Fully-automated multi-objective optimization for fitting a neuronal model with real morphology

Abstract

Morphologically realistic models have successfully been used to elucidate many complex mechanisms in neuronal dendrites. However, the tuning of such models to match experimental data remains challenging. Here we introduce a fully automated parameter optimization methodology that uses the Python programming language to control the NEURON simulator in parallel on a high performance computing cluster. Using targeted experimental protocols, including sub- and supra-threshold somatic as well as dendritic voltage recordings, we constrain a model hippocampal CA1 pyramidal cell built with a complete reconstructed morphology. The optimization is performed using the non-dominated sorting genetic algorithm (NSGA-II), and model fitness is evaluated by directly comparing the simulated and recorded voltage traces. In order to impose minimal a priori assumptions, we use a multi-objective framework, which tunes all of the free parameters with respect to all of the experimental objectives simultaneously. Furthermore, the multi-objective approach avoids the pitfalls of overfitting, because the algorithm produces a diverse family of solutions on the so-called Pareto-optimal front. To facilitate model selection, we have developed a clickable interface for visually browsing the set of optimal solutions, which permits the explicit and rapid identification of trade-offs among the fitting objectives and the biophysical parameters that govern variability in the solution set. Figure 1 Clickable interface for browsing the Pareto front of solutions produced by our multiobjective optimization platform. Top row: Scatterplot of error scores across a 6-dimensional error space. The abscissa and ordinal represent errors in the objectives shown ...