Several distinct patterns of population activity can be recorded in the hippocampus in vivo. These neural activity patterns depend on the behavioral state of the animal, and include theta-modulated gamma oscillations as well as low-rate irregular activity with periodically occurring large-amplitude sharp wave-ripple (SWR) events. During SWRs, neuronal populations in the hippocampus have been found to "replay", on a faster time scale, activity recorded during theta-gamma activity in the exploring animal. Such replay may be important for the establishment, maintenance and consolidation of long-term memory. Our aim was to develop a mechanistic understanding of cellular and network mechanisms underlying the generation of SWRs, spatio-temporal sequence replay during SWRs, and the observed switching to other types of population dynamics such as gamma oscillations, based on in vitro and in vivo experimental observations.

A recently developed hippocampal slice preparation, in which SWRs arise spontaneously, has allowed the collection of a large and diverse set of data regarding the properties of SWRs, as well as the characterization of several cell types and synapses which are critical in their generation. The same types of data were collected after the addition of a cholinergic agonist, which alters the global dynamics, giving rise to gamma oscillations. Directly measured properties of cells and synapses were used to construct full-scale models of area CA3 of a mouse hippocampal slice in specific modulatory states. The model consisted of 8000 pyramidal cells (PCs), 150 parvalbumin-containing basket cells (PVBCs), and 100 other interneurons, all modeled as adaptive exponential integrate-and-fire neurons, whose parameters were tuned using our recently developed software tool, which facilitates the optimization of neural models to fit experimental observations.

We used a combination of experimental and theoretical tools including physiology, pharmacology, optogenetics, analytical calculations and computer simulations, to understand the dynamics of the CA3 network in the SWR state and during gamma oscillations. Excitatory and inhibitory activity were tightly coordinated in both networks states, and the firing of PCs remained sparse, while PVBCs discharged at higher rates, strongly coupled to the local field potential. Several lines of evidence indicated that these two cell types were the critical players in both SWRs and gamma oscillations, but the mechanisms responsible for the generation of fast network oscillations were markedly different in the two states. While a feedback loop between PCs and PVBCs underlay the generation of gamma oscillations in area CA3, ripple oscillations were generated mainly through synchronization of reciprocally connected PVBCs driven to fire at high rates by increased activity in the PC population. The initiation of SWR occurred stochastically, and normally required a critical amount of synchronous firing followed by exponential build-up of activity in the excitatory recurrent network of CA3 PCs, but could also be triggered by optogenetic activation of parvalbumin-containing neurons.

In simulations where all relevant neuronal populations were randomly connected with uniform or randomly varying weights with the experimentally measured mean, we observed stochastically initiated population bursts with associated ripple-frequency oscillations, but PCs were activated non-selectively at non-physiological high rates during these events. Introducing structure in the form of more strongly interacting PC assemblies led to selective activity and lower average PC rates during SWRs, but participating PCs still fired at unrealistically high rates. When cellular and synaptic properties were altered to reflect measurements made under a high cholinergic tone, the model network switched to globally coherent gamma oscillations, in agreement with experimental observations.

When we used spike-timing-dependent plasticity with simulated place cell activity during exploration to set up the strength of recurrent synapses, the model network produced spontaneous SWRs during which learned place cell sequences were replayed on a faster time scale, and the level of activity was also physiological. We then manipulated the synaptic weight matrix in various ways to reveal which aspects of its structure were critical in determining the stability of global activity levels, the emergence of coherent fast oscillations, and the generation of activity sequences. When we shuffled the synaptic weights in ways that preserved their local statistics but changed the global structure, replay was eliminated as expected, but steady-state firing rates and the properties of oscillations were also radically altered. More sophisticated manipulations of the weight matrix allowed us to selectively probe the effects of different aspects of its structure on the network dynamics. We concluded that the detailed structure of weights, as established during learning, is not only essential for the expression of meaningful neural representations, but is also a major determinant of population-level dynamics in cortical networks, and is, in particular, critical in producing the experimentally measured characteristics of SWRs in the hippocampus.