Deep Learning for PK/PD Modeling
Objective
Reduce the time and effort required by expert modelers to do PK/PD system identification in pharmacology. Accurate PK/PD modeling in pharmacology supports drug discovery and personalized dosing. Our work with a global pharma company focused on develop deep learning models with synthetic data samples from “forward” PK/PD models to reduce time and modeling expertise required for solving “inverse” PK/PD modeling—e.g., chemical coefficient estimation from drug response time series.
Approach
In a matter of months, using internal design patterns and open source tools like PyTorch and MLFlow, developed the Hemetox Deep Learning Platform. Developed synthetic data generator to generate input/output data for viable regimes in PK/PD coefficient space. Rapidly experimented with different deep learning architectures for PK/PD system identification. Developed model performance metrics across experiments to understand error and develop insights for improvements.
Impact
Advanced ability to train architectures for in-vitro and in-vivo problems with low error estimation over a large range of parameters by efficiently completing thousands of experiments. Developed Singularity image deployed on HPC cluster at pharma company. Empowered modeling and simulation teams with self-service workflows for rapid experimentation.