PBPK/PD modeling and machine learning approaches to support the development of new drugs and regimens against tuberculosis.
Federico Reali (1), Roberto Visintainer (1), Anna Fochesato (1,2), Chanchala Kaddi (3,+), Karim Azer (3,#), Micha Levi (3), Shayne Watson (3), Veronique Dartois (4), Luca Marchetti (1,5)
(1) Fondazione The Microsoft Research - University of Trento Centre for Computational and Systems Biology (COSBI), Rovereto, Italy; (2) University of Trento, Department of Mathematics, Trento, Italy; 3) Bill & Melinda Gates Medical Research Institute, Cambridge, MA, USA; (4) Center for Discovery and Innovation, Hackensack Meridian Health, NJ, USA; (5) University of Trento, Department of Cellular, Computational and Integrative Biology – CIBIO, Trento, Italy; (+) current affiliation: Translational Disease Modeling, Data and Data Science, Sanofi, USA; (#) current affiliation: Axcella Therapeutics, Cambridge, MA, USA
Objectives: Tuberculosis (TB) is a major global health problem that causes millions of new cases and deaths every year. Several drug development programs are working on fighting the emergence of resistant strains, shortening treatment time, and increasing the patient-adherence. In this regard, in silico pipelines are instrumental in closing the gap between in vitro and in vivo, guiding go/no-go decisions, and prioritizing candidates towards clinical applications [1].
Methods: A minimal PBPK model (mPBPK) tracking drug exposure in TB-key organs was developed and implemented in Matlab to fit mouse and rabbit PK data. Internal and literature data were organized in an online navigation tool that supports data visualization and graphical comparisons with model simulation results. The mPBPK model was then coupled with a pharmacodynamic (PD) two-compartment disease model describing Mycobacterium tuberculosis growth dynamics in log-phase and semi-dormant states. The unknown kinetic and kill rates for the PD module were determined using parameter estimation procedures integrating mice, rabbit, and hollow fiber system (HFS) data. Currently, this integrated framework includes several compounds under development, as well as a few marketed drugs, and different regimens [2,3]. In parallel, the mPBPK model was embedded in a Machine Learning (ML) pipeline to predict the penetration of historical and new anti-TB compounds at the target sites starting from physicochemical and in vitro information [4,5]. The hybrid ML platform uses a minimal set of easy retrievable physicochemical properties as input and mPBPK-predicted AUCs in plasma and TB-lesions as labels to learn functional dependencies between input and target during the training phase. Six ML algorithms were tested to accurately regress penetration values in rabbits of anti-TB compounds in the lesion and caseum.
Results: We have developed a comprehensive PBPK/PD modeling framework based on an ad hoc mPBPK model and on a disease model that connects drug exposure in blood, lung, pulmonary lesion, and caseum compartments with the treatment bactericidal and bacteriostatic effects. Informed by both in vivo (mice and rabbit) and in vitro (HFS) preclinical data, this integrated modeling approach allows evaluating the sterilizing efficacy of several single drugs and regimens in terms of reducing bacterial burden. Additionally, we have designed a machine learning pipeline for forecasting PK metrics that leverages a mechanistic-based protocol for data augmentation, validated through a repeated k-fold-cross-validation schema. The stability and the accuracy obtained when generalizing penetration levels to unseen anti-TB compounds guarantee a proficient usage of the mPBPK/ML tool as part of preclinical screening.
Conclusions: The proposed computational approaches effectively support the preclinical and translational stages of drug development against tuberculosis. Overall, our frameworks are providing accurate information on drug exposure and efficacy to facilitate dose selection and experimental study design and, ultimately, can be used to prioritize regimens to be translated into clinical development based on an established target product profile.
References:
[1] Azer et al., History and Future Perspectives on the Discipline of Quantitative Systems Pharmacology Modeling and Its Applications, Front. Physiol. (2021), 12:637999
[2] Strydom et al., Tuberculosis drugs’ distribution and emergence of resistance in patient’s lung lesions: A mechanistic model and tool for regimen and dose optimization, PLoS Med. (2019), 16(4): e1002773
[3] Wicha et al., Forecasting Clinical Dose-Response From Preclinical Studies in Tuberculosis Research: Translational Predictions With Rifampicin, Clin. Pharmacol. Ther. (2018), 104(6): 1208-1218
[4] Pillai et al., Machine-learning-guided early drug discovery of small molecules. Drug Discov. (2022).
[5] Sarathy et al., Prediction of Drug Penetration in Tuberculosis Lesions. ACS Infect. Dis. (2016). 2(8): 552–563.
