the German validation cohort (GVC). The therapy-end model and the other signatures were used to derive TE time points. Results are depicted for both patient groups over the course of treatment where probabilities ≥0.5 would be associated with successful therapy end. a) TB22; b) Anderson et al. [20], 43 genes; c) Berry et al. [21], 87 genes; d) Kaforou et al. [22], 27 genes; e) Kaforou et al. [22], 44 genes; f) Kaforou et al. [22], 53 genes; g) Laux da Costa et al. [23], three genes; h) Maertzdorf et al. [24], three genes; i) Penn-Nicholson et al. [25], six genes (risk 6 score); j) Sambarey et al. [26], 10 genes; k) Singhania et al. [27], 20 genes; I) Suliman et al. [28], four genes; m) Sutherland et al. [29], four genes; n) Sweeney et al. [30], three genes; o) Thompson et al. [13], nine genes; p) Thompson et al. [13], 13 genes; q) Thompson et al. [13], 32 genes; r) Zak et al. [31], 16 genes.