Statistical and Modelling Techniques to Build Confidence in the Investigation of Immunology through Agent-Based Simulation

For over a century immunological research has made striding advancements that have substantially improved the wellbeing and longevity of the population. There remain, however, many pathogens and diseases to which the immune system and immunologists currently have no answer. Immunological research is as active now in understanding the immune system and its response as it ever has been. Extraordinary technological breakthroughs allow researchers to examine the genes and proteins expressed in individual cells, and the molecules expressed on their cell surfaces. Advancements in imaging techniques allow researchers to observe immune cell interactions at levels of detail never before possible. Yet, despite impressive technological advancements, there is sentiment within immunology that reductionist techniques alone will not be sufficient to explain manner in which system-level behaviours arise from low-level components. Such an understanding is necessary to develop the best strategies for fighting disease. An increasing number of immunologists are combining traditional wet-lab research programs with computational methods that consolidate wet-lab data and attempt to reconstruct the system-level phenomenon observed in the real-world immune system. These computational techniques facilitate hypothesis formation and exploration, allow for predictive experimentation that is not possible in the real-world, and can guide wet-lab work towards potentially fruitful avenues of research. However, not well appreciated in the literature is the issue of simulation validity: confidence that in silico results are representative of the real-world system that simulations attempt to capture. There is considerable uncertainty surrounding many immunological phenomenon, which can complicate the creation of simulations that are themselves highly abstract entities. The results of simulation will not necessarily translate directly into the real-world domain. This thesis investigates modelling and statistical techniques that establish confidence in in silico results being representative of the real-world immune system. A case study in the murine autoimmune disease experimental autoimmune encephalomyelitis (EAE), a model for multiple sclerosis, is undertaken as a means to explore these techniques. The disease is subject to rigorous domain modelling prior to its representation in simulation. Modelling and simulation calibration are performed in close collaboration with a domain expert in EAE. A means of grading simulation executions by the same scale employed in the wet-lab, through examination of an entire mouse, is created. A novel technique is developed whereby the relationship between the accuracy of averaged simulation results and the number of simulation executions sampled in compiling them is established. The completed simulation is explored using a global sensitivity analysis, and a novel robustness analysis technique. These analyses reveal aspects of the simulation that are highly influential to its overall system-level behaviours. The extent to which simulation behaviours are reliant on parameters being assigned particular ranges of values is investigated using robustness analysis. This thesis theorises how these techniques may be considered and applied in the context of domain-specific knowledge to interpret in silico results into the original domain. Next, in silico experimentation into the nature of EAE is performed, resulting in several predictions concerning the role of particular cells and the spleen in mediating recovery from disease. Lastly, this thesis reflects upon the contribution of these modelling and statistical techniques in providing confidence that simulation results are representative of real-world EAE. A novel approach to development that guides simulations to appropriate levels of abstraction, and demonstrates this to be the case, making extensive use of metaheuristic search and real-world experimental data is proposed.