Estimating robustness and parameter distribution in compartmental models of neurons Noam Peled 1 and Alon Korngreen 1 1 The Mina and Everard Goodman Faculty of Life Sciences and the Leslie and Susan Gonda Multidisciplinary Brain Research Center, Bar-Ilan University, Ramat-Gan, Israel. (e-mail: korngra@mail.biu.ac.il ) Abstract. Our understanding of the input-output function of single neurons has been advanced by biophysically accurate multi-compartmental models. The large number of free parameters in these models requires the use of automated approaches for finding their optimal values in a very multi dimensional parameter-plane. Due to inherent noise in measuring equipment and surroundings determination of the accuracy of the obtained parameters is near impossible. Here we show that finding the parameters distribution via Monte Carlo approach simulations reveals the sensitivity of each parameter to noise. This method allows for the reduction of the complexity of the parameter-plane which in turn enables finding the sensitivity of the model to each parameter and the sensitivity of each parameter to noise. Keywords: Multi-compartmental models, Monte Carlo, Robustness, Noise sensitivity, Global minimum 1 Introduction The properties and functions of neurons in the CNS have been intensively studied over the past decade [Johnston et al., 2003; Migliore and Shepherd 2002; Stuart et al. 1999] providing exciting new information that revived the discussion about the computational properties of single neurons [Häusser and Mel, 2003; Mel, 2002; Poirazi and Mel, 2001; Polsky et al. 2004]. Clearly, understanding complex neuronal computations requires functional models which must take into account several types of neuronal properties, e.g., voltage-gated conductances, conductances distributions etc. This complexity resulted in many parameters in neuronal models being tuned by hand [Rhodes and Llinás, 2001; Schaefer et al., 2003]. However, the large number of free parameters causes these models to be ill constrained [Rhodes and Llinás 2001; Schaefer et al., 2003]. Previous studies have shown that parameter search methods can be effective in finding matches between single-neuron models and a target data set [Eichler West, 1996; Vanier and Bower, 1999]. Nevertheless, not all parameter search methods perform equally well, and the relative performance of the different methods depends on the model being optimized. Stochastic algorithms such as simulated annealing [Kirkpatrick et al., 1983] and genetic algorithms (GAs) [Mitchell, 1996] have been