Despite great advances in our knowledge of the mechanisms underlying synaptic plasticity, there are gaps in our understanding of learning and memory between the cellular and systems levels. In the last 5 years, my lab has made three advances at the synaptic, circuit, and systems levels that are beginning to fill in some of these gaps.
Nanoconnectomic Upper Bound on the Variability of Synaptic Plasticity
In collaboration with Kristen Harris and Chandra Bajaj at the University of Texas at Austin, we reconstructed a 6x6x5 mm3 piece of the rat CA1 hippocampal neuropil from electron microscopic (EM) sections. Our goal was to achieve the highest fidelity geometric reconstruction possible, including a restoration of the extracellular space, which is lost during dehydration of the tissue. (See below for a video of the reconstruction.) Our reconstruction is accurate at the nanometer scale to keep track of biochemical interactions in nanoscale compartments, whereas most microconnectomic reconstructions only achieve a wiring diagram.
The release of neurotransmitter at excitatory synapses in the cerebral cortex and the hippocampus is probabilistic and because synaptic plasticity depends on the history of presynaptic and postsynaptic activity, the amount of information that can be stored at a synapse in the form of a retrievable synaptic efficacy is limited. The trial-by-trial variability may, however, mask a more reliable intrinsic precision. We approached this question by taking advantage of a previously established correlation between the size of the spine head on pyramidal dendrites and the efficacy of a synapse, which allowed us to estimate the variability of synaptic plasticity.
In our EM reconstruction of hippocampal neuropil we found 10 single axons making two or more synaptic contacts onto the same dendrites, having shared histories of presynaptic and postsynaptic activity (Figure 1). The postsynaptic spine heads, but not the spine necks, of these pairs were nearly identical in size, for the small as well as the large spine heads. This finding was confirmed in two other rats. This unexpectedly low variability allows 26 different values of the synaptic strength to be distinguished, corresponding to 4.7 bits of information, across synapses that vary in size over a factor of 60. This is an order of magnitude greater precision than previous estimates and requires postsynaptic averaging over minutes of activity because of stochastic variability. This raises two interesting questions: How is this precision achieved at the molecular level and why is the precision so high if most release probabilities are so low?
Selective Memory Generalization by Spatial Patterning of Protein Synthesis
Protein synthesis is necessary to form long-term memories and mRNA and ribosomes are abundant within dendrites regulate the long-term anatomical changes of synaptic efficacy. The time scale for these changes is hours, much longer than the millisecond time scale of electrical signals. Most neural network models ignore these longer time scales. We asked whether new computational principles might emerge from these long time scales.
We developed a computational model of synaptic plasticity that included synaptic tagging and capture, in which a weakly potentiated synapse can be “rescued” by a nearby synapse that is strongly potentiated within an hour. Spatially patterned protein synthesis within dendrites can enable selective consolidation of some memories but forgetting of others, even for simultaneous events that are represented by the same neural population. Key factors regulating selectivity include the functional clustering of synapses on dendrites and the sparsity and overlap of neural activity patterns at the circuit level. Based on these findings we proposed a novel two-step model for selective memory generalization during rapid eye movement (REM) and slow-wave sleep. Experimental confirmation for the branch-specific formation of dendritic spines during sleep, appeared shortly after it was published.
Large-Scale Spatiotemporal Structure of Thalamocortical Spindle Oscillations
Sleep, which consumes about a third of our lives, is thought to have a critical role in memory processes. The effects of sleep on memory consolidation were first observed in behavioral studies where there was an improvement in the performance for various memory tasks after sleep compared to a similar awake period. This improvement has been observed in declarative, procedural and emotional memory tasks. Although behavioral studies have shown a significant effect of sleep on memories, the neural mechanisms are not known. During slow-wave sleep, the cortex is decoupled from external inputs and can be devoted to consolidating previously acquired labile memories into stable memories. Recently, memory replay has been demonstrated during sleep, and associated with characteristic oscillations, giving rise to the hypothesis that these may form the critical neural substrate for memory consolidation.
Electrocorticographic (ECoG) recordings from the surface of the cortex in humans are used to locate the focus of seizures in epilepsy patients. In collaboration with Syd Cash at the Massachusetts General Hospital in Boston, we have analyzed human ECoG recordings during sleep. We were particularly interested in sleep spindles that occur during stage 2 of non-REM sleep. Spindles are 10-14 Hz repetitive bursts of activity that originate in the thalamus last for 1-2 seconds and entrain the cortex. Sleep spindles are known to be critical for sleep-dependent consolidation of long-term memory. My laboratory has developed detailed biophysical models of the ionic mechanisms underlying the generation and termination of sleep spindles.
Spindle activity was thought to occur synchronously across the cortex. We developed new analysis methods to track the phase of traveling waves in the cortex and found that that cortical spindles form repeating circular wave-like patterns across the entire cortex. The short-range association fibers could support this global pattern neuronal activity, with time delays of tens of milliseconds between neighboring regions, a timescale highly relevant for spike-time dependent plasticity (STDP): If spindles were synchronous across the cortex, then spikes from nearby regions of cortex would arrive after a delay of 5-10 ms, within the range for long-term depression from STDP. But with a circular wave the delayed spikes from one area would arrive at about the same time because of the phase delay in the traveling wave. STDP is effective when repeated at 10-20 Hz, in the range of spindle frequencies. Spindle oscillations could therefore serve as a system for globally organizing sleep-dependent memory consolidation.
This research was supported in part by the National Institutes of Health, the Office of Naval and the National Science Foundation.
As of April 22, 2016