Glutamate release is inhibitory and unnecessary for the long-term potentiation of presynaptic function

Long-term potentiation (LTP) and long-term depression (LTD) of transmitter release probability (Pr) are thought to be triggered by the activation of glutamate receptors. Here, we demonstrate that glutamate release at CA3-CA1 synapses is in fact inhibitory and unnecessary for increases in Pr. Instead, at active presynaptic terminals, postsynaptic depolarization alone can increase Pr by promoting the release of nitric oxide from neuronal dendrites in a manner dependent on L-type voltage-gated Ca2+ channels. The release of glutamate, in contrast, decreases Pr by activating presynaptic NMDA receptors (NMDAR). Thus, net changes in Pr are determined by the combined effect of both LTP-promoting and LTD-promoting processes, that is, by the amount of glutamate release and postsynaptic depolarization that accompany presynaptic activity, respectively. Neither of these processes directly depends on the activation of postsynaptic NMDARs. We further show that presynaptic changes can be captured by a simple mathematical framework, in which the role of presynaptic plasticity is to ensure that the ability for a presynaptic terminal to release glutamate is matched with its ability to predict postsynaptic spiking.


Introduction
Learning and memory are thought to require synaptic plasticity, which refers to the capacity for synaptic connections in the brain to change with experience. The most frequently studied forms of synaptic plasticity are long-term potentiation (LTP) and longterm depression (LTD), which involve long-lasting increases and decreases in synaptic transmission. LTP and LTD can be expressed either postsynaptically, as changes in AMPA receptor (AMPAR) number or presynaptically, as changes in glutamate release probability (P r ) [1][2][3][4][5][6]. Traditionally, postsynaptic NMDA receptor (NMDAR) activation is believed to be important for both pre-and post-synaptic forms of plasticity [2,7]. Postsynaptic changes, in particular, have been causally and convincingly linked to NMDAR-dependent Ca 2+ influx which, via the activation of postsynaptic Ca 2+ -sensitive kinases and phosphatases, triggers changes in the number of synaptic AMPARs [7]. The link between NMDAR activation and presynaptic plasticity, however, is not as well studied. In the case of presynaptic LTP induction, it is traditionally thought that Ca 2+ influx through postsynaptic NMDARs triggers the synthesis and release of a retrograde signal, most likely nitric oxide (NO), which in turn triggers increases in P r [8,9]. Several studies, however, have suggested that presynaptic enhancement can actually be induced in the presence of NMDAR antagonists. This form of plasticity relies on L-type voltage-gated Ca 2+ channels (L-VGCCs) [10][11][12][13][14], but still depends on NO signalling [15]. These findings suggest that presynaptic plasticity requires neither the activation of NMDARs nor NMDAR-dependent NO synthesis.
Glutamate release is inevitably necessary to drive the postsynaptic depolarization required for LTP. This function of glutamate is not site-specific, since depolarization triggered by one synapse will spread to another. The necessity for site-specific release of glutamate in the induction of LTP is instead imposed by postsynaptic NMDARs, since NMDAR-mediated Ca 2+ signalling is needed for the postsynaptic expression of LTP [2,7] .
Consequently, manipulations that enhance postsynaptic NMDAR signalling at the level of the single synapse reliably augment the induction of postsynaptic LTP [7,[16][17][18]. However, that NMDARs are unlikely to be necessary for the induction of presynaptic LTP, suggests that the role of synapse-specific glutamate release in presynaptic plasticity may be different. Indeed, a common finding across a number of studies is that high P r synapses are more likely to show presynaptic depression whereas low P r synapses are more likely to show presynaptic potentiation [19][20][21][22][23]. Moreover, glutamate release is known to activate presynaptic NMDARs [24], which can induce presynaptic LTD [25]. Thus enhanced glutamate release at a presynaptic terminal, unlike as the dendritic spine, may not necessarily result in enhanced potentiation, but instead promote depression. Several studies have also demonstrated that presynaptic terminals initially releasing little or no glutamate are reliably potentiated following tetanic stimulation [19-24, 26, 27]. How low P r synapses, including those that are putatively silent, can undergo such activitydependent potentiation raises questions as to the exact role of glutamate in presynaptic plasticity.
Here we re-examined the mechanisms underlying activity-dependent presynaptic changes at CA3-CA1 hippocampal synapses, with a particular focus on understanding the role of glutamate in presynaptic plasticity. We bidirectionally manipulated glutamatergic signalling during synaptic activity and examined the resulting consequences on P r at single synapses using Ca 2+ imaging. Remarkably we found that site-specific glutamatergic signalling was unnecessary for the induction of presynaptic LTP. Postsynaptic depolarization alone could increase P r at active presynaptic terminals by promoting the release of NO from neuronal dendrites in a manner dependent on L-VGCC activation. This increase was both Hebbian, in that it required presynaptic activity to precede postsynaptic depolarization, and site-specific, in that it did not spread to inactive terminals. Glutamate release, in contrast, promoted decreases in P r by activating presynaptic NMDARs. Our findings support a simple mathematical model in which net changes in P r at a presynaptic terminal depends on the amount of glutamate release and postsynaptic depolarization that accompanies presynaptic activity, suggesting that LTP-promoting processes and LTD-promoting processes do not operate separately, but rather jointly to tune P r at individual synapses.

High frequency presynaptic activity inhibits presynaptic LTP and augments presynaptic LTD
We were interested in understanding the mechanisms underlying activity-dependent presynaptic changes at CA3-CA1 synapses, with a particular focus on understanding the role glutamate plays in presynaptic plasticity. We started by examining how manipulating glutamatergic signalling at synapses would affect activity-driven changes in presynaptic function. We recorded excitatory postsynaptic potentials (EPSPs) in CA1 neurons in hippocampal slice cultures. Cells were recorded using patch electrodes (4-8MΩ) and EPSPs were evoked by Schaffer-collateral stimulation. Baseline EPSP recordings were kept short (< 5 minutes) to minimize postsynaptic dialysis which impedes LTP induction. Following baseline recording, we induced LTP by pairing presynaptic stimulation with postsynaptic depolarization. Pairings were causal, in that presynaptic stimuli preceded postsynaptic spiking by 7-10ms. A single pairing was repeated 60 times at 5Hz. For postsynaptic depolarization, we injected current of sufficient amplitude to generate 3-6 postsynaptic spikes over a 50ms time course, with the first spike starting 7-10ms following the start of current injection (see Figure 3A for example). Spikes often tended to broaden over the time course of injection. The resulting waveform resembled a complex spike, which is known to efficiently drive LTP in vitro [28][29][30][31] and has been recorded in the hippocampus in vivo [32,33]. We found that this pairing protocol produced robust and reliable LTP (fold ΔEPSP slope : 1.88±0.24; n=6; p<0.01; Figure 1A,B), which had a presynaptic component of expression, as assessed by a decrease in the paired pulse ratio (PPR) (ΔPPR: -0.39±0. 15; n=6; p<0.01; Figure 1C). We next examined the effects of elevating glutamatergic signalling during LTP induction on presynaptic plasticity. Under physiological conditions, elevated glutamate signalling arises from increased presynaptic activity. We therefore repeated our LTP experiments, but during induction, in the place of single presynaptic pulses, we delivered short, high frequency bursts of presynaptic stimuli stimuli to drive more glutamate release. The burst consisted of two pulses, delivered 5ms apart, and resembled high-frequency bursting activity recorded in CA3 neurons in vivo [34]. We found that this protocol produced significantly less LTP compared to single pulse pairings (fold ΔEPSP slope : 1.36±0.13; n=6; vs. single pulse pairings: p<0.05), and was accompanied by no significant changes in PPR (ΔPPR: 0.00±0.04; n=6; p>0.99; Figure 1C), suggesting that LTP was likely to be expressed exclusively postsynaptically. These findings suggest that high frequency presynaptic activity can inhibit the induction of presynaptic LTP.

Glutamate photolysis inhibits presynaptic LTP and augments presynaptic LTD
We next tested whether the effects of high frequency presynaptic stimulation on presynaptic plasticity were in fact due to the presynaptic terminal releasing more glutamate, as opposed to other effects, such as an increased Ca 2+ influx in the presynaptic terminal. To do so, we opted to use glutamate uncaging, which would enable us to elevate glutamate release at synapses during LTP induction whilst keeping the frequency of presynaptic stimulation constant. To keep the experiment as physiological as possible, we restricted glutamate uncaging to single synapses. We then monitored activity-dependent changes in P r at these synapses by imaging postsynaptic Ca 2+ transients, as previously described [27,35]. This technique relies on the fact that at most CA3-CA1 synapses, uniquantal glutamate release, through AMPAR-mediated depolarization, generates sufficient Ca 2+ influx from NMDAR and voltage-gated Ca 2+ channels (VGCCs) to be detected by Ca 2+ -sensitive dyes [1,36,37]. Consequently, the proportion of trials in which single presynaptic stimuli generate postsynaptic Ca 2+ transients can be used to calculate P r at single synapses [37]. CA1 pyramidal neurons were loaded with the Ca 2+ sensitive dye Oregon Green BAPTA-1, and a fluorescently-labelled glass electrode was used stimulate Schaffer-collaterals in the vicinity of the imaged dendrite ( Figure 2A). Dendritic spine fluorescence was examined in order to identify synapses that were responsive to stimulation. To increase the likelihood of visually identifying responsive synapses, especially those with low basal release probabilities, we always delivered two presynaptic stimuli, 70ms apart, to transiently increase P r ( Figure 2B). When a synapse was found that responded to stimulation, it always responded in an all-or-none manner, with Ca 2+ transients largely restricted to the spine head. As expected, Ca 2+ transients were also more likely to be elicited by the second of the two presynaptic stimuli because of the effects of short-term plasticity. P r was calculated as the proportion of trials in which the first of the two presynaptic stimuli generated a fluorescent increase in the spine head; the second of the two presynaptic stimuli was ignored.
Because of the additional time requirements of these experiments, cells were recorded from using sharp microelectrodes (80-120MΩ) to minimize dialysis that otherwise compromises LTP induction. Following baseline measurements of P r , we induced LTP as before, by delivering 60 individual presynaptic stimuli at 5Hz, each paired with postsynaptic depolarization Consistent with our electrophysiological results, this protocol evoked an increase in P r (ΔP r : 0.19±0.03; n=14; Figure 2B,D,F) and a long-lasting potentiation of the recorded EPSP (fold ΔEPSP slope :1.97±0.13; n=12; Figure 2G,I). We then repeated the experiment but this time, during LTP induction, each presynaptic stimulus was paired with photolysis of caged glutamate at the synapse. We adjusted the laser power to ensure that photolysis mimicked the fluorescent changes elicited by uniquantal glutamate release evoked by electrical stimulation (ΔF/F photolysis vs. stimulation: 0.38±0.08 vs. 0.43±0.09; n=12; p=0.72; Figure 2A) Figure 2G,I), suggesting that the failure for the imaged synapse to support LTP could not be attributed to the failure of the recorded cell or slice to support LTP. In five experiments, LTP induction was repeated for a second time at the same synapse, in the absence of caged glutamate, but with photolytic laser exposure; under these conditions, the expected increase in P r was observed (ΔP r : 0.18±0.02; n=5; post-photolysis control in Figure 2D). Increases in P r were also observed in control experiments, in which LTP induction was conducted in the presence of caged glutamate, but in the absence of photolytic laser exposure (ΔP r : 0.21±0.08; n=3). These results suggest that the inhibitory effect of photolysis on P r was due to glutamate release, as opposed to non-specific effects of uncaging.
We also examined the effects of glutamate photolysis delivered in the absence of postsynaptic depolarization (unpaired stimulation). Delivery of 60 presynaptic stimuli at 5 Hz, as before, had no effect on the recorded EPSP (fold ΔEPSP slope : 1.03±0.10; n=9; p>0.99; Figure 2H,I), consistent with our previous result, and produced no changes in P r at the majority of synapses imaged ( Figure 2C,E,F). We did, however, notice that synapses with initially high release probabilities (Pr>0.5), showed a modest decrease in P r following unpaired stimulation ( Figure 2E); this decrease was not likely to be detected by electrophysiological recordings because high P r synapses comprise an estimated <10% of synapses in our preparation [35]. Remarkably, when we coupled each presynaptic stimulus with glutamate photolysis, we now observed decreases in P r at all imaged synapses, regardless of their initial P r (ΔP r photolysis vs. control: -0.33±0.08 vs. -0.12±0.06; n=9,10 p=0.037; Figure 2C,E,F). Consistent with our hypothesis, these findings suggest that elevated glutamate release at a synapse inhibited presynaptic LTP and, in the absence of postsynaptic depolarization, drove presynaptic LTD.

Presynaptic LTP can be induced in complete glutamate receptor blockade
Given that augmenting glutamatergic signalling inhibited the induction of presynaptic LTP, we asked if glutamate release was required at all for driving increases in P r during paired stimulation. We reasoned that although the activation of glutamate receptors would ultimately be necessary to drive the postsynaptic depolarization required for LTP induction, presynaptic potentiation may not actually require any one presynaptic terminal to trigger glutamate release, provided that its activity coincides with postsynaptic depolarization, which could be triggered by glutamate release at other co-active synapses.
To test this possibility we attempted to induce LTP at CA3-CA1 synapses in hippocampal slices with all known glutamate receptors (AMPARs, KainateRs, NMDARs, and mGluRs) pharmacologically inhibited (10µM NBQX, 50-100µM AP5, 0.5-1mM MCPG, 100µM LY341495). Given the additional time requirements for these experiments, we recorded from CA1 neurons using high-resistance patch electrodes (18-25MΩ) to limit postsynaptic dialysis. Following abolishment of the EPSP, synaptic activity was causally paired as before with complex spikes ( Figure 3A). The antagonist cocktail was washed out, and the EPSP was allowed to recover. As expected, drug washout was never complete ( Figure 3C,D) and so it was necessary to compare the EPSP recorded from the pathway receiving paired stimulation to a second, independent control pathway recorded simultaneously ( Figure   3A,B). We found that pairing induced a robust enhancement of the EPSP in the stimulated pathway (fold ΔEPSP slope paired vs. control: 1.12±0.13vs. 0.71±0.12; n = 7; p < 0.05; Figure   3B,D), which lasted for the duration of the recording (up to 40-90 minutes post-pairing).
Pairing resulted in a 1.72±0.21 fold potentiation (p<0.05), which we estimated by normalizing the fold change in the EPSP of the paired pathway to that of the control pathway. Notably, EPSP recovery of the control pathway was not significantly different from experiments in which drugs were applied in the absence of paired stimulation (control vs. drugs-only: 0.71±0.12 vs. 0.54±0.11; n = 7 and 5; p = 0.33; Figure 3C,D), suggesting that LTP was restricted to only synapses that were active during the pairing.
We next examined the locus of LTP expression. We found that LTP induction in glutamate receptor blockade was accompanied by a significant increase in the co-efficient of variation parameter (CV -2 ) (paired vs. control: 2.80 ±0.79 vs. 0.89 ±0.10; n = 7; p<0.01; Figure 3E), and a significant decrease in PPR (paired vs. control ΔPPR: -0.28±0.06 vs. 0.03±0.03; n = 6; p<0.01; Figure 3F). Both of these changes are consistent with a presynaptic component of LTP expression, and both were found only in the paired pathway, suggesting that LTP induction was site-specific.
We also confirmed that presynaptic enhancements could be induced under glutamate receptor blockade in acute hippocampal slices (ΔEPSP slope paired vs. control: 0.65 ±0.18 vs. 0.43±0.14; n = 6; p<0.05; Supplemental Figure 1 and 2). Pairing resulted in a 1.56±0.11 fold potentiation (p<0.01), which we again estimated by normalizing the fold change in the EPSP of the paired pathway to that of the control pathway. As with slice cultures, these enhancements were accompanied by significant decreases in the PPR (ΔPPR  Figure 1 and 2) that were only evident in the paired pathway, suggesting that the changes were both presynaptic in origin and site-specific.

LTP induction in glutamate receptor blockade is associated with an increase in P r
We then returned to Ca 2+ imaging to determine whether we could directly observe increases in P r at single synapses associated with the induction of LTP in glutamate receptor blockade (Figure 4). To minimise dialysis for these experiments during drug wash-in, imaging was conducted in the absence of electrophysiological recordings on CA1 neurons that were pre-loaded with Ca 2+ indicator dye (see Methods). Following an initial assessment of P r , glutamate receptor antagonists (APV, NBQX, MCPG, and LY341495) p<0.01) ( Figure 4B,C). The induction of presynaptic LTP in the absence of glutamatergic signalling was therefore Hebbian, requiring presynaptic activity to be causally paired with strong postsynaptic depolarization.

Postsynaptic depolarization increases P r by promoting dendritic release of nitric oxide in a manner dependent on L-type voltage-gated Ca 2+ channels
We next investigated the mechanism by which paired stimulation could trigger increases in P r in the absence of glutamatergic signalling. The requirement for postsynaptic depolarization in the induction of presynaptic potentiation suggests a need for a diffusible retrograde messenger. Perhaps the most promising, albeit still controversial, retrograde signal implicated in LTP induction is nitric oxide (NO) [14]. Although NO synthesis has classically been associated with the activation of postsynaptic NMDARs [9], there is some suggestion that Ca 2+ influx from L-type voltage-gated Ca 2+ channels (L-VGCCs), which have previously been implicated in presynaptic LTP [12,13,15], may also trigger NO production [15,38,39]. If NO synthesis in neuronal dendrites can be triggered by L-VGCC activation, then NO production could occur in a manner dependent on postsynaptic depolarization, but independent of glutamatergic signalling. To test this hypothesis, we first asked whether presynaptic LTP, induced in glutamate receptor blockade, was dependent on L-VGCC activation and NO signalling. Consistent with our hypothesis, we found that pairing-  Figure 1 and 2), suggesting that, as in cultured slices, presynaptic efficacy in acute slices was similarly regulated by NO signalling.
We then examined whether NO production depended on L-VGCC activation. We transiently patched CA1 neurons in order to load them with the NO-sensitive dye, DAF-FM, and then measured fluorescent changes in the apical dendrites prior to and following postsynaptic depolarization in glutamate receptor blockade. Given the poor signal-tonoise ratio associated with DAF-FM imaging, and to circumvent the problem of intracellular dialysis, we non-invasively evoked strong postsynaptic depolarization by elevating extracellular K + , as previously described [38,39]. Under these conditions, we observed robust fluorescent increases in neuronal dendrites (ΔF/F: 0.38±0.04; n=5) ( Figure   5B,C). These increases were dependent on NO synthesis as they could be prevented by To then examine whether NO release alone was sufficient for the induction of presynaptic LTP, we examined whether increases in P r could be elicited when presynaptic stimulation was paired with NO release, in the absence of postsynaptic depolarization. We paired 30- causal pairing: p<0.05; Figure 5E,F), suggesting that LTP did not result from non-specific effects associated with photolysis. Moreover, when presynaptic stimuli followed NO photolysis no significant change in P r was observed (ΔP r : -0.01±0.04; n=8; vs. causal pairing: p<0.01; Figure 5D-F), suggesting that NO-mediated potentiation was Hebbian, requiring presynaptic activity to precede, rather than follow, NO release.

Glutamate release decreases P r via activation of presynaptic NMDARs
If glutamatergic signalling at a presynaptic terminal is not required for its potentiation, then what is the role of glutamate release in presynaptic plasticity? To investigate this we compared changes in P r produced by activity in the presence and absence of glutamate receptor blockade. Remarkably, we found glutamate receptor blockade augmented increases in P r produced by paired stimulation (ΔP r blockade vs. control: 0.34±0.04 vs. 0.18±0.02; n = 10; p < 0.05; Figure 6A-C) and abolished decreases in P r generated by unpaired stimulation (ΔP r blockade vs. control: 0.00±0.03 vs. -0.21±0.05; n=10, 9; p<0.01; Figure 6D,E). These results suggest that glutamate release during synaptic activity promotes decreases in Pr, regardless of the accompanying levels of postsynaptic depolarization.
How might glutamate release depress P r ? We have previously reported that presynaptic NMDARs are found at CA3-CA1 synapses; these receptors act as reliable detectors for uniquantal glutamate release [24] and, at neocortical synapses, have been implicated in presynaptic LTD [40][41][42][43]. We therefore examined whether these receptors mediated the inhibitory effects of glutamate observed on presynaptic function.
Given the difficulties associated with selectively blocking pre-, as opposed to post-, synaptic NMDARs, several groups have investigated the role of presynaptic NMDARs in plasticity by comparing the effects of bath application of AP5 or MK801, which blocks both pre-and post-synaptic NMDARs, with that of intracellular MK801 application, which selectively blocks postsynaptic NMDARs [43][44][45][46]. We sought to use a similar approach.
However, because MK801 does not readily washout, and since postsynaptic NMDARs greatly contribute to spine Ca 2+ influx [36,37,47], we first examined whether the permanent loss of postsynaptic NMDAR signalling affected our ability to measure P r using postsynaptic Ca 2+ imaging. We found that at about 50% of synapses, NMDAR blockade reduced, but did not entirely abolish synaptically-evoked Ca 2+ transients (Supplemental Figure 3). The residual Ca 2+ transients were mediated by activation of voltage-gated Ca 2+ channels (VGCCs) in response to AMPAR-mediated depolarization, and could be used to accurately measure P r (Supplemental Figure 3). Importantly, the average P r of these synapses did not significantly differ from that of synapses lacking a residual Ca 2+ transient in NMDAR blockade (ΔP r : AP5-sensitive vs. AP5-insensitive: 0.42=0.07 vs. 0.47=0.11; p=0.67; Supplemental Figure 3). These findings suggest that, in NMDAR receptor blockade, VGCC-dependent spine-Ca 2+ influx can be used as a means of calculating P r at a sizeable and representative proportion of presynaptic terminals.
Incidentally, increases in P r were not only rescued by bath application of MK-801 but were significantly greater than in control conditions (ΔP r control: 0.20±0.03; n=12; vs. bath MK-801: p<0.05; Figure 7A,B). This is likely because bath application of MK-801, in addition to blocking the inhibitory effects of caged glutamate, additionally blocked the inhibitory effect of endogenous glutamate release that was present in control conditions ( Figure   6B,C). As expected, bath, but not intracellular, application of MK-801 also prevented photolysis-induced augmentation of LTD, in which stimuli were delivered in the absence of postsynaptic depolarization (ΔP r : bath MK-801 vs. photolysis: -0.02±0.03; vs. -0.29±0.06; n=7 vs 9; p<0.01; intracellular MK-801: -0.28±0.08; n=7; vs. photolysis: p=0.92; Figure   7C,D). Bath MK-801 application additionally prevented LTD induction present at high P r synapses under control condition ( Figure 7C,D) again, likely by blocking the inhibitory effects of endogenous glutamate release ( Figure 6D,E).

A simple mathematical framework predicts activity-dependent changes in P r
Our findings thus far suggest that changes in P r are driven by two processes: 1) postsynaptic depolarization, which promotes increases in P r through L-VGCC-dependent release of NO from neuronal dendrites, and 2) glutamate release, which promotes decreases in P r through presynaptic NMDAR activation ( Figure 8A). We have demonstrated that net changes in P r will depend on both processes. This suggests that changes in P r at a presynaptic terminal are driven by a mismatch between 1) the amount of postsynaptic depolarization and 2) the amount of glutamate release that accompanies presynaptic activity. We therefore asked whether these two variables could be incorporated into a mathematical framework that could predict change in P r (ΔP r ) in our data set. For ease of calculation, we chose to quantify both these variables in terms of probabilities, that is, 1) the probability that presynaptic activity was accompanied by strong postsynaptic depolarization (P depol ) and 2) the probability that glutamate was released at the synapse (P glu ). The simplest mathematical framework to model changes in P r would be: ΔP r = η (P depol -P glu ), where net changes in P r would be proportional to the relative difference or mismatch between P depol and P glu as defined by (P depol -P glu ), multiplied by some constant η, defined as the learning rate ( Figure 8B).
We applied the model ΔP r = η(P depol -P glu ) to our data set of imaged synapses across all experimental conditions. As a reminder, to induce plasticity we had delivered 60 presynaptic stimuli at 5Hz, where each stimulus was either consistently paired or unpaired with postsynaptic depolarization. For conditions with paired stimulation, we set P depol to 1, since depolarization was present for every presynaptic stimulus. However, for conditions in which NO signalling was inhibited, P depol was set to 0, as the effects of depolarization would be absent for every presynaptic stimulus. For all other conditions, in which presynaptic stimuli were not paired with presynaptic stimulation, we set P depol to 0. We then determined values for P glu . For experiments involving glutamate photolysis, P glu was set to 1, as glutamate was available for every presynaptic stimulus. For experiments in which presynaptic NMDARs were inhibited, P glu was set to 0, as the inhibitory effects of glutamate were absent for every presynaptic stimulus. For all other conditions, P glu had to be calculated for each synapse. Although P glu will likely be proportional to the initial P r of a synapse, it will not be equivalent to initial P r owing to short-term plasticity effects induced during the 5Hz stimulation train [24]. To obtain a more accurate estimate of P glu , we conducted an additional set of experiments, in which we used Ca 2+ imaging to examine glutamate release at single synapses stimulated with 60 presynaptic pulses delivered at 5Hz (Supplemental Figure 4). We calculated P glu as the total number of glutamate release events in the train divided by the number of pulses in the train (i.e. 60). When plotted against initial P r , we obtained a linear relationship with which to relate P glu to initial P r (P glu = 0.475*initial P r + 0.2175; Supplemental Figure 4) . We used this relationship to derive estimates of P glu for synapses in the remainder of our data set. Finally, we determined the learning rate η, which obtained the best fit of the model (η=0.35).
To assess the goodness of fit of our model, we divided the remainder of our data (n=216 synapses) into 6 categories, depending on whether the effects of postsynaptic depolarization were present or absent, whether glutamate photolysis was present or absent, and whether presynaptic NMDAR blockade was present or absent. These 6 categories, along with the models predictions for ΔP r are shown in Figure 8C and are summarized in the table below.
No. Effects present P depol P glu ΔP r = η (P depol -P glu ) Unpaired stimulation + preNMDAR blockade 0 0 0 * Actual value will vary across synapses in this condition and is therefore represented as the variable P glu .
For each category, we plotted initial P r against final P r for all synapses within the category, and superimposed the model's predictions as a red trend line ( Figure 8D). The model's predictions were compared to a simple linear model, derived as a line of best fit (grey line; Figure 8D) for the data in the each category. The model consistently achieved a predictive power that was comparable to that of the line of best fit ( Figure 8D,E), with the exception of category 4. Synapses in this condition received glutamate photolysis during unpaired stimulation and therefore underwent augmented depression. The model tended to overestimate the amount of depression for synapses with a low initial P r and underestimate the amount of depression for synapses with a high initial P r . However, the model's deviations from the line of best fit did not achieve statistical significance (p=0.11; Figure 8E), likely owing to a small number of data points (n=13) in this category. Across all 6 categories the model's explanatory power was not different from the lines of best fit (p=0.44; Figure 8E), despite only having a single free parameter (η, the learning rate) as compared with a total of 12 free parameters used by the lines of best fit (2 free parameters per line, the slope and the intercept). Consequently, the model achieved a substantially better (i.e. lower) Bayesian Information Criterion (BIC) than the lines of best fit (BIC lines of best fit -BIC model = 34.97). This was also the case when we combined categories having similar trends (category 1 and 6) to minimize the free parameters used by the lines of best fit from 12 to 10 (BIC lines of best fit -BIC model = 24.25). Thus, a simple and parsimonious mathematical learning rule is capable of effectively predicting changes in P r across a range of experimental conditions.

Discussion
We have explored the mechanisms of presynaptic plasticity at CA3-CA1 synapses in hippocampal slices. Based on our findings we present a unified framework of presynaptic plasticity, which suggests that, at active presynaptic terminals, changes in P r are driven by two processes: 1) postsynaptic depolarization, which promotes increases in P r through the L-VGCC-dependent release of NO from neuronal dendrites (LTP-promoting process), and 2) glutamate release, which promotes decreases in P r through presynaptic NMDAR activation (LTD-promoting process). Both processes operate together to tune presynaptic function, with net changes in P r depending on the strength of each process during presynaptic activity. (Figure 8A,B). Consequently, we found that pairing presynaptic activity with strong postsynaptic depolarization in glutamate receptor blockade produced the greatest increase in P r , whereas pairing presynaptic activity with glutamate photolysis in the absence of postsynaptic depolarization produced the greatest decrease in P r .

Presynaptic LTP can occur in the absence of synapse-specific glutamate signalling
Importantly, the mechanism of plasticity proposed in this study enables presynaptic terminals releasing little or no glutamate to become potentiated provided that their activity is accompanied by strong postsynaptic depolarization. This feature of presynaptic LTP may have important implications for plasticity. Most central synapses have low glutamate release probabilities, with some synapses appearing to release no glutamate in response to presynaptic stimulation [48,49] This is true for synapses recorded in both in vitro preparations from young rodents and ex vivo preparations from adult rodents. In fact, a recent electron microscopy study has identified that a significant portion of synapses (up to 35-50%) in the adult rodent hippocampus have presynaptic zones lacking synaptic vesicles in their near proximity (<170nm); these so-called "nascent zones" have been hypothesized to be functionally silent [50]. Although the existence of bona fide presynaptically silent synapses remains controversial [48], the low release probabilities of central synapses suggests that it is possible that activity at a presynaptic terminal may not elicit glutamate release at the synapse, but may still coincide with strong postsynaptic depolarization, driven by glutamate release at other co-active synapses. Under such conditions the presynaptic mechanisms of plasticity elucidated in this study, could enable the efficient induction of LTP at the presynaptic terminal, thereby allowing Hebb's postulate to be fulfilled, even under conditions of little or no glutamate release.
Our finding that presynaptic enhancements can occur without glutamatergic signalling at the synapse raises the question as to why many studies show that LTP induction can be abolished or impaired by blockade of one or more glutamate receptor subtypes [47,51,52]. Reconciliation of this discrepancy is readily achieved by recognizing that antagonists to AMPA, NMDA, and mGlu receptors can all reduce the level of postsynaptic spiking generated by synaptic stimulation, since all of these receptors have been shown to contribute to postsynaptic depolarization [33,[53][54][55]. Given that we find presynaptic changes rely on the voltage-dependent release of NO, it is possible that blockade of any of these glutamate receptor classes would abolish or reduce presynaptic LTP in an indirect way, by reducing postsynaptic depolarization and the activation of L-VGCCs. This may explain, in part, why experimental manipulations that augment the levels of postsynaptic depolarization reliably rescue LTP in AMPAR [47,56], NMDAR [8,[11][12][13][57][58][59][60][61] and mGluR blockade [62]. Importantly, our LTP induction protocol circumvented the need for any glutamate receptor-dependent depolarization during paired stimulation, as strong postsynaptic spiking was elicited by somatic current injection. Based on these results, we would argue that the physiological role of glutamate release in presynaptic potentiation is for driving postsynaptic spiking as opposed to conveying a synapse-specific signal; this contrasts with the role of glutamate release in postsynaptic plasticity, in which synapsespecific activation of postsynaptic NMDARs is necessary for LTP induction.
While our approach for inducing LTP resembles that of traditional spike-timing dependent plasticity (STDP) protocols, which rely on NMDAR activation [63], there is a key difference.
In our study, postsynaptic depolarization took the form of complex spikes, which included a brief period (7-10 ms) depolarization before the first spike. This period of subthreshold depolarization is known to facilitate the induction of LTP, possibly by inactivating voltagegated potassium channels within the dendrite, which otherwise impede action potential backpropagation [64][65][66][67][68]. Moreover, these emulated potentials, like complex spikes recorded in vivo [32], contained broadened action potentials, which likely reflect strong depolarization in the dendrites [66,68]. Consequently, the postsynaptic waveforms used in our study were likely to generate stronger levels of postsynaptic depolarization, and in a manner independent of glutamate release and NMDAR activation, than those used in STDP studies.

Presynaptic LTP requires L-VGCC and nitric oxide signalling
In our study, we demonstrate the importance of L-VGCC signalling in presynaptic potentiation, consistent with findings from other laboratories [10][11][12][13]. Since L-VGCCs have high voltage activation thresholds, presynaptic potentiation will require strong levels of postsynaptic depolarization, which may explain why presynaptic enhancements are not always reported in the literature [8]. It is very likely that the successful induction of presynaptic LTP depend on the levels of postsynaptic depolarization achieved during tetanus [10,12], which in turn will be influenced by a variety of experimental factors, including the frequency and intensity of tetanic stimulation. L-VGCCs also exhibit voltagedependent inactivation during extended periods of depolarization (>1s) [69,70]. As such, the use of pairing protocols, in which the postsynaptic cell is voltage clamped at 0mV for prolonged periods of time (>10s) during presynaptic stimulation, are likely to be unsuitable for driving presynaptic enhancements, despite being commonly used to drive postsynaptic enhancements [14].
It has long been recognized that the induction of LTP at the presynaptic locus requires a retrograde signal [71]. The most promising candidate is NO [9]. The role of NO in LTP has been a source of much controversy, and some studies have concluded that NO signalling is not necessary in LTP induction [9]. However, given that NO is likely to be important for presynaptic strengthening, the effect of NO signalling on synaptic plasticity will depend on whether presynaptic enhancements are obtained following LTP induction [14]. Indeed, studies that actually confirm presynaptic changes following LTP induction, including our own, consistently demonstrate that presynaptic enhancements depend on the synthesis and release of NO in both acute and cultured hippocampal preparations [72][73][74][75].
It has generally been assumed that NO synthesis is dependent on Ca 2+ influx from postsynaptic NMDARs [9]; however, several studies have demonstrated that induction of presynaptic LTP is possible in NMDAR blockade suggests that this signalling pathway is not necessary for presynaptic potentiation [11][12][13]. Consistent with this notion, in our experiments, we found that specific blockade of postsynaptic NMDARs using intracellular application of MK-801 had no effect on changes in P r induced by paired stimulation.
Instead, we found an alternative pathway for NO synthesis that was crucial for presynaptic strengthening, and that depended on strong postsynaptic depolarization and activation of L-VGCCs. are not only restricted to active synapses, but specifically at synapses whose activity precede, rather than follow, NO release; thus, the requirements of NO signalling are consistent with those of Hebbian and spike-timing dependent plasticity [63]. It is important to mention that some groups have found no effect of exogenous NO application on synaptic plasticity [8,9]. However, like glutamate, the effects of NO will depend on the spatiotemporal dynamics of NO signalling and the pattern of concurrent synaptic activity, which will likely vary across studies; therefore it is not surprising that NO application, like that of glutamate, can potentiate, depress, or have no effect on synaptic input depending on experimental conditions [14].

Glutamate drives presynaptic LTD via presynaptic NMDAR signalling
At active presynaptic terminals, whereas postsynaptic depolarization drives increases in P r , we show, unexpectedly, that glutamate release drives decreases in P r by acting on presynaptic NMDA receptors. We found that presynaptic NMDAR signalling operated both during LTP and LTD induction paradigms to reduce P r . This finding suggests that the potentiating effects of postsynaptic depolarization and the depressing effects of endogenous glutamate release occur simultaneously during synaptic activity, regardless of the nature of postsynaptic depolarization. Thus, the processes underlying LTP and LTD induction are not temporally distinct mechanisms as traditionally believed, but operate jointly to tune synaptic function during presynaptic activity. Our results may explain why sometimes the same pairing protocol that produces LTP at low P r synapses, produces LTD at high P r synapses; presumably the postsynaptic depolarization achieved by such protocols is not of sufficient magnitude to prevent the depressing effects of glutamate release at high P r synapses [22,23]. At neocortical synapses, presynaptic NMDARs have been implicated in the induction of LTD; although their pharmacological inhibition does not appear to effect LTP [41,78]. It is possible that the low frequency (0.2Hz) of presynaptic stimulation used during LTP induction in these studies did not result in sufficient levels of glutamate release to elicit presynaptic depression via presynaptic NMDAR activation. By contrast, in our study LTP induction involved presynaptic stimulation at a theta frequency, which is effective at promoting glutamate release at the synapse [24]. As such, the inhibitory effects of presynaptic NMDARs on LTP may only be evident at higher stimulation frequencies.
Presynaptic NMDARs may, however, operate at lower stimulation frequencies in the context of spike-timing dependent plasticity. Recently, it was shown that low-frequency (0.2Hz), anti-causal pairing of pre-and post-synaptic spiking induced presynaptic LTD at hippocampal synapses, in a manner dependent on presynaptic NMDAR signalling [25].
This form of LTD, in addition to glutamate release, required endocannabinoid and astrocytic signalling, similar to spike-timing dependent LTD in the neocortex [42].
Previously we have demonstrated that presynaptic NMDARs at hippocampal synapses facilitate transmitter release during theta stimulation [24]. Given our current findings, presynaptic NMDARs appear to be important for presynaptic facilitation in the short-term, but presynaptic depression in the long-term. Presynaptic NMDAR in the neocortex also appear to facilitate both evoked and spontaneous glutamate release, and yet are similarly implicated in presynaptic LTD [44]. It may appear peculiar for a single protein to mediate seemingly disparate functions; however, another way to view the presynaptic NMDAR is as a dynamic regulator of presynaptic activity, appropriately tuning glutamate release depending on the patterns of pre-and postsynaptic activity. As such, the receptor may aid glutamate release during theta-related activity, but, triggers presynaptic LTD when this release fails to elicit sufficiently strong postsynaptic depolarization.

Presynaptic plasticity optimally tunes glutamate release
We found a simple mathematical framework [ΔP r = η (P depol -P glu )] could predict changes in P r based on a mismatch between 1) the probability that presynaptic activity is accompanied by strong depolarization (P depol ) and 2) the probability that presynaptic activity is accompanied by glutamate release (P glu ) during synaptic stimulation. According to this model, presynaptic plasticity is driven only when there is a mismatch between P glu and P depol . Plasticity, by changing P r , works to drive the value of P glu closer to that of P depol .
This has a very intuitive interpretation. P glu is a measure of the synapses ability to drive postsynaptic activity. P depol is a measure of the synapse's ability to predict postsynaptic activity. By driving the value of P glu closer to P depol , presynaptic plasticity changes a synapse's ability to drive postsynaptic activity, so that it matches its ability to predict postsynaptic activity.
It is important to recognize that P glu will not only depend on basal P r , but also on the pattern of presynaptic activity, owing to the effects of short-term plasticity [79].
Moreover, for any given pattern of pre-and post-synaptic activity, there exists a value of P r for which P glu will equal P depol . That is, there exists a value of P r for which the probability of glutamate release for a given pattern of presynaptic stimulation (P glu ) will be equal to the probability that presynaptic stimulation is accompanied by strong postsynaptic depolarization (P depol ). The model predicts that this value of P r represents a target value that all synapses will tend to for a given pattern of pre-and post-synaptic activity.
Synapses below this target value will be potentiated because P depol >P glu , whereas those above this target value will be depressed because P depol <P glu . This can explain why in the literature, P r has previously been reported to tend to a certain value when a particular pattern of pre-and post-synaptic stimulation is applied [22]. According to our model, the target P r value given by pairing high frequency presynaptic stimulation with strong postsynaptic depolarization is lower than pairing low frequency presynaptic stimulation with strong postsynaptic depolarization. In both cases, P depol is high, meaning that presynaptic plasticity will continue to drive P r until P glu is also high.
However, with high frequency stimulation, a low target P r achieves a high P glu . In contrast, with lower frequency stimulation, a higher target P r is required to achieve a similarly high P glu . Consequently, as we showed in our experiments with Figure 1, when paired with strong postsynaptic depolarization, low frequency presynaptic stimulation will produce greater presynaptic potentiation than higher frequency stimulation. That P r is kept low at synapses where presynaptic bursts are good predictors of postsynaptic spiking ensures that only presynaptic bursts mobilize glutamate, rather than individual presynaptic stimuli, which may not be good predictors of postsynaptic spiking. Therefore, presynaptic plasticity appears to tune P r , such that the patterns of presynaptic activity that predict postsynaptic spiking are the ones that most efficiently drive glutamate release.

Cultured hippocampal slices
Cultured hippocampal slices (350µm) were prepared from male Wistar rats (P7-P8), as previously described [27]. MgATP, 0.4 Na 3 GTP and 50U/mL creatine phosphokinase) [57]. The recording method used in a given experiment is clearly indicated in the text. In imaging experiments, unless otherwise stated, low-resistance patch electrodes (4-8MΩ) containing standard internal solution were used to transiently load cells with dye. Cells were then subsequently repatched for the purposes of LTP induction.

Electrophysiology and analysis
All electrophysiological data was recorded using WinWCP (Strathclyde Electrophysiology Software) and analyzed using Clampfit (Axon Insturments) and Excel (Microsoft). The initial EPSP slope, calculated during the first 3ms of the response, was used to analyze changes in the EPSP throughout the recording. All data was normalized to the average EPSP slope recorded during baseline to yield ΔEPSP slope. Paired pulse ratio (PPR) was calculated as the average EPSP slope evoked by the second stimulation pulse divided by the average EPSP slope evoked by the first stimulation pulse, as previously described [80]; averages were calculated from 10-20 paired pulse trials. Decreases in PPR are thought to reflect increases in release probability [81]. The coefficient of variation parameter CV -2 , which reflects the mean 2 /variance, was calculated using the EPSP slopes collected over 25-30 trials. The CV -2 calculated from 25-30 stimulation trials taken 30 minutes following LTP induction was normalized to the CV -2 calculated from 25-30 stimulation trials at baseline to yield ΔCV -2 . For each experiment, ΔCV -2 was plotted against mean ΔEPSP slope. When ΔCV -2 > mean ΔEPSP slope, activity-induced enhancements in EPSP slope are thought to predominantly reflect increases in release probability [82]. To control for the effects of incomplete drug washout, for experiments with glutamate receptor blockade, ΔEPSP slope and ΔCV -2 for the tetanized pathway were normalized to values obtained for the control pathway.

Ca 2+ imaging and analysis
In experiments only requiring Ca 2+ imaging without electrophysiological recordings, CA1 pyramidal neurons were loaded with Ca 2+ sensitive dye using a low resistance patch electrode (4-8 MΩ) containing 1mM Oregon Green Bapta-1 (Invitrogen) dissolved in standard internal solution. The cell was loaded for 45-60s, after which the patch electrode was slowly withdrawn over the course of 1-2 minutes using a piezoelectric drive.
Withdrawal of the electrode was very rarely associated with a rise in intracellular Ca 2+ , suggesting that the procedure caused minimal damage to the cell.
A stimulating glass electrode (4-8MΩ) was then brought near (5-20µm) to a branch of imaged dendrite within stratum radiatum. For visualization purposes, electrode tips were coated with bovine serum albumin Alexa 488 conjugate (Invitrogen), as previously described [83]. Briefly, a 0.05% BSA-Alexa 488 solution was made with 0.1M phosphate- Synapses with initial P r values of 0-0.7 were used for LTP experiments. Since our LTD protocol did not elicit depression in low P r synapses (Figure 2)

Mathematical model
The model ΔP r = η(P depol -P glu ) was used to predict changes in P r in our data set of imaged synapses, across all experimental conditions. Data was divided into 6 categories, depending on whether the effects of postsynaptic depolarization were present or absent, whether the effects of glutamate photolysis was present or absent, and whether the effects of presynaptic NMDAR signalling was present or absent. These 6 categories, along with the models predictions for ΔP r are shown in Figure 8C and are summarized in the Unpaired stimulation + preNMDAR blockade 0 0 0 * Actual value will vary across synapses in this condition and is therefore represented as a variable.
The experimental data in each category are as follows.
Category 1 -included data from experiments in which the effects of postsynaptic depolarization and glutamate photolysis were present during presynaptic stimulation ( Figure 2D, 7A). This included experiments in which MK-801 was applied intracellularly during photolysis ( Figure 7A). For this category, P depol was set to 1 and P glu was set to 1.
Category 2 -included data from experiments in which the effects of postsynaptic depolarization and endogenous glutamate release were present during presynaptic stimulation ( Figure 2D, 4B, 6B). This included experiments in which MK-801 was applied intracellularly ( Figure 6B). For this category, P depol was set to 1 and P glu was calculated for each synapse based on: P glu = 0.475*initial P r + 0.2175, which was experimentally derived from data in Supplemental Figure 4.  Figure 6B). These also included pairing experiments involving glutamate photolysis, but in the presence of MK-801 in the bath ( Figure 7A). For this category, P depol was set to 1 and P glu was set to 0.
Category 4 -included data from experiments in which the effects of postsynaptic depolarization were absent but the effects of glutamate photolysis was present during presynaptic stimulation ( Figure 2E, 7C). This included experiments in which MK-801 was applied intracellularly during photolysis ( Figure 7C). For this category, P depol was set to 0 and P glu was set to 1.
Category 5 -included data from experiments in which the effects of postsynaptic depolarization were absent but the effects of endogenous glutamate release were present during presynaptic stimulation ( Figure 2E, 6D). This included experiments in which MK-801 was applied intracellularly during stimulation ( Figure 6D). For this category, P depol was set to 0 and P glu was calculated for each synapse based on: P glu = 0.475*initial P r + 0.2175, which was experimentally derived from data in Supplemental Figure 4. or MK-801 present in the bath, and in which postsynaptic depolarization was absent ( Figure 6D). 5) Experiments done with bath application of MK-801, and in which postsynaptic depolarization was absent but glutamate photolysis was present. ( Figure   7C). For this category, P depol was set to 0 and P glu was set to 0.
The learning rate η was determined to be 0.35, this gave the best fit of the model on a small training set of 10 synapses, 5 of which that underwent paired stimulation and 5 of which that underwent unpaired stimulation.
For each category, the initial P r was plotted against final P r for all synapses within the category. The model's predictions were compared to that of a line of best fit, derived from simple linear regression. The model's predictions for final P r were set to 1 if they were >1, and were set to 0 if they were <0. Differences in the variance explained by both the model and line of best fit was assessed for significance using an F-test. Bayesian Information Criterion (BIC) was calculated for the model and the lines of best fit as BIC = n*ln(RSS/n) + k*ln(n). Here, n is the number of data points (n=216), ln is the natural logarithm, RSS is the residual sum of squares and calculated by Σ(model predictions -actual value) 2 , and k is the number of free parameters, which was 1 for the proposed model, and 12 for the lines of best fit (2 free parameters per line of best fit, the slope and the intercept). The BIC aids model selection by favouring models that have a high predictive power (explained variance) but a small number of free parameters. A difference of BIC > 10 between two models strongly favors the model with the smaller BIC value .

Statistical analysis
The statistical significance of comparisons was mainly assessed using two tailed, Mann-Whitney or Wilcoxon matched pairs tests, depending on whether the data was unpaired or paired, respectively. One sample t-tests were used to determine if average data significantly differed from an expected value. Pearson correlation coefficients were calculated to determine the significance of linear trends. Significance between trend lines was determined by using a Wilcoxon matched-pairs tests to examine the statistical difference between the two sets of residuals that resulted from fitting a given data set      Asterisks denotes significance differences between the first group in the graph (**p<0.01; Mann-Whitney test).    Importantly, NO drives an increase in P r , but only at presynaptic terminals whose activity precedes its release. The detection of such an event requires an effector (Δt) that is which ΔP r = η (P depol -P glu ). P depol is the probability that presynaptic activity is casually accompanied by strong postsynaptic depolarization during plasticity induction. P glu is the probability that glutamate was released at the synapse during plasticity induction. η is a constant defined as the learning rate, which was determined to be 0.35 to achieve the best fit. (C) Data (n=216 synapses) were divided into 6 categories based on experimental manipulations during plasticity induction in which 60 presynaptic stimuli were delivered at 5Hz. These manipulations include all experimental conditions in which the effects of postsynaptic depolarization were present or absent, whether glutamate photolysis (yellow dot) was present or absent, and whether presynaptic NMDAR blockade (red cross) was present or absent. These categories, along with the associated value of P depol, P glu, and ΔP r , are summarized in table form. Where P glu did not take the value of 0 or 1, it was calculated using the following formula: P glu = 0.475*basal P r + 0.2175; see Supplemental Figure