The structural aspects of neural dynamics and information flow.

Abstract

Neurons have specialized structures that facilitate information transfer using electrical and chemical signals. Within the perspective of neural computation, the neuronal structure is an important prerequisite for the versatile computational capabilities of neurons resulting from the integration of diverse synaptic input patterns, complex interactions among the passive and active dendritic local currents, and the interplay between dendrite and soma to generate action potential output. For this, characterization of the relationship between the structure and neuronal spike dynamics could provide essential information about the cellular-level mechanism supporting neural computations. This work describes simulations and an information-theoretic analysis to investigate how specific neuronal structure affects neural dynamics and information processing. Correlation analysis on the Allen Cell Types Database reveals biologically relevant structural features that determine neural dynamics-eight highly correlated structural features are selected as the primary set for characterizing neuronal structures. These features are used to characterize biophysically realistic multi-compartment mathematical models for primary neurons in the direct and indirect hippocampal pathways consisting of the pyramidal cells of Cornu Ammonis 1 (CA1) and CA3 and the granule cell in the dentate gyrus (DG). Simulations reveal that the dynamics of these neurons vary depending on their specialized structures and are highly sensitive to structural modifications. Information-theoretic analysis confirms that structural factors are critical for versatile neural information processing at a single-cell and a neural circuit level; not only basic AND/OR but also linearly non-separable XOR functions can be explained within the information-theoretic framework. Providing quantitative information on the relationship between the structure and the dynamics/information flow of neurons, this work would help us understand the design and coding principles of biological neurons and may be beneficial for designing biologically plausible neuron models for artificial intelligence (AI) systems.