Reinforcement Learning Based Model Selection and Parameter Estimation for Pharmacokinetic Analysis in Drug Selection

Abstract

Selecting effective drug candidate is a crucial procedure in drug discovery and development. Dynamic Positron Emission Tomography (dPET) is an ideal imaging tool for pharmacokinetic analysis in drug selection, because it offers possibilities to tract the whole procedure of drug delivery and metabolism when the drug is radio-labeled properly. However, various challenges remain: 1) the kinetic models for drugs are generally very complicated and selecting a proper model is very difficult, 2) solving the kinetic models often needs special mathematical considerations, 3) dPET imaging suffers from poor spatial and temporal resolutions, 4) blood sampling is required in pharmacokinetic analysis, but it is very hard to generate an accurate one. In this paper, we propose a reinforcement learning based model selection and parameter estimation method for pharmacokinetic analysis in drug selection. We first utilize several physical constraints to select the best possible model from a bank of models, and then estimate the kinetic parameters based on the selected model. The method highly improves the accuracy in model selection and can estimate corresponding kinetic parameters even with an inaccurate blood sampling. The quantitative accuracy of our method is tested by experiments using digital phantom and Monte Carlo simulations. Furthermore, 3 cases of patient studies on model selection and parameter estimation are also provided to show the potentials to reduce drug development cycle and save money for the pharmaceutical industry.