Development of population pharmacokinetic models associated with machine learning algorithms to optimize tuberculosis treatment

• Grant number: 25/02626-1
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: October 01, 2025
• End date: November 30, 2028
• Field of knowledge: Biological Sciences - Pharmacology - Clinical Pharmacology
• Principal Investigator: João Paulo Bianchi Ximenez
• Grantee: Matheus de Lucca Thomaz
• Host Institution: Faculdade de Ciências Farmacêuticas de Ribeirão Preto (FCFRP). Universidade de São Paulo (USP). Ribeirão Preto , SP, Brazil
• Associated research grant: 24/05744-2 - Gut microbiome in Tuberculosis: a challenge in antibiotic pharmacokinetics, AP.JP

Abstract

Tuberculosis (TB) treatment is often monitored clinically in a pragmatic manner, using subjective symptom assessments and objective but imprecise measures such as weight gain. The treatment failure rate is 14%, however, in HIV-positive patients, the failure rate increases to approximately 21%. Our main objective is to optimize TB treatment through model-informed precision dosing by developing population pharmacokinetic models and applying machine learning approaches. To achieve this, we will recruit TB patients with a negative HIV diagnosis (TB-HIV¿ group; n = 20) and TB patients with a positive HIV diagnosis (TB-HIV+ group; n = 20) after the second month of treatment with rifampicin, isoniazid, pyrazinamide, and ethambutol, administered according to body weight. Serial blood samples will be collected over a 24-hour period after drug administration. In the present project, the population pharmacokinetics of these antibiotics will be modeled using nonlinear mixed-effects models through the NONMEM software. The influence of covariates on the pharmacokinetic parameters of the drugs will be assessed stepwise by inclusion/removal. The predictive performance of the model will be evaluated using graphical and statistical criteria. We will develop and validate XGBoost, RandomForest, MARS, and GLMNET algorithms using different combinations of simulations generated from previously developed population pharmacokinetic models. The simulation data will be divided into a training set (75% of the data) and a test set (25%) for model development and validation. Once optimized, the algorithms will be evaluated on the test set to select the best-performing model. This approach will enable a lower error in the determination of pharmacokinetic parameters and reduce bias in simulations, allowing us to explore multiple dosing scenarios and drug combinations used in TB treatment, ultimately leading to a higher therapeutic success rate.