Scientific Understanding of Consciousness
Consciousness as an Emergent Property of Thalamocortical Activity

Memory Retrieval Via Prefrontal Cortex Projection to the Hippocampus


Nature  526, 653–659 (29 October 2015)

Projections from neocortex mediate top-down control of memory retrieval

Priyamvada Rajasethupathy,

Department of Bioengineering, Stanford University, Stanford, California 94305, USA

CNC Program, Stanford University, Stanford, California 94305, USA

Neuroscience Program, Stanford University, Stanford, California 94305, USA

Department of Psychiatry and Behavioral Sciences, Stanford University, Stanford, California 94305, USA

Howard Hughes Medical Institute, Stanford University, Stanford, California 94305, USA


Top-down prefrontal cortex inputs to the hippocampus have been hypothesized to be important in memory consolidation, retrieval, and the pathophysiology of major psychiatric diseases; however, no such direct projections have been identified and functionally described. Here we report the discovery of a monosynaptic prefrontal cortex (predominantly anterior cingulate) to hippocampus (CA3 to CA1 region) projection in mice, and find that optogenetic manipulation of this projection (here termed AC–CA) is capable of eliciting contextual memory retrieval. To explore the network mechanisms of this process, we developed and applied tools to observe cellular-resolution neural activity in the hippocampus while stimulating AC–CA projections during memory retrieval in mice behaving in virtual-reality environments. Using this approach, we found that learning drives the emergence of a sparse class of neurons in CA2/CA3 that are highly correlated with the local network and that lead synchronous population activity events; these neurons are then preferentially recruited by the AC–CA projection during memory retrieval. These findings reveal a sparsely implemented memory retrieval mechanism in the hippocampus that operates via direct top-down prefrontal input, with implications for the patterning and storage of salient memory representations.

AC–CA: a direct top-down projection 

To identify direct top-down inputs to the hippocampus, we injected a retrograde tracer capable of labelling afferent neurons with tdTomato (RV-tdT) into the hippocampus. We observed robust tdT labelling in brain regions with known inputs to the hippocampus, including the medial septum, contralateral CA3 and entorhinal cortex. Additionally, we identified a previously uncharacterized input arising from the dorsal anterior cingulate cortex (AC) and adjacent frontal cortical association cortex, both of which are reciprocally connected with the mediodorsal thalamic nucleus—a defining feature of the prefrontal cortex (PFC) in rodents. Injection of RV-tdT in the AC also sparsely labelled neurons bilaterally in the dorsal hippocampus, consistent with potential bidirectional communication between the AC and hippocampus. To validate further the existence of this novel PFC-to-hippocampus projection, we injected an anterograde label (adeno-associated virus 5-enhanced yellow fluorescent protein (AAV5-eYFP)) into the dorsal AC and detected fluorescence-filled projection terminals bilaterally in the striatum and ipsilaterally in the medial dorsal thalamic nucleus (both areas are known to receive projections from the PFC), but also bilaterally in the hippocampus.

To determine if these prefrontal projections gave rise to direct monosynaptic drive of hippocampal neurons, we transduced the AC with an AAV encoding a channelrhodopsin (ChR), and performed patch-clamp recordings of light-driven excitatory postsynaptic currents (EPSCs) in CA1/CA3 cell bodies. Cells in both CA1 and CA3 reliably responded to light pulse trains, and generated evoked EPSC amplitudes sufficient to drive action potentials. Responses when present were fast, with mean latency of 3.2ms in CA1 (n = 26) and 2.7ms in CA3 (n = 13); along with the observation of sustained evoked spikes after 10Hz stimulation. This finding was consistent with the presence of a direct and efficacious monosynaptic connection from the AC onto hippocampal pyramidal cells in the CA3/CA1 subfields, which we accordingly term the AC–CA projection.

We have identified a direct projection from the AC to the hippocampus with properties that mediate retrieval of recently encoded memory traces. The ability of the AC to select appropriate targets among hippocampal pyramidal neurons during context retrieval points to potential reciprocity between these two regions; the hippocampus is well positioned to inform the AC regarding context through bottom-up pathways, after which the AC can access and mobilize engrams in the hippocampus for top-down control of retrieval. The AC may also form an independent representation during training, as suggested by the speed (within 1 day post-training) with which the AC–CA projection can mediate top-down retrieval. This rapid retrieval also demonstrates that the AC is engaged early during the memory encoding process, as has been suggested, which, together with the previous finding of hippocampal engagement even at later time points, suggests a shift towards viewing the AC and hippocampus, and their bidirectional communication, as having shared involvement in both early and late stages of memory. The speed and specificity of the top-down retrieval also underscores likely plasticity at the AC–CA terminals for dynamic access to recently allocated neurons in the hippocampus.

We have found that this top-down control over contextual memory involves preferential recruitment of a sparse set of hippocampal neurons that are highly correlated in the local network and that tend to lead population-wide synchronous events, but only in the memory context and in a manner only seen after training. Sparsification of memory-associated networks during learning has been observed and computationally predicted, since hierarchical networks with few extensively connected hub-like nodes, as observed here, are well suited for memory stability (random degradation of the network will probably affect non-hub nodes that are less consequential to the memory representation). But beyond memory stability, the emergence of contextual memory-specific HC neurons could facilitate efficient access to engrams, in a process suitable for memories that are strongly encoded via repetition, reward, or emotional saliency. This circuit property could be broadly relevant, helping to explain how even single cortical neurons can in some cases drive network activity and behaviour.