Ephron S. Rosenzweig, Carol A. Barnes & Bruce L. McNaughton
 
The authors are in the Department of Psychology, Division of Neural Systems, Memory and Aging, Room 384 Life Sciences North Building, Tucson, Arizona 85724, USA
e-mail: bruce@nsma.arizona.edu

A new finding suggests that memory removal may be an active process, as blocking NMDA receptors after induction of long-term potentiation prolongs LTP and memory retention.

In this issue, Villarreal et al. report the effects of the NMDA-receptor antagonist CPP on the decay of LTP and the retention of spatial memory³. Because NMDA receptors are thought to be critical in the induction of many forms of LTP, many researchers have attempted to correlate the effects of NMDA-receptor antagonists on LTP and memory (for review, see ref. 4). In those studies, however, NMDA receptors were generally blocked before LTP induction or behavioral training, and the antagonist's ability to hinder LTP or memory was assessed. In contrast, Villarreal et al. found that an NMDA-receptor antagonist administered after LTP induction prolongs LTP. Similarly, when the drug is administered after training on a radial eight-arm maze, it enhances the retention of spatial memory in the task. These results, at the very least, strengthen the link between LTP and memory, because the drug treatment affects LTP and spatial memory in the same qualitative manner. Perhaps more importantly, the results help us understand the forgetting process at a synaptic level, as they suggest that memory decay is an active, NMDA-receptor-dependent process.

It is widely accepted that the brain uses associative modification of the strengths of the synaptic connections between neurons to store multiple memories in the same network. The principle can be most easily understood by considering the case of binary synapses whose strength (or 'weight') can take on values of 0 or 1. According to Hebb's theorem, simultaneous activation of presynaptic and postsynaptic cells leads to an increase of the synaptic weight from 0 to 1 (assuming it is not already 1)⁵. With the addition of some biologically plausible filtering properties (based on inhibitory interneurons, for example), such networks can retrieve complete versions of previously stored patterns from any unique subset of the original pattern⁶. If, for example, one had stored the pattern 101010 in a network, one could present the partial input 1010—, and the network would correctly retrieve the full original pattern. This phenomenon, called pattern completion, is an important characteristic of associative neural networks, but places important constraints on network storage capacity. In particular, the more patterns that are stored and/or the more overlap among the patterns, the more likely is the spurious activation of cells not belonging to the target pattern. One well-known partial solution to this problem is sparse encoding of the memories, which both reduces the rate at which synapses are used and reduces overlap among the stored patterns. Still, this may not be sufficient and it is likely that there exists some means for active removal of old memories.

Before discussing the possible mechanisms for memory erasure, we shall briefly review LTP, the process that is currently the best experimental approximation of memory encoding. In 1973, Bliss, Lomo and Gardner-Medwin observed a long-lasting increase of synaptic strength in the rabbit dentate gyrus in response to high-frequency stimulation¹. Their discovery set off a flurry of research to determine the mechanisms behind the process (which eventually became known as LTP) and the correlation, if any, of LTP with learning and memory. This process was shown early on⁷ to embody the associative principle postulated by Hebb. Several years later, the core mechanism of this associativity was shown to center on the NMDA receptor. Briefly, the NMDA receptor requires both its preferred ligand (glutamate) and a large membrane depolarization to allow Ca²⁺ ions to flow through the receptor and into the postsynaptic cell. The Ca²⁺ influx begins a cascade that results in both short-term and long-term changes at the synapse (for review, see ref. 8). The requirement of concurrent depolarization and glutamate binding satisfies the associativity principle.

The problem, of course, is that LTP only increases synaptic strength. Without a way to decrease synaptic strength, neural networks would soon experience the saturation problem alluded to above, in which too many synaptic weights are set to 1 and information stored in the network is degraded (Fig. 1a). For example, artificial saturation of LTP can lead to memory impairments⁹.

Long-term depression, or LTD, seems to be the necessary counterpart to LTP (for review, see ref. 10). Interestingly, like LTP, LTD is induced by Ca²⁺ flow through NMDA receptors. It seems that moderate levels of Ca²⁺ influx induce LTD, and high levels of Ca²⁺ influx induce LTP. According to one view, this occurs because the balance between net addition and net removal of AMPA receptors shifts as a function of intracellular Ca²⁺ concentration (for example, ref. 11). Thus, in theory, the same events that lead to strengthening of some synapses can lead to weakening of other, previously strengthened synapses, if they are not participants in the encoding of the new event. As new events are stored in the network, saturation of the network is prevented by the preferential erasure of previously encoded memories (Fig. 1b). Initial evidence in support of this model was recently provided, when it was observed that new spatial experiences could accelerate the decline of LTP previously established by electrical stimulation¹².

Villarreal et al.³ have addressed the next critical question: if the normal acquisition of new memories erases older ones, can one preserve a given memory by blocking the NMDA receptors after that memory has been encoded? The answer seems to be yes. Daily intraperitoneal injections of CPP, at doses shown by the experimenters to block LTP induction, reduced LTP decay in the dentate gyrus of awake, freely behaving rats. This was true whether the injections were begun 1 hour or 48 hours after LTP induction. Similar drug treatments also enhanced spatial memory retention on a radial 8-arm maze. The task required animals to remember, first, which arms had already been visited in a given session and, second, which four arms contained reward, as the same arms were baited on every session. Retention of the task was tested six days after completion of training. CPP-treated rats had fewer working-memory errors (re-entries into already visited arms) and reference-memory errors (entry into never-rewarded arms) than did controls. Figure 1c shows schematically how the manipulations used by Villarreal et al. might affect memory storage and erasure.

As is common in this field, the new results are not entirely consistent with some previous findings. Norris and Foster, for example, found no significant differences between drug and control groups¹³. However, they used MK-801, an uncompetitive NMDA-receptor antagonist with kinetics different from those of CPP. In addition, they tested their animals in the Morris water maze rather than the eight-arm radial maze, and the literature contains multiple examples of manipulations having different effects on these two tasks. Finally, the retention test was performed 24 hours after a single dose of MK-801. The retention enhancements observed by Villarreal et al. may appear only after a longer blockade of LTP, when the memories in the control animals have had sufficient time to decay. Another study found a mixture of improvements and impairments, depending on the exact timing of the drug administration¹⁴. In that study, animals were also tested in the water maze, and the differences in the drug administration schedule make comparison with Villarreal et al. difficult. Perhaps most crucially, the animals were treated with the drug LY326325, whose primary effect is the blockade of AMPA receptors, and only indirectly prevents Ca²⁺ influx through NMDA receptors. Such disparities in results and methods, though difficult to interpret at present, may eventually reveal new subtleties in the mechanisms of learning and memory.

A final note concerns the ongoing attempts to prove that LTP is the experimental analog of memory formation in the brain⁴. McNaughton and Morris proposed three critical experiments that would be necessary to support the hypothesis that LTP underlies some forms of memory¹⁵. The first was that saturation of LTP should disrupt existing memories and prevent further acquisition. The second was that blockade of LTP should impair acquisition without disrupting retention. The third was that selective erasure of LTP should disrupt retention without preventing further acquisition. While there is support for the first two conjectures, to our knowledge, no method for selective erasure has yet been achieved that would enable testing the third. The study by Villarreal et al.³, however, represents the complement to the third postulate: prevention of the naturally ocurring erasure of LTP prevents the loss of a memory. These results have thus added another critical piece of information to the puzzle of the relationship between synaptic plasticity mechanisms and memory dynamics.

REFERENCES