Single neurons have RAM-like activity

In his epic poem Visions of the Daughters of Albion, William Blake asks: "Where goest thou O thought? To what remote land is thy flight?" More than two centuries later, memory remains as one of the enduring mysteries of neuroscience, and despite decades of intensive research using modern techniques, we still have no answer to the questions posed by Blake.

Traditionally, memory has been regarded as consisting of several distinct processes or storage systems. Short-term memory (sometimes referred to as working memory) stores information that is required for the task at hand, but is severely limited in its capacity and duration. Long-term memory serves as a relatively permanent store of information, and information from the short-term store can be transferred to it by the process of rehearsal.

This distinction still holds true, but most of the research to date has focused on long-term memory. The prevailing view is that memories are encoded over long periods of time by the strengthening of connections in distributed networks of nerve cells, and we now have some understanding of the underlying mechanisms. However, very little is known about how nerve cells encode information for short periods of time.

Now researchers have identified a cellular memory mechanism whereby single neurons in the prefrontal cortex of mice can maintain persistent activity for short periods of time, much like a computer's random access memory (RAM), which temporarily stores data to improve the machine's performance. They suggest that this mechanism underlies working memory, and that their findings may lead new treatments for drug addiction,  and other conditions in which working memory and decision-making processes are disrupted.

In the study, Donald Cooper and his colleagues investigated the functional properties of nerve cell proteins called metabotropic glutamate receptors (mGluRs). The word metabotropic denotes the mechanism of action of these receptors. mGluR molecules are embedded in the cell membrane and associated with a G-protein, which is tethered to that part of the receptor which is exposed to the inside of the cell. Binding of the neurotransmitter glutamate to the receptor causes uncoupling of the G-protein, which then activates various biochemical signalling cascades, known as second messenger systems, within the cell.

mGluRs and other metabotropic receptors mediate slow neurotransmission, because the signalling pathways they activate take many seconds or minutes to exert their effects on a cell. This mechanism differs from that of the ion channel proteins responsible for generating nervous impulses. Although ion channels are also activated by the binding of a neurotransmitter molecule, they are deactivated within approximately one thousandth of a second; this quick inactivation is essential for the proper functioning of the neuron, as it enables the cell to generate trains of up to 1,000 electrical impulses per second.

The researchers used a technique called patch-clamping to electrically stimulate single neurons in slices of prefrontal cortex dissected from the mouse brain, while at the same time recording their responses to the stimulation. This involves impaling each cell with two microelectrodes. One of these (the stimulating electrode) can be used to inject a solution of ions into the cell. Because the solution carries a current, this alters the electrical properties of the cell by activating membrane proteins which are sensitive to changes in the membrane voltage. The other electrode is attached to a tiny patch of membrane on the outside of the cell, and forms a tight seal, so that the flow of currents through proteins in the membrane, which changes in response to the applied current, can be measured.

In these experiments, small currents which would not normally lead to a nervous impulse were found to elicit stable, low frequency electrical activity in the cells, which persisted for up to a minute after the injection of the current had ceased. When the experiments were repeated using cells from mutant mice lacking mGluR1, the same activity was still observed, ruling out the involvement of this receptor subtype. Then, by using chemicals which either inhibit or enhance other receptors, the researchers determined that the mGluR5 receptor subtype was responsible for the novel activity that they had observed. Further investigation revealed that the activity of this receptor is reduced by the neurotransmitter dopamine, via a second messenger system which involves a molecule called protein kinase A (PKA).

The  prefrontal cortex is involved in cognitive processes such as attention, decision-making and working memory, and chronic use of cocaine and other psychostimulants adversely affects this part of the brain. This causes impairments in the cognitive functions it performs - for example, cocaine addicts are known to choose immediate rewards which carry high risks. Cooper's group had previously found that repeated administration of cocaine and short-term withdrawal from the drug leads to increased PKA activity in neurons of prefrontal cortex. They therefore investigated if the drug would have any effect on mGluR5 activity and, sure enough, found that short-term cocaine withdrawal significantly reduced its ability to convert small electrical stimuli into a stable train of nervous impulses.

Studies carried out on monkeys in the 1970s revealed that neurons in the prefrontal cortex become active in working memory tasks, during the "delay period" which follows the presentation of a stimulus but precedes the behavioural response elicited by it. They are therefore believed to transiently encode information that is relevant to the tasks, and this study describes a candidate mechanism by which they do so. The new findings suggest that cells in layer 5 of the prefrontal cortex function as a temporary memory storage buffer, by converting brief stimuli into neuronal activity which persists for about one minute and then decays over a period of several seconds.

As well drug addiction, disruption of the interactions between mGluR5 and dopamine receptors is implicated in schizophrenia and attention deficit disorder (ADD). Drugs which enhance mGluR1 activity may therefore maintain the transient memory traces thought to be encoded in these cells, and possibly enable patients with these conditions to remain focused on what they are doing, ignore their impulses and make better decisions.   


Sidiropoulou, K. et al (2009). Dopamine modulates an mGluR5-mediated depolarization underlying prefrontal persistent activity. Nat. Neurosci. DOI: 10.1038/nn.2245.

More like this

This essay I wrote was shortlisted in the Association of British Science Writers competition in 2002. It was the first thing I posted on the old blog. It was written as an introduction to what were generally believed to be the fundamentals of brain function, starting from the molecular level and…
Continuing with my BIO101 lecture notes (May 08, 2006). As always, please correct my errors and make suggestions in the comments. -------------------------------------------------BIO101 - Bora Zivkovic - Lecture 1 - Part 4 Cell-cell interactions Cells do not exist in complete isolation. For a…
Learning and memory are widely thought to involve long-term potentiation (LTP), a form of synaptic plasticity in which a neuron's response to the chemical signals it receives is enhanced. This leads to a strengthening of the neuronal circuit, so that the memory encoded in the circuit can persist…
Optogenetics is a newly developed technique based on a group of light-sensitive proteins called channelrhodopsins, which were isolated recently from various species of micro-organism. Although relatively new, this technique has already proven to be extremely powerful, because channelrhodopsins can…

Neat! Being more electrically- than neurologically-inclined, I've always liked to think of neurons and memory as being similar to RAM, or some of the related nanotech like single-electron memories (using qubits) or molecular memories.


I think we ought to be very careful with such analogies. Learning and information processing in neural networks is fundamentally different from linear computation on von Neumann architectures.

The 'information' that is changed in the system when it learns new information is not stored in units, but in the distributions of synaptic weights. There is no language of thought or hierarchy of languages like binary, machine code etc... the processing is massively parallel.

Interestingly, we can prove mathematically that a sufficiently complex neural net with real-valued synaptic weights is turing complete - and that a neural network with irrational or complex numbered synaptic weight values has trans-turing capacity.

The insights from this article might be an inspiration for new computational models... usually, biological networks are modeled simply as ordered nodes with biases, synaptic connections with specific weights, and activation functions (or, more recently, as self-organizing maps, growing self-organising maps, support vector machines, with spin-glass models etc)

(I recommend Simon Haykin's wonderful "Neural Networks: A comprehensive foundation" for a basic introduction into the information processing of neural networks)

But with these insights, maybe one could develop an improved neural network-model, where the nodes act as 2 gateways at once - one for high-speed traffic and one for low-speed traffic. Since the connectivity-structure determines the information processesing, this means that essentially the same information can be processed by these two pathways... though I'm just musing here.


Recently someone I attended a seminar on neurophilosophy with (someone studying for his PhD in systemic neuroscience) told me that a guy he know did some research on the patch-clamp method... and found that the scales are small enough that the metal from the microelectrodes changes the overall elecrical properties of the cell so that the measurement might be affected significantly.

If repeatable, such observations could mean that we have, for a long time now, been using methodologically distorting methods to gather data. Ah well, seeing how useful it has proven already it seems we can handle it.

Cool study. Though, there is lots of evidence that you could have working memory without the PFC, so this mechanism can't be the only short-term memory buffer (or it exists in other brain areas also).

@Alan: Thanks - it's one of my favourite poems, so I'll think of any excuse to quote from it.


Could you please suggest me some papers showing evidences of working memory without PFC? Are they maybe ausiliary pathways which can arise for example in temporal lobe after a selective lesion to PFC in neuropsychological patients?

Thank you very much.

Quite an informative article. As to the question raised by MPhil there's obviously a possibility of an 'uncertainty' of Heisenberg type, given the nanoscale nature of memory networks. Though von Neumann architecture is apparently not seen, a Hopfield type networking is evident in the hippocampus. I have elaborated the DRAM like functioning in my article 'LTP Ensures That Memories Are Forever ' and also gone, may be a bit too far in speculating that memories may be forever in 'Do We Really Forget? Fathoming The Esoteric Realms of Memory '. Here I have resorted to the basic principles of quantum mechanics to come to such a far fetched possibility.