Drug combination screening using machine-learning and PK/PD modeling
Joseph Cicchese1, Shuyi Ma2,3, Suraj Jaipalli4, Awanti Sambarey4, Jonah Larkins-Ford5, Jenny Lohmiller2, Bree B. Aldridge5, David R. Sherman2,3, Denise Kirschner6, Jennifer Linderman1, Sriram Chandrasekaran4
1 - Department of Chemical Engineering, University of Michigan, Ann Arbor, Michigan, USA; 2 - Department of Microbiology, University of Washington, Seattle, Washington, USA; 3 - Center for Global Infectious Disease Research, Seattle Children’s Research Institute, Seattle, Washington, USA; 4 - Department of Biomedical Engineering, University of Michigan, Ann Arbor, Michigan, USA
5 - Department of Molecular Biology and Microbiology, Tufts University School of Medicine, Boston, Massachusetts, USA
6 - Department of Microbiology and Immunology, University of Michigan Medical School, Ann Arbor, Michigan, USA
The search for new TB therapies is hindered by the large number of drug combinations, dosing choices, and complex pharmaco-kinetics/dynamics (PK/PD). Here we present a pipeline that incorporates both machine learning and PK/PD modeling to design multi-drug regimens. We utilize a machine-learning model, INDIGO-MTB, to predict drug interactions using transcriptome profiling of individual drugs. The INDIGO-MTB model contains 164 drugs with anti-TB activity and identified synergistic two-, three-, and four-drug regimens from over a million possible combinations. By linking INDIGO-MTB with a model of drug PK/PD and pathogen-immune interactions called GranSim, we calculate an in vivo interaction score for drug combinations. The interaction score incorporates dynamics of drug diffusion, spatial distribution, and activity within lesions against various pathogen sub-populations. Predictions from our model significantly correlates with both experimental interaction scores and efficacy metrics from clinical trials. Our framework enables efficient optimization of combination therapies (Ma et al, mBio, 2019; Cicchese et al, Biorxiv, 2020).