Uncertainty quantification in neuroscience

Abstract

The complexity of the nervous system has made computational science an invaluable tool in order to understand how the nervous system functions. The overarching goal of this thesis has been to develop software tools to improve areas of neuroscience that are currently lacking, which include uncertainty analysis of computational models (Paper I), data storage (Paper IV), and education (Paper V). The major area of focus has been that of uncertainty analysis. Computational models always contain parameters that describe the system to be modeled. These parameters are for various reasons often uncertain. An uncertainty analysis provides rigorous procedures to quantify how the model depends on this parameter uncertainty. To reduce the barrier of performing uncertainty analysis in neuroscience we have created a toolbox for uncertainty analysis (Paper I). We then used this toolbox on a selected set of models (Paper I, II and III). In Paper I we introduced Uncertainpy, a Python toolbox for performing uncertainty quantification and sensitivity analysis. Uncertainpy is tailored for neuroscience applications by its built-in capability for calculating characteristic features in the model output. We provided a detailed user guide for Uncertainpy and illustrated its use by showing four different case studies. In Paper II we presented a reimplementation of a model for endocrine pituitary cells in rats. We qualitatively replicated the computational results in the original publication and confirmed the key conclusions, namely that big conductance K+ (BK) ion channels are important for the bursting activity of endocrine pituitary cells in rats. Additionally, we performed an uncertainty analysis of the model using Uncertainpy, which further strengthened the findings in the original publication. In Paper III we created a computational model for endocrine pituitary cells in medaka, a species of Japanese rice fish. The reimplementation and results in Paper II were used as a basis for the computational work in this paper. We discovered that the BK conductance has the opposite effect on the action potential shape in medaka pituitary cells compared to in the rat pituitary cells in Paper II. The BK channels makes the action potentials generated in the medaka model narrower, but they make the action potentials generated in the rat model broader. An uncertainty analysis of the two models was performed in order to examine differences in the sensitivity of the models to changes in their ion channel conductances. In Paper IV we developed a specification for organizing data in a hierarchy by using file-system directories to represent the hierarchy. We used the same data abstraction as in the HDF5 file format. We provided a reference implementation in Python and described how to use this implementation. In Paper V we introduced Neuronify, an educational app for easily creating neural networks by dragging and dropping neurons onto the canvas and then simulating the networks. Neuronify is available for iOS and Android, as well as Mac, Linux, and Windows.