Brute force parameter fitting for a simulation of an in vitro neuronal network

Abstract

Event Abstract Back to Event Brute force parameter fitting for a simulation of an in vitro neuronal network Kerstin Lenk1*, Kay Barthold1 and Barbara Priwitzer1 1 Lausitz University of Applied Sciences, Faculty of Engineering and Computer Science, Germany We developed a pulsing model called INEX [1] which simulates neuronal activity as observed in neuronal networks cultivated on multielectrode array (MEA) neurochips. In an in vitro experiment approximately 500,000 cells of the frontal cortex of embryonic mice are cultivated on such a MEA neurochip (experiments are carried out by NeuroProof GmbH, Rostock, Germany). Circa 10,000 neurons and 90,000 glia cells of the total amount survive. The INEX model is based on an inhomogeneous Poisson process [2] to simulate neurons which are active without external input or stimulus as observed in neurochip experiments. It is accomplished using an Ising model [3] with Glauber dynamics [4] and has four parameter types: basic activity c, the excitatory weights w_e of the synapses which have positive values, the inhibitory weights w_i of the synapses which have negative values and a factor f indicating the varying probability for the occurrence of spikes depending on the previous activity of the neuron. Using an automated fitting method, these four parameter types were chosen in the following way: First, the ranges for the four parameter types were set. Then we simulated twenty times a neuronal network with 800 excitatory and 200 inhibitory neurons with parameter type values in different intervals but within the range. For each spike train twelve activity describing features were calculated. The same features were calculated from the spike trains of twenty MEA neurochip experiments with cell cultures originated from frontal cortex of embryonic mice after 28 days in vitro. Per experiment 15 to 50 spike trains were available. The simulated and the experimental spike strains had a length of 30 minutes. The mean value and the standard deviation were calculated for all features. The parameter fitting was considered as successful if the calculated features of the simulated spike trains lay within the standard deviation of the MEA neurochip features. The parameter types are scalable. For the choice of the parameter range we developed a heuristic based on spike rates and burst rates observed in experiments and the dependency between the parameters. The described brute force parameter fitting was applied within the intervals according to the derived parameter dependency. Results were shown in several isoline 3D plots depending on the value of f with the mean basic activity of all 1,000 neurons on the x-axis, the mean inhibitory weights on the y-axis and the mean excitatory weights on the z-axis. The simulated spike trains show typical synchronous spike and burst patterns as known from MEA neurochip experiments with frontal cortex neurons. Calculated features adapted from spikes and bursts indicate that the presented model simulates neuronal activity similar to activity as observed on MEA neurochips. Acknowledgements We thank Matthias Reuter (Clausthal University of Technology, Germany), Lars Schwabe (University of Rostock, Germany) and last not least Olaf Schröder (NeuroProof GmbH, Rostock, Germany).