![Sample our Mathematics & Statistics journals, sign in here to start your FREE access for 14 days]({"type":"image" , "src" : "/sda/5943/TOC-Subject-Sample-2021-Mathematics-Statistics.jpg"})
Abstract
For computational agent-based simulation, to become a serious tool for investigating biological systems requires the implications of simulation-derived results to be appreciated in terms of the original system. However, epistemic uncertainty regarding the exact nature of biological systems can complicate the calibration of models and simulations that attempt to capture their structure and behaviour, and can obscure the interpretation of simulation-derived experimental results with respect to the real domain. We present an approach to the calibration of an agent-based model of experimental autoimmune encephalomyelitis (EAE), a mouse proxy for multiple sclerosis (MS), which harnesses interaction between a modeller and domain expert in mitigating uncertainty in the data derived from the real domain. A novel uncertainty analysis technique is presented that, in conjunction with a latin hypercube-based global sensitivity analysis, can indicate the implications of epistemic uncertainty in the real domain. These analyses may be considered in the context of domain-specific knowledge to qualify the certainty placed on the results of in silico experimentation.
Acknowledgements
Paul Andrews is funded by EPSRC grant EP/E053505/1. Work in Dr. Kumar's laboratory has been supported by grants from the National Institutes of Health, USA.
Notes
1. The EPSRC funded Complex System Modelling and Simulation infrastructure (CoSMoS) project. http://www.cosmos-research.org/.
2. With regard to simulation, predictive results are those which highlight possible scenarios that could arise in the real domain being modelled.
3. This is achievable in the real domain through a number of different interventions [21], whilst in the simulation it is trivial to revoke the ability of CD8Tregs to kill CD4Th1 cells, hence disabling regulatory activity.
4. One may not be certain that the prediction definitely does rely on this, since abstraction dictates that the relationship between simulation parameters and domain parameters is not exact. The analysis can, however, indicate that caution must be exercised when interpreting simulation results.
5. Samples from parameter space obtained using latin hypercube sampling can be computationally expensive to execute in the simulation. For example, parameters might be chosen that generate a lot of cells that do not die quickly. As such, only 300 simulation executions are performed here.
6. Available upon request from the corresponding author.