Optimizing conductance parameters of cortical neural models via electrotonic partitions.

Development of automated methods for fitting computational models to observed biological data is an important challenge of neural modeling. Previous work has focused on generalized search techniques combined with distance measures tuned to specific neural morphologies. We propose general analysis techniques to guide construction of distance measures across a broader range of cell types. Specifically, we evaluate the use of multiple external stimuli to evoke characteristic behaviors of underlying active channel densities on a simple three-compartment model. We also examine the use of frequency analysis to smooth search space distortions induced by temporal shifts in recorded voltage traces. We propose a novel method of parameter optimization that is characterized by linear regression over the conductance densities using channel permissiveness as a basis of ionic current. We derive this method and demonstrate, given known anatomy and kinetics, it will solve all conductance densities in an N compartment model given N spatially distinct membrane potential traces with minimal error. We compare the regression method with the covariance matrix adaptation evolutionary strategy (CMA-ES) over a two-compartment cortical neuron and empirically show that regression over electrotonic partitions solves the cortical model near-optimally. We also show that electronic partitioning significantly improves search performance of CMA-ES on the cortical model.