Data driven building of realistic neuron model using IBEA and CMA evolution strategies

Recently the building of large neuronal circuits from realistic neuron models has gained traction. This bottom-up approach relies on the accurate description of the primitive elements composing the brain such as neurons and astrocytes, that are then aggregated into larger and larger circuits. However, as of today, this data-intensive approach is slowed done by the lack of complete biological description that would be needed to build such models. In the present study, we compare the use of different optimizers in the context of building numerical neuron models presenting realistic electrical behaviours despite only having a sparse description of the original neuron behaviour. To do so, we perform single and multi-objective optimization of the neuron model parameters using as targets electrical features (e-features) extracted from voltage recording obtained through patch-clamp experiments. The purpose of the optimizers is therefore to find the optimal set of parameters for the neuron model such that it presents a firing behaviour similar to the original neurons when exposed to the same stimuli. This neuron model building approach is not new [7, 10], however, to the authors knowledge, it is the first time that multi-objective covariance matrix adaptation evolution strategy is used in such a highly-dimensional parameter space using e-features obtained from experimental recordings.