Uncertainty quantification and sensitivity analysis in brain modelling
Uncertainty quantification and sensitivity analysis are important aspects of computational modelling, due to the need to assess the validity and precision of model predictions. In the first phase of Brain-IT we formed a prototype for a workflow to address this need (see Eriksson et al, 2018), consisting of i) parameter estimation from experimental data with uncertainty quantification of the estimates, ii) propagation of the parameter uncertainty to a corresponding uncertainty in the predictions, and iii) global sensitivity analysis of the prediction based on the parameter uncertainty. We used a Bayesian method (Approximate Bayesian Computation, ABC) for the uncertainty quantification and variance-decomposition methods (Soboll-Saltelli) for sensitivity analysis.
This approach has been shown to offer scientific insights and promising results for semi-large subcellular models such as those built in Multi-scale simulations of synaptic plasticity. The models tested had a large parameter uncertainty but at the same time they could display well constrained high-level aspects. The method could also provide guidance for further experiments. We intend to develop this prototype further in thefollowing directions:
- Making the current ABC approach more efficient, generally applicable and user friendly
- Developing an exact Bayesian method for subcellular models
- Investigating how global sensitivity analysis on non-orthogonal parameter distributions are best performed and relate pathway component sensitivity to evolution
- Investigating how to make structural uncertainty predictions and study the effects of bad structural assumptions on parameter estimation in different models
Uncertainty in structural dependencies (last point) is an area that needs more attention, and may benefit from experiences in Comparative Genomics. We would like to explore how evolutionary analysis may guide the investigation of structural uncertainty: the hypothesis is that pathway components that are more variable (during evolution) are also less important, while more conserved components are under higher (negative) selection and therefore more important for the system. A system to study may be the evolution of synapse-related proteins (see Grant 2016).