2003
   Verona, Italy

Applications of Distributed computing in Drug development

Mark Sale

Clinical Pharmacology and Discovery Medicine, GlaxoSmithKline, 5 Moore Drive, RTP NC

Objective: Improve speed and robustness of NONMEM analysis.

Methods: Distributed computing is a collection of methods that permit the application of many computers to solve a computationally intensive problem. The key to the ability to use distributed computing is the ability to break the problem into appropriately sized pieces. In biological sciences the most common use for distributed computing is in computational chemistry. Two applications in this area are BLAST [1] and GOLD [2]. Two applications in modeling and simulation have found distributed computing to be useful. Entelos's Physiolab uses 145 computers to perform Monte Carlo Simulations for complex physiological models [3]. In addition to applications in computational chemistry, GSK has applied distributed computing to automated model selection with NONMEM, using a Genetic algorithm based approach.

Genetic Algorithm is a discrete space search algorithm based on the mathematics of evolution/survival of the fittest. Preliminary work suggests that this approach is more robust than the conventional forward addition method of model selection [4]. The limitation of this approach is that it is inefficient, requiring the evaluation of thousands of models to insure robustness in finding the optimal solution. However, Genetic algorithm is very well suited to distributed computing, as a "population" of several hundred NONMEM models is created at a time, each being independent of the others. Each model in a "generation" then can be sent to a separate computer for minimization and the results returned. The application of distributed computing to Genetic algorithm model selection in NONMEM has made the approach practical, with typical analyses, comprised of several thousand model runs requiring only a few days.

Results: Distributed computing increases the speed of Genetic Algorithm model selection in NONMEM by up to two orders of magnitude. This permits the assessment of a sufficient number of models (typically 5000) to insure robustness of the analysis.

Conclusions: Genetic Algorithm with distributed computing is a promising technology to provide dramatic improvements in speed and robustness of NONMEM analysis.

References:

[1] Madden T, Tatusov R, Zhang J. Applications of network BLAST server Meth. Enzymol. 266:131-141. 1996

[2] Jones G, Willett P, Glen RC. Molecular recognition of receptor sites using a genetic algorithm with a description of desolvation. Journal of Molecular Biology. 245(1):43-53, 1995

[3] Metz, C. Grid computing: Case Study - Entelos. PC Magazine Jan 2003.

[4] Bies R, Sale M, Smith, G, Muldoon, M. Lotrich F, Pollock, B. Outcome of NONMEM analysis depends on modeling strategy, Clin Pharm Ther, 73(2) 2003 (abstract)