Transcriptomic signature for precise prediction of 3HP related systemic drug reaction for latent tuberculosis infection treatment: A prospective study

Background: Systemic drug reaction (SDR) is a major safety concern of weekly rifapentine-based treatment (known as 3HP) for latent tuberculosis infection (LTBI). Understanding predictors of SDR and identifying risk subjects before treatment can improve public acceptance and cost-effective implement of LTBI program. Methods: We first prospectively recruited 187 subjects receiving 3HP and randomly selected a pilot cohort for generating whole-blood transcriptomic data. After integrating the hierarchical systems biology model (HiSBiM) and a therapy-biomarker pathway approach, candidate genes were obtained and evaluated by reverse transcription-quantitative polymerase chain reaction (RT-qPCR). Then, we developed interpretable machine learning models and established universal cut-off value (shapley additive explanations value) for each SDR prediction model. An independent cohort was used to evaluate the performance of these predictive models. Findings: Based on the transcriptomic profile of the pilot cohort, a total of 19 candidate genes were identified. Among them, six were sieved out by RT-qPCR results of a small number of samples from the training cohort. By applying the SHAP values of various combinations of the 6 gene expression signature, and the best model consisted of three genes, had an G-mean of 0.959, sensitivity of 0.972 and specificity of 0.947 for the joint of the two cohorts, and worked well across different subgroups. Four groups: 8 samples before 3HP treatment from SDR group (SB), 8 samples after 3HP treatment from SDR group (SA), 8 samples before 3HP treatment from non-SDR group (NB), and 12 samples after 3HP treatment from non-SDR group (NA)