Modeling the Electrical Diversity of Cortical Interneurons

Neuron models, in particular conductance-basedcompartmental models, often have numerous parameters thatcannot be directly determined experimentally and must beconstrained by an optimization procedure. A common prac-tice in evaluating the utility of such procedures is using apreviously developed model to generate surrogate data (e.g.,traces of spikes following step current pulses) and then chal-lenging the algorithm to recover the original parameters (e.g.,the value of maximal ion channel conductances) that wereused to generate the data. In this fashion, the success or failureof the model fitting procedure to find the original parameterscan be easily determined. Here we show that some modelfitting procedures that provide an excellent fit in the caseof such model-to-model comparisons provide ill-balancedresults when applied to experimental data. The main reason isthat surrogate and experimental data test different aspects ofthe algorithm’s function. When considering model-generated S. Druckmann(B) · I. SegevInterdisciplinary Center for Neural Computationand the Department of Neurobiology, Institute of Life Sciences,The Hebrew University of Jerusalem, Edmond J. Safra Campus,Givat Ram, 91904 Jerusalem, Israele-mail: [email protected]. Segeve-mail: [email protected]. K. Berger · S. Hill · F. Schürmann · H. MarkramBrain Mind Institute, Ecole Polytechnique Fédérale de Lausanne(EPFL), 1015 Lausanne, Switzerlande-mail: [email protected]. Hille-mail: [email protected]. Schürmanne-mail: [email protected]. Markrame-mail: [email protected] data, the algorithm is required to locate a perfectsolution that is known to exist. In contrast, when consideringexperimental target data, there is no guarantee that a perfectsolution is part of the search space. In this case, the optimiza-tion procedure must rank all imperfect approximations andultimately select the best approximation. This aspect is nottested at all when considering surrogate data since at leastone perfect solution is known to exist (the original parame-ters) making all approximations unnecessary. Furthermore,we demonstrate that distance functions based on extractinga set of features from the target data (such as time-to-first-spike, spike width, spike frequency, etc.)—rather than usingthe original data (e.g., the whole spike trace) as the target forfitting—are capable of finding imperfect solutions that aregood approximations of the experimental data.