Imaging of brain structures in living mice reveals that learning new tasks leads to persistent remodelling of synaptic structures, with each new skill associated with a small and unique assembly of new synapses.
The notion that structural changes in brain circuitry underlie certain forms of learning is widely accepted, yet this belief has been frustratingly difficult to establish experimentally. Two studies, one by Xu et al.1 (page 915) and one by Yang et al.2 (page 920) published in this issue, provide compelling evidence that learning new motor tasks (and acquiring new sensory experiences) is associated with the formation of new sets of persistent synaptic connections in motor (and sensory) regions of the mouse brain. These findings suggest that synapse assemblies, rather than cell assemblies, might be viewed as the elementary entities (engrams) of stored memories.
At a basic level, the brain can be viewed as a vast network of neurons connected to each other by specialized structures known as synapses. Most synaptic connections are formed between axons (the slender and elongated extensions that carry the signals generated by neurons) and dendrites (highly branched extensions that are specialized for receiving signals originating in other neurons). Most axodendritic synapses are found on dendritic spines, which are tiny protrusions that extend from dendritic shafts in large numbers (Fig. 1a, overleaf).
It has long been assumed that structural changes in this complex circuitry provide the basis for long-term memory formation or the learning of new tasks3. The development of new imaging tools4 over the past decade has opened the door to an experimental evaluation of this assumption as it has allowed for longitudinal studies of axonal and dendritic morphology in the brain of living animals (mainly mice). Somewhat surprisingly, it has become evident that overall neuronal morphology is remarkably stable over long periods (months and beyond). What such studies have revealed, however, is that structural changes do occur at the level of individual synapses, manifested by the appearance of new dendritic spines and the disappearance of others over the course of hours and days3,5,6. Interestingly, the extent of spine remodelling can be altered by experimental manipulations, such as depriving the animals of sensory input (from the whiskers or eyes)3,5,6. Yet a direct relationship between spine remodelling and learning had yet to be demonstrated.
The two studies in this issue1,2 provide strong evidence for this link. In both studies, mice were trained to perform a new motor task (reaching for a single seed1 or remaining on an accelerating rotating rod2), and two-photon microscopy4 of the living animals was used to investigate whether successful training was associated with changes in spine remodelling beyond those observed in untrained mice. Both studies revealed that by the end of the first 1–2 days of training (and as soon as one hour after training1), twice as many new spines had appeared in the brain of trained mice compared with untrained mice. Continuous training was subsequently followed by increased rates of spine elimination, and so, after 1–2 weeks, total spine numbers did not differ between trained and untrained animals. Remarkably, behavioural performance correlated well with the numbers of new spines formed shortly after training (and with the extent of pre-existing-spine elimination) when these parameters were compared across all trained animals.
Yang and colleagues2 also examined how exposure to an enriched environment (altering patterns of bead strings hanging from cage tops) affected spine remodelling. In essence, spine remodelling was affected in the same manner. Here, however, remodelling was confined to particular regions of the cerebral cortex concerned with sensory input from the whiskers, unlike spine remodelling associated with motor task learning, which was confined to specific cerebral cortex regions concerned with forelimb movement1,2.
Although these findings strongly support the notion that some forms of learning are 'encoded' by changes in brain circuitry, several points warrant further discussion. In agreement with other studies3,5,6, most (96–98%) of the new spines in both trained and untrained animals were short-lived (days), with <1% persisting for many months2. Therefore, only a tiny fraction of the new spines formed over a 2-day period would be expected to survive for long durations. Yang and colleagues2 calculated that at the end of the mouse's life (assumed to be 36 months) these new spines would make up 0.04% of the total number of spines in the particular cortical areas examined. Although this tiny fraction is suggested to be a sufficiently large number of spines to encode a learned behaviour, one wonders how baseline changes in spine number, shown in the same studies1,2 to be about 3–5% per day (two orders of magnitude greater), can occur without some loss of coherent brain function.
One possible explanation might be that most short-lived spines are only weakly functional7,8. It is intriguing to speculate that this large pool of weak synapses could serve as a substrate for selective processes that would promote the stabilization and maturation of a minority of such synapses according to certain functionality criteria. These synapses would then survive and eventually underlie future behaviour. In this respect it is worth mentioning that training was not associated with the additional proliferation of dendritic filopodia1,2 (often considered to represent spine precursors9) but was associated with increased conversion of filopodia to spines (Fig. 1b, spine 1) and the subsequent stabilization of the converted structures1, which would be congruent with such selective3,9,10 processes. As compelling as this possibility is, however, the data are also consistent with the possibility that new spines formed through instructive processes, for example, by spine extension at particular locations in response to specific cues (Fig. 1b, spine 2).
Of particular interest is the finding that training-associated spine remodelling was not observed in mice that failed to learn the reaching task or could not reach the seeds1, or in animals exposed to a slowly rotating rod task2 (which does not require learning). Whereas the latter case can be explained by the failure of new spines to satisfy a certain short-term functionality criterion (such as causally linked pre- and postsynaptic activities11), the former can only be explained by a behaviourally driven spine formation and stabilization criterion, operating at longer timescales (hours and days, rather than minutes). The most likely criterion would be a reward-related one12 — spines would be stabilized only if their functionality resulted in rewarded actions.
This strongly implicates the involvement of diffuse modulatory brain systems — small groups of neurons located in specific brain regions (mainly the brainstem), whose projections cover large brain regions13. Most notable in this respect is the dopaminergic system, whose activation has been shown time and again to have key roles in various forms of reinforcement, or reward-driven learning. Indeed, it was recently reported14 that the elimination of dopaminergic terminals within the same cerebral cortex regions concerned with forelimb movement specifically impaired the acquisition of a motor skill (reaching for a food pellet, a task similar to that studied by Xu and colleagues1) but not the execution of a previously acquired skill. Novelty, another attribute ascribed to these modulatory brain systems, also seemed to be extremely important, as re-exposure of trained animals to the original training tasks did not lead to particular spine remodelling, whereas learning new motor tasks did1,2.
Given the sparseness of the task-related synapse groups1,2, it is not surprising that Xu et al. found that the overlap between synapse groups related to two different tasks was negligible (Fig. 1c). This finding, the fact that most task-related remodelling involved new spines, and the lack of clear size changes in pre-existing spines following training1, would seem to be at odds with certain tenets of the cell-assembly hypothesis of memory11,15. The cell-assembly hypothesis assumes that what distinguishes one memory trace from another is the composition of cells that are co-activated within a given network. Most notably, this hypothesis posits that these distinct compositions are created during memory formation by retuning the strengths of all the synapses already in the network, along a continuum of synaptic strengths. The new results, however, point to strong associations between learned tasks and specific, non-overlapping groups of novel synapses whose strength might remain relatively constant once formed. These 'synapse assemblies' might thus be viewed as the fundamental engrams for these specific motor tasks. Synapse assemblies in a given network would hold the keys to the functioning of that network; for every learned task, the network would function with a connectivity scheme that had been formed and optimized specifically for that task.
Clearly, further studies are required to determine whether the experience-associated spine remodelling described here can be generalized to other forms of memory and other brain systems. Investigating the involvement of diffuse modulatory systems in the process is another intriguing challenge. And although it remains to be shown conclusively that these forms of spine remodelling are essential components of long-term learning and not merely distant echoes of other, yet to be discovered processes, these exciting studies make a convincing case for a structural basis to skill learning and reopen the field for new theories of memory formation.
Xu, T. et al. Nature 462, 915–919 (2009).
Yang, G., Pan, F. & Gan, W.-B. Nature 462, 920–924 (2009).
Holtmaat, A. & Svoboda, K. Nature Rev. Neurosci. 10, 647–658 (2009).
Denk, W., Strickler, J. H. & Webb, W. W. Science 248, 73–76 (1990).
Bhatt, D. H., Zhang, S. & Gan, W. B. Annu. Rev. Physiol. 71, 261–282 (2009).
Alvarez, V. A. & Sabatini, B. L. Annu. Rev. Neurosci. 30, 79–97 (2007).
Knott, G. W., Holtmaat, A., Wilbrecht, L., Welker, E. & Svoboda, K. Nature Neurosci. 9, 1117–1124 (2006).
Nägerl, U. V., Köstinger, G., Anderson, J. C., Martin, K. A. & Bonhoeffer, T. J. Neurosci. 27, 8149–8156 (2007).
Jontes, J. D. & Smith, S. J. Neuron 27, 11–14 (2000).
Holtmaat, A., Wilbrecht, L., Knott, G. W., Welker, E. & Svoboda, K. Nature 441, 979–983 (2006).
Hebb, D. O. A Neuropsychological Theory (Wiley, 1949).
Crow, T. J. Nature 219, 736–737 (1968).
Kety, S. S. in Neurosciences: Second Study Program (ed. Schmitt, F. O.) 324–336 (Rockefeller Univ. Press, 1970).
Molina-Luna, K., Pekanovic, A., Röhrich, S., Hertler, B. & Schubring-Giese, M. PLoS ONE 4, e7082 (2009).
Hopfield, J. J. Proc. Natl Acad. Sci. USA 79, 2554–2558 (1982).
About this article
Functional Brain Network Topology Discriminates between Patients with Minimally Conscious State and Unresponsive Wakefulness Syndrome
Journal of Clinical Medicine (2019)
Pioneering topological methods for network-based drug–target prediction by exploiting a brain-network self-organization theory
Briefings in Bioinformatics (2018)
Modelling Self-Organization in Complex Networks Via a Brain-Inspired Network Automata Theory Improves Link Reliability in Protein Interactomes
Scientific Reports (2018)
Can local-community-paradigm and epitopological learning enhance our understanding of how local brain connectivity is able to process, learn and memorize chronic pain?
Applied Network Science (2017)
Comparison of Local Information Indices Applied in Resting State Functional Brain Network Connectivity Prediction
Frontiers in Neuroscience (2016)