Towards a Scalable, Future-Proof Platform for Dynamical Modeling in Biology

Thesis (Ph.D.)--University of Washington, 2019Mathematical models are used widely throughout the sciences, and often influence not only research, but our daily lives. For example, weather prediction is made possible by numerical weather models that have advanced steadily since computers became widely available in the 1970's. The improvements in weather models have been so consistent that some researchers have coined a "Moore's law" for weather models. In its original formulation, Moore's law states that the number of transistors in an integrated circuit such as a CPU approximately doubles every year. The weather forecasting equivalent states that the accuracy of numerical weather models improves by ten percent every ten years [291,181]. Systems biology models do not advance to a steady drumbeat as weather models do. It can hardly be claimed that a ten year period yields a ten percent universal improvement in systems biology. In fact, current systems biology models are in many ways more primative than weather models¹. Systems biology models (especially detailed, mechanistically-accurate models) are underutilized in synthetic biology and are almost completely absent from the clinic. This is unfortunate, because systems biology models have the potential to aid in drug discovery [57], cancer treatment [28,159,117], disease biology [227, 307], and production of biofuels [56, 170, 46]. Why are systems biology models so underutilized, despite considerable advances in computing power and experimental data? There are at least three major factors: difficulty in model reuse, lack of scalability, and lack of technological advances for simulation. Specifically, these three factors are due to the following problems: 1. Perhaps more so than any other field, models in systems biology need to be reusable and reproducible ². Due to the complexity of biology, no single research team can specialize in every subsystem of the cell. Therefore, models of cellular subsystems must be developed, validated, and analyzed by different research teams, and combined into a single, larger model of the cell. This can only happen if researchers use uniform standards to store their work, and provide a means for others to reuse and incorporate their models. 2. As the size of a model increases, so do the resources required to simulate it. This is not simply an issue of convenience, since fitting a model requires simulating it many times. The strain on computational resources has caused many groups to use simpler "constraint-based" modeling approaches. However, this approach trades detail for performance. Continued advancement of systems biology requires the development of mechanistically accurate kinetic models, which is currently hampered by scalability constraints. 3. Finally, despite decades of innovation in computer simulation, kinetic models are still commonly simulated using 40-year old solvers such as LSODA [231, 124]. It is natural to ask wether advancements in computer technology could be used to provide a better approach to simulating models. Indeed, this thesis considers state-of-the-art specialized silicon hardware specifically designed to simulate kinetic models. This thesis seeks to address these factors through technological innovations which enable the construction of larger, more accurate, and more robust models. This is accomplished through providing better solutions for encoding and reusingg models, providing a scalable solution for optimizing large, challenging kinetic models, and providing a way to simulate models on special-purpose hardware. Taken together, these foundational advances in modeling technology provide a pathway toward building larger, more complex, and more accurate models