Systematic validation of common approaches for cortical network modeling

The ultimate goal of Computational Neuroscience is the development of models with high predictive power, capable of explaining brain dynamics and operational principles under different circumstances. To achieve this, models should be developed with the aim to simultaneously incorporate the increasing number of experimentally observed features. In Computational Neuroscience, however, the common approach is to construct spiking neural network models, which are focused on a narrow set of phenomena, while the model validation on the independent set of features is typically omitted. This often results in the partly unrealistic network properties and dynamics, in particular when the model is subjected to a more thorough analysis. Such phenomenological modeling reduces the model predictive reliability and makes it hard to merge multiple highly specialized models into an unified model of the brain. Furthermore, since the first spike-based cortical models were introduced almost three decades ago, the significant fraction of the published biologically-realistic cortical spiking models can be reliably assigned to one of the two classes. The most prominent class includes the model variations of the classic balance-random network (BRN, Brunel 2000) and is based on the implementation of quasi-random synaptic connectivity structure, where unstructured stochastic synaptic input occasionally drives neurons towards spiking threshold. The second numerous model class includes the variations of the “synfire” chain model (Abeles, 1991) and is based on the implementation of structured synaptic map, where neurons are driven towards the spiking threshold by the excitation originating from the coordinated activation of pre-synaptic neurons. These two model classes are orthogonal with respect to the basic mechanism underlying network dynamics, thus making it possible, that the models of at least one class are based on the biologically-implausible assumption. The latter would reduce the predictive reliability of those models. Experimental evidences supporting either of these two model classes are vague, making it difficult to resolve which of them better represents cortical dynamics. To address these issues, I systematically analyze the most conspicuous structural and dynamical features of cortical circuits, and derive a set of criteria, which can be further used to constrain future cortical models, thus making their predictions more reliable. Then, I select three prominent computational models, including BRN, and evaluate them on the basis of these criteria. I show, that each model possesses some of these features, but none combines all features with parameters constrained by biology. Specifically, the crucial role of external input in sustaining activity in biology-constrained BRN model contradicts experimental observations. Because of this draw back in the performance, I search for additional experimental evidences in support of the stochasticity and/or coordination of input. I show, that neuronal membrane potential dynamics in cortical circuits in vivo possess the features characteristic for coordinated, but not stochastic input. Furthermore, based on the coordinated input approach, I suggest the biologically-realistic spiking network model of local cortical circuit, that satisfies all the derived criteria and outperforms BRN. This model is also capable of generating Up-Down oscillations, propagating across the model in awave-like fashion, resembling cortical waves. In contrast to other models capable of generating Upstates or propagating waves, the present model does not rely on unrealistic neuronal or synaptic properties. I conclude, that a systematic constraint of computational models is important to increase their predictive power. For example, the dominant role of synaptic input stochasticity in neuronal spiking, underlying the majority of the reported spiking cortical models, was not systematically verified. The presented here analysis of experimental data, as well as revealed difficulties of networks with unstructured synaptic input to explain network excitability and smallness of membrane potential fluctuations within a biology-plausible parameter range, argue against the major role of synaptic input stochasticity in driving neuronal spiking. Therefore, the predictive reliability of the models with unstructured synaptic input is questionable and requires thorough investigation. In turn, coordination of synaptic input finds additional support in this work. The here presented model of Up-Down oscillations can serve as an example of how multiple experimentally observed structural and dynamical cortical features can be systematically integrated into a single model