Message Passing Computational Methods with Pharmacometrics Applications

Abstract

A pharmaceutical company needs to invest in the costly and tightly regulated multi-year drug development process early on. While many compounds are considered initially, only a few make it to the final phase where the newly developed drug is made available to the wide population. Time and effort is lost on the majority of candidate compounds as these turn out to not be efficacious and no return on investment is made. For this reason, pharmaceutical companies are interested in methodologies and approaches to better detect the ones that have more promise as soon as possible. In addition, the exclusivity granted by a patent is limited in time, and speeding up the process translates into material financial gains. Model Based Drug Development, a quantitative approach where gathered data is leveraged in an online fashion to improve decision making, has been suggested as one way to optimize the process. The domain of Pharmacometrics, a key component of Model Based Drug Development, is concerned with modeling human-compound interactions. One of the challenges in Pharmacometrics is that the computational requirements of the models preclude agility and timeliness. Currently, modelers switch between different projects to avoid stalls as typical computational runs can take up to weeks, but arguably this hampers swift modeling due to the long feedback cycle. Recent computing systems have seen a surge in the number of explicitly exposed parallel resources due to the limits being reached in single processor systems. Leveraging the computational resources of these systems is far from trivial. Pharmacometrics, like other branches of computational science, is a multidisciplinary field. Scientists that specialize in the models are rarely equipped with the right Computer Science background to write efficient computational codes. Therefore, the common approach is to rely on software packages that provide a toolbox of mathematical and statistical methods. However, contemporary packages lack in efficiency when deployed on a parallel system. This motivates the need for new approaches like those explored in this thesis. In the context of computational modeling, there are two prominent strategies for parallelization. First, computations on models are fit using an iterative optimization routine where multiple processors can be kept busy within each iteration by evaluating multiple candidate parameters concurrently. This part of the computation is referred to as the back-end. While this approach can hide the parallel constructs within the routine improving the usability of these routines for scientists from other domains, it requires the optimization routine to be designed to run in parallel. However, this might not always be feasible. Second, in the front-end a single candidate parameter can be evaluated in parallel if permitted by the dependency structure of the model, a strategy suitable both for more sequential optimization routines as well as parallel optimization routines where it further improves performance. Even if a task can be decomposed into smaller concurrently executable tasks, doing so manually is tedious, errorprone and requires the right parallel computing background. Arguably, the scientist concerned with building these models is in an even worse position; their expertise is probably not in parallel computing and more automated approaches are preferable. This thesis proposes novel ways to leverage parallelism in both the front-end and the back-end. In the former, two approaches are presented to parallelize evaluations without any input from the user. In the latter, changes to two existing state-of-the-art Markov Chain Monte Carlo samplers are presented that allow to better deal with large parallel systems in the message passing paradigm. Improvements for samplers running with data-bound models are explored as well. One of the main properties of Pharmacometrics models is that computation time required for parts of the model depends highly on the choice of model parameters. For this reason, common approaches fail to perform well in this regard as they assume a more uniform execution time across evaluations. By neglecting this property, idle times are introduced resulting in poor use of available resources. Performance gains observed by the presented techniques vary greatly from 10% all the way to many hundred fold reductions in execution time. It is important to note that these improvements depend not only on the targeted algorithms, but also on the computation models and on the platform. Nevertheless, the ideas are presented at a reasonably abstract level to support generalizations to other domains and computational models.