Whole-Blood 3-Gene Signature as a Decision Aid for Rifapentine-based Tuberculosis Preventive Therapy

Abstract

Background: Systemic drug reaction (SDR) is a major safety concern with weekly rifapentine plus isoniazid for 12 doses (3HP) for latent tuberculosis infection (LTBI). Identifying SDR predictors and at-risk participants before treatment can improve cost-effectiveness of the LTBI program.

Methods: We prospectively recruited 187 cases receiving 3HP (44 SDRs and 143 non-SDRs). A pilot cohort (8 SDRs and 12 non-SDRs) was selected for generating whole-blood transcriptomic data. By incorporating the hierarchical system biology model and therapy-biomarker pathway approach, candidate genes were selected and evaluated using reverse-transcription quantitative polymerase chain reaction (RT-qPCR). Then, interpretable machine learning models presenting as SHapley Additive exPlanations (SHAP) values were applied for SDR risk prediction. Finally, an independent cohort was used to evaluate the performance of these predictive models.

Results: Based on the whole-blood transcriptomic profile of the pilot cohort and the RT-qPCR results of 2 SDR and 3 non-SDR samples in the training cohort, 6 genes were selected. According to SHAP values for model construction and validation, a 3-gene model for SDR risk prediction achieved a sensitivity and specificity of 0.972 and 0.947, respectively, under a universal cutoff value for the joint of the training (28 SDRs and 104 non-SDRs) and testing (8 SDRs and 27 non-SDRs) cohorts. It also worked well across different subgroups.

Conclusions: The prediction model for 3HP-related SDRs serves as a guide for establishing a safe and personalized regimen to foster the implementation of an LTBI program. Additionally, it provides a potential translational value for future studies on drug-related hypersensitivity.

Keywords: interpretable machine learning; latent tuberculosis infection; rifapentine; systemic drug reaction; transcriptome.

Figure 1.

Overview of the case enrollment and analysis plan. The main steps included case enrollment, biomarker derivation, SDR model construction, and validation. S_A represents samples from the SDR group after 3HP treatment; S_B represents samples from the SDR group before 3HP treatment; N_A represents samples from the non-SDR group after 3HP treatment; and N_B represents samples from the non-SDR group before 3HP treatment. ^∗Seven cases had acute upper respiratory infection, 2 had urinary tract infection, 2 had influenza, 1 had pneumonia, and 1 had cellulitis. ^#Please see Supplementary Methods. Abbreviations: 3HP, weekly rifapentine plus isoniazid for 12 doses; BS, biological similarity score; HiSBiM, hierarchical system biology model; LTBI, latent tuberculosis infection; qPCR, quantitative polymerase chain reaction; SDR, systemic drug reaction; SHAP, SHapley Additive exPlanations.

Figure 2.

Gene expression signature and therapy–biomarker pathway for predicting SDR in participants with latent tuberculosis infection before treatment with 3HP. A, Heat map and hierarchical clustering of gene expression (left) and pathway (right) for 19 potential biomarkers in S_A, S_B, N_A, and N_B samples. Of the 19 genes, 4 are significantly upregulated after 3HP treatment (dark green), whereas the other 15 genes are not (light green). Among these genes, the 6 selected potential biomarkers are marked in red. B, Therapy–biomarker pathways for illustrating potential genes associated with SDR development under 3HP treatment. For better visualization, the 19 potential biomarkers are underlined with the 6 selected genes marked in red. C, Bar chart for the expression of ATP5PF, GABARAPL2, ATP6V0E1, PIGX, SPCS1, and DDT in 3 SDR (red) and 2 non-SDR (blue) samples collected before 3HP treatment. The expression levels were validated through reverse-transcription quantitative polymerase chain reaction. Abbreviations: 3HP, weekly rifapentine plus isoniazid for 12 doses; MHC, major histocompatibility complex; N_A, samples from the non-SDR group after 3HP treatment; N_B, samples from the non-SDR group before 3HP treatment; NAD, nicotinamide adenine dinucleotide; SDR, systemic drug reaction; S_A, samples from the SDR group after 3HP treatment; S_B, samples from the SDR group before 3HP treatment.

Figure 3.

RF model and SHAP models of the 6 selected genes to discriminate pretreatment samples collected from participants with and without SDR in the training cohort. A, Box plot of the G-mean of sensitivity and specificity of the RF model (green) and SHAP model (yellow) in 50 random unbalanced testing sets (8 SDR and 27 non-SDR samples) and balanced testing sets (8 SDR and 8 non-SDR samples). The P values were calculated using the Mann–Whitney U test. B, Box plot of the G-mean of the RF model (green) and SHAP model (yellow) in 50 random testing sets (8 SDR and 27 non-SDR samples) under various model parameters, including default, number of trees (500), number of genes (6), minimum number of samples required to be at a leaf node (leaf = 5), and minimum number of samples required to split an internal node (split = 5). The P values were calculated using the Mann–Whitney U test. C, Box plot of the expressions of ATP5PF, ATP6V0E1, PIGX, SPCS1, GABARAPL2, and DDT for the 28 SDR (pink) and 104 non-SDR (blue) training samples. Boxes indicate the sample median and interquartile range, whereas bars and colored dots indicate the range and outliers, respectively. Data were analyzed using the Mann–Whitney U test. D, Box plot of the SHAP output values of the 4 best performing models in SDR (orange) and non-SDR (steel blue) training samples. Abbreviations: G-mean, geometric mean; RF, random forest; SDR, systemic drug reaction; Sen, sensitivity; SHAP, SHapley Additive exPlanations; Spe, specificity.

Figure 4.

SHAP models of the 2 selected models to discriminate pretreatment samples collected from participants with and without SDR in the testing cohort. A, Receiver operating characteristic curve and AUC of the 3-gene (ATP6V0E1-PIGX-SPCS1) and 4-gene (ATP6V0E1-PIGX-SPCS1-DDT) models in the testing cohort. B, Box plot of the SHAP output values of 2 models for the 8 SDR (orange) and 27 non-SDR (steel blue) testing samples. Boxes indicate median and interquartile range, whereas bars and colored dots indicate the range and outliers, respectively. Data were analyzed using the Mann–Whitney U test. C, Interpretation of SDR predictive models with universal cutoffs for SDR (orange) and non-SDR (blue) testing samples. The SHAP output value in each sample is a red diamond, and the universal cutoff is 0 (deep red line). G-mean represents the geometric mean of sensitivity and specificity. Abbreviations: AUC, area under the curve; SDR, systemic drug reaction; SHAP, SHapley Additive exPlanations.

Figure 5.

Forest plots of the performance of ATP6V0E1-PIGX-SPCS1 (A) and ATP6V0E1-PIGX-SPCS1-DDT (B) models used to predict SDR from pretreatment samples in the joint population of training and testing cohorts as well as various subgroups. Abbreviations: eGFR, estimated glomerular filtration rate; SDR, systemic drug reaction.