Not a current user.

GPU-Accelerated Model-Based Drug Development

Drug development is a very time-consuming, complex and expensive process. In fact, most of the cost and time spent in drug development are in clinical trials, and translation of scientific discoveries into innovative safe and effective drug therapies is rate limiting. The U.S. Food and Drug Administration recently established a new approach to assess the efficacy and safety of investigational drugs through its Critical Path Initiative in order to overcome problems associated with drug development. One aspect of this initiative is to transition from empirical to model-based drug development. This new approach emphasizes the use of modeling and simulation technologies to integrate drug-specific pharmacokinetic (PK), pharmacodynamics (PD), and disease-specific information into the overall drug development process. This strategy quantitatively and qualitatively utilizes our growing understanding of pharmacology and the underlying disease process to improve the efficiency of drug development by selecting optimal doses, designing efficient clinical trials, and developing evidence-based and well-informed drug development strategies.

Population pharmacokinetics, pharmacodynamics, and disease process is the important link between dose and clinical responses. Developing a quantitative model to understand these processes is an essential step in the model-based drug development, and allows us to 1) design optimal dosage regimen of the promising drug candidate in order to achieve maximum clinical benefit-to-risk ratio in patient population, and 2) make an early decision to no longer pursue less viable drug candidate for avoiding expensive later stage failure. However, development of complex population PK/PD/disease models is one of the most unique numerical challenges in the biological/medical fields.

In contrast to the computational challenge of the molecular mechanics that typically focuses on studying the complex interaction of a single molecule with a large system of ordinary differential equations (ODEs). The complex population PK/PD/disease model development involves repeatedly solving millions of identical ODEs system for thousands of studied subjects within the integration framework of nonlinear-mixed effect statistical (NLME) methods consisted of complex high-dimensional integrals with no closed-form solution. The model development process may take months to years, therefore the decision for the next drug development phase is often made empirically before the analysis is completed. This practice significantly limits the impact of insights gained from performing population PK/PD/disease modeling in drug development.

The recent advancement of computational technologies with massively parallel multi-core and GPU architectures provides the opportunity to significantly decrease the model development times and allows timely access to the modeling and simulation results in supporting key decisions making for drug development. Recently, using a single workstation with a NVIDIA Tesla C1060 GPU computing card, our group developed and reported a first prototype of hybrid CPU/GPU-based parallelized Monte-Carlo Parametric Expectation-Maximization algorithm (MCPEM) for simple population PK data analysis (Ng CM, AAPS J, 15:1212-21,2013). By converting a single computational steps of the MCPEM algorithm from serial to parallel execution using a simple high level Matlab code, the hybrid CPU/GPU-based algorithm is about 30-folds faster than the same algorithm implemented in CPU for completing the data analysis with 100 simulated subjects. This preliminary result suggests that innovative computing approaches to modify existing estimation algorithms can lead to vast speed-up in developing complex population PK/PD/disease model for model-based drug development.

Therefore, the objective of this project is to develop the FIRST massively parallel computing platform for complex population PK/PD/disease model development. The proposed numerical algorithms subjected to parallelization including MCPEM and iterative-two-stage methods. These developed methods will then be combined with other data mining tools (i.e., artificial intelligence and others) to form a novel FIRST many-task and high-performance computing platform that can significantly improve the efficiency and lower the cost of drug development, and may ultimately lead to lower drug price and economic burdens of patients and society.

Researcher Participants

Faculty: Prof. Chee M. Ng
Postdoc: Dr. Chin F. Ng

Computational methods

Monte Carlo, MC-PEM, iterative two-stage Bayesian method

Software

C, C++, gcc, CUDA, GPUs, MATLAB, OpenMP, OpenACC, vim

UK/non-UK collaborators

Currently no UK or non-UK collaborators

Browser not supported