Introduction

Short term synaptic plasticity (STP) governs changes in the amplitude of synaptic signals following recent neuronal activity. These changes are often rapid and large, and they potently shape information transfer across synapses. Two main components of STP are synaptic facilitation and synaptic depression (Zucker and Regehr, 2002).

Synaptic depression occurs after a burst of high frequency presynaptic firing and associated synaptic release (Betz, 1970; Castillo and Katz, 1954; Charlton et al., 1982). This link with previous presynaptic activity led to the view that synaptic depression is largely due to a decrease in the size of the readily releasable pool (RRP) of synaptic vesicles (SVs), following the loss of SVs by exocytosis (Kavalali, 2006; Regehr, 2012; Zucker and Regehr, 2002). In addition, a number of other mechanisms including desensitization of postsynaptic receptors (Rossi et al., 1995; Trussell et al., 1993), activation of presynaptic inhibitory receptors (Redman and Silinsky, 1994), decreased voltage-dependent calcium entry (Xu and Wu, 2005), decreased SV release probability (Wölfel et al., 2007; Wu and Borst, 1999), heterogeneous release probabilities among release sites (Neher, 2015; Trommershäuser et al., 2003), and non-linear presynaptic calcium buffers (Bolshakov et al., 2019), contribute to synaptic depression in various proportions depending on the preparation.

In more recent years, several studies have uncovered a form of synaptic depression that may differ from the RRP depletion pattern. In invertebrate synapses (Aplysia (Doussau et al., 2010); crayfish (Silverman-Gavrila et al., 2005)), as well in mammalian central synapses (hippocampal CA3 to CA1 (Abrahamsson et al., 2007) and CA3-CA3 (Saviane et al., 2002) synapses, climbing fiber to Purkinje cell synapses (Rudolph et al., 2011); parallel fiber to Purkinje cell synapses (Doussau et al., 2017); calyx of Held (Lin et al., 2022; Müller et al., 2010)), repetitive stimulation at low frequency (0.1 Hz to 5 Hz) induces synaptic depression. This form of synaptic depression (low frequency depression, or LFD) does not require a previous period of intense presynaptic activity, so that it is unclear whether it can be explained by the classical RRP depletion hypothesis, or by any of the alternative mechanisms mentioned above.

Recent studies suggest that the RRP is made up of a heterogeneous group of SVs, and that the relevant parameter for synaptic plasticity is not the RRP size as such, but rather the proportion between two components of the RRP (primed vs. unprimed or superprimed SVs; loosely vs. tightly docked SVs; or docked vs. replacement SVs) (Blanchard et al., 2020; Doussau et al., 2017; Eshra et al., 2021; Fukaya et al., 2023; Kim et al., 2024; Kobbersmed et al., 2020; Koppensteiner et al., 2022; Lin et al., 2022; Miki et al., 2020, 2016; Taschenberger et al., 2016). These results open new perspectives for the interpretation of STP (Neher, 2023; Neher and Brose, 2018; Pulido and Marty, 2018; Schmidt, 2019; Silva et al., 2021).

Electron microscopy studies of hippocampal synapses in culture showed that, following a single presynaptic action potential (AP) stimulation, the distance between SVs and the active zone (AZ) membrane varies as a function of time in a complex manner (Kusick et al., 2020; Ogunmowo et al., 2025). Following the immediate loss of SVs by exocytosis, the number of SVs directly in contact with the AZ membrane rapidly increases (within ∼15 ms), reflecting the movement of SVs from a cytosolic location to the membrane (‘calcium-dependent docking’). Within ∼100 ms, SVs move back into the cytosol (‘calcium-dependent undocking’). In the following several seconds, SVs partially move forward again to the AZ membrane as the system returns to its basal state. While SV docking has been associated with facilitation (Miki et al., 2016; Schmidt, 2019; Silva et al., 2021), the functional role of subsequent undocking has remained unclear. Here we explore the possibility that undocking could underly LFD.

Testing the undocking hypothesis of LFD necessitates a functional assay of various subclasses of RRP SVs, a notoriously difficult task (Neher, 2015). Our laboratory has developed a method to monitor the release of individual SVs at single AZ synapses between parallel fibers (PFs) and molecular layer interneurons (MLIs) in cerebellar slices (‘simple synapse recording’ (Malagon et al., 2016)). Simple synapse recording has provided models of synaptic facilitation and of high frequency depression based on changes in occupancy of docking sites and associated SV groups inside and outside the RRP (Miki et al., 2016; Tanaka et al., 2021; Tran et al., 2022). In the present work, we show that LFD reflects a specific SV re-equilibration inside the RRP, as well as an inhibition of SV entry into the RRP. We further propose a general model of synaptic depression which explains differences between low and high stimulation frequencies.

Results

Biphasic dependence of paired-pulse ratio on inter-stimulus interval

PF-MLI synapses are considered facilitating (Atluri and Regehr, 1998). In the present work, we ask whether single PF-MLI synapses may turn depressing at low frequency. Using simple synapse recordings at PF-MLI junctions, short trains of 4 presynaptic APs were applied at various inter-stimulus intervals (ISIs) in a scrambled order. Each train was separated from the next by a 10 s-long time interval, considered long enough to reset the system (Tran et al., 2023) (Fig. 1A). The resulting EPSCs were deconvolved using the mean quantal EPSC as kernel to yield the numbers of released SVs per AP (s_i, where i stands for AP number) (Malagon et al., 2016). Results of an example experiment are shown in Fig. 1 B-C. Fig. 1C, top shows average EPSC traces for the first AP (average s₁) and for the second AP at various ISI values. Fig. 1C, bottom shows, for each ISI, the ratio of mean released SV numbers for AP #2 over the corresponding mean for AP #1 (paired pulse ratio; PPR: blue dots; results for AP #3-4 are shown in Supplementary Fig. 1). In this experiment, the PPR defined in this manner displayed values > 1 for ISIs ranging from 10 to 100 ms, and values < 1 for ISIs ranging from 200 to 1600 ms. In group results, the PPR was consistently > 1 for ISI values of 100 ms or less, and < 1 for ISI values of 800 or 1600 ms (Fig. 1D). A biexponential fit to the data (red curve in Fig. 1D) intersected the PPR = 1 line for an ISI near 300 ms. These results show that the synapse changes from facilitating to depressing depending on the ISI. They differ from the single facilitation component previously reported for this synapse (Atluri and Regehr, 1998). This discrepancy presumably stems from differences in experimental conditions (room temperature, stimulation of multiple presynaptic PFs and 2 mM external Ca²⁺ concentration in the previous work, vs. near-physiological temperature, single presynaptic stimulation and 3 mM external Ca²⁺ here; see below). Here, the PPR is the sum of a fast facilitation component and of a slow depression component (PPR = 1 + A_fast exp(-t/τ_fast) - A_slow exp (-t/ τ_slow), with A_fast = 0.94, τ_fast = 230 ms, A_slow = 0.30, τ_slow = 2 100 ms).

Figure 1:

Low frequency depression (LFD) reflects a decrease in DS occupancy but no change in RS occupancy or in IP size

We previously developed a sequential 2-step model of docking at PF-MLI synapses (Miki et al., 2018, 2016; Tran et al., 2022) (Fig. 2A). In this model, each AZ contains a fixed number of docking/release sites, called docking units (DUs) in the present study, with a mean value of 4 DUs per AZ (Miki et al., 2017). Each DU comprises two SV binding sites: a docking site (DS), which is directly attached to the AZ membrane, and an associated replacement site (RS). The RS is able to accommodate an SV regardless of whether the associated DS is occupied or not, meaning that a DU can simultaneously bind two SVs (Silva et al., 2024). The RRP is the sum of the SVs bound to all DUs (RS and DS) in the same AZ. High frequency AP trains induce a flow of SVs from an upstream intermediate pool (IP) to the RRP. Once in the RRP, SVs transit from the RS to the DS, and eventually undergo exocytosis (downward red arrows in Fig. 2A).

Figure 2:

The RS/DS model of Fig. 2A has been used to interpret phenomena of STP during and after high frequency AP trains. In response to stimulation at 100 or 200 Hz, the s_i plot displays an initial increase up to i = 2-4, reflecting synaptic facilitation, followed by a gradual decrease reflecting synaptic depression (Miki et al., 2016; Tran et al., 2022). According to the model, the DS occupancy, δ, increases after a stimulus over its basal value, following Ca²⁺-dependent movement of SVs from RS to DS. This explains the initial facilitation. Subsequent synaptic depression reflects a δ decrease following the gradual exhaustion of IP SVs and consequently, a decrease in the flow of SVs into the RRP. Recovery requires the refilling of the IP, and it occurs on a time scale of hundreds of ms.

To explore the mechanisms underlying STP as a function of AP frequency, we alternated 8-AP stimulations at 100 Hz and at 2 Hz. Immediately after the 2 Hz train, we repeated the 100 Hz train, before repeating the entire protocol again after a pause of 10 s, to allow for synapse recovery (Fig. 2B). According to our previous work (Tran et al., 2022), we can establish the occupancy of the RRP and the intermediate pool (IP) during the low frequency stimulation by comparing the s_i curves of the 100 Hz trains before (control) and immediately after (recovery) the 2 Hz train. Under the hypothesis that the release probability of docked vesicles (p_r) remains constant throughout the AP train, changes in δ, in RS occupancy (ρ) and in IP size are reflected by changes in the number of released SVs for AP #1 (s₁), for AP #2 (s₂), and for the average of AP #5-8 (s_5-8) respectively. This analysis was derived under conditions of elevated release probability. Therefore, in most of our experiments, we used an increased external Ca²⁺ concentration (3 mM) to facilitate the interpretation of the results in terms of changes in RRP and IP occupancy; results obtained under physiological external Ca²⁺ concentration conditions will be reported below.

The results of these experiments are shown in Fig. 2C-E. In keeping with the results of Fig. 1, when stimulating at 2 Hz in 3 mM external Ca²⁺ concentration, a clear synaptic depression was observed (PPR = 0.74 ± 0.05; blue traces in Fig. 2C; absolute SV numbers in Fig. 2C top, and values normalized to control s₁ value in Fig. 2C bottom). The s_i curve for 2 Hz trains could be approximated to an exponential decay with an asymptote of 0.71 ± 0.03 and a time constant of 0.51 ISIs (255 ms; red curve in Fig. 2C bottom). We then compared release during control and recovery trains to investigate the origin of the synaptic depression (Fig. 2D). The value of s₁ at the onset of the recovery train replicated the stable low s_i values observed at the end of the low frequency train. It displayed a decrease with respect to the s₁ value of control trains (to 70 % of this control; p = 0.02, one-tailed paired t-test; Fig. 2C bottom). As s₁ is a proxy for δ under the present experimental conditions, this result suggests that LFD involves a decrease in δ, as previously suggested for high or moderate frequency depression (HFD) (Borges-Merjane et al., 2020; Lin et al., 2022; Miki et al., 2016; Tanaka et al., 2021). Strikingly, s_i values during the recovery train caught up with control values immediately after the first stimulation, so that neither s₂ nor s_5-8 values differed from those of control 100 Hz trains (s₂ = 1.30 ± 0.09 for control and 1.16 ± 0.10 for recovery; p > 0.05; s_5-8 = 5.05 ± 0.68 for control and 5.02 ± 0.63 for recovery; p > 0.05, one-tailed paired t-test; compare yellow and black s_i curves in Fig. 2D; summary graph in Fig. 2E). These results indicate that neither ρ nor the IP size were decreased at the end of the low frequency train. Taken together, the results suggest similarities and differences between the mechanisms of synaptic depression at high frequency, as studied before, and at low frequency, as studied here. In both cases, synaptic depression is accompanied by a decrease in δ. However, while ρ and IP pool size accompany the δ decrease during synaptic depression at high stimulation frequency (Tanaka et al., 2021; Tran et al., 2022), only δ is changed during LFD.

LFD as a function of stimulation frequency and external calcium concentration

Having documented LFD at 2 Hz stimulation frequency and in 3 mM external Ca²⁺ concentration as shown in Fig. 2, we next examined the results of similar experiments when changing the stimulation frequency to 1 Hz, or when reducing the external Ca²⁺ concentration to near physiological concentration (1.5 mM). When stimulating at 1 Hz, LFD was obtained (Supplementary Fig. 2A, blue triangles; fractional response at steady state: 0.78 ± 0.02, n = 5). As was the case with 2 Hz stimulations, the s_i curve recovered immediately after the first stimulus during the recovery train (yellow curves in Supplementary Fig. 2A). Next, we examined synaptic responses at 2 Hz in 1.5 mM external calcium, finding no depression (Supplementary Fig. 2B). At this calcium concentration, synapses facilitate more than at 3 mM (compare control black normalized traces in Fig. 2C bottom in 3mM Ca²⁺, and Supplementary Fig. 2B-C right in 1.5 mM Ca²⁺), potentially causing a longer-lived facilitation that would mask the depression. To test this theory, we performed the same protocol at 0.5 Hz in 1.5 mM Ca²⁺. While the PPR was not significantly < 1 (PPR = 0.89 ± 0.17, n = 5), the rest of the train led to a reduced number of released SVs compared to control s₁ (fractional response at steady state: 0.66 ± 0.09), indicating LFD. An exponential fit to the 0.5 Hz data (red curve in Supplementary Fig. 2C, right) indicated a τ of 2.8 stimuli (5.6 s), slower than that observed for higher calcium concentrations at 2 Hz. The recovery curve (yellow, Supplementary Fig. 2C) indicated a full recovery to control values once high frequency stimulation was resumed.

In conclusion, LFD occurs at near-physiological external Ca²⁺ concentration (1.5 mM) but a lower frequency is required compared to 3 mM Ca²⁺ (0.5 Hz instead of 2 Hz). These experiments further confirm that LFD only involves a decrease in δ, without accompanying changes in ρ or in IP size.

Modelling the dependence of PPR on ISI

The facilitation/depression curve of Fig. 1 matches the docking/undocking sequence reported in flash-and-freeze experiments (Kusick et al., 2020). Therefore, we next asked whether the RS/DS model (Fig. 2A) could account for the dependence of the PPR on ISI on the basis of δ changes (constant p_r hypothesis). The sequence of events following an AP is shown in Fig. 3. Immediately after an AP, docked SVs were released with a probability p_r, resulting in a drop in δ (grey area, Fig. 3A). We modelled subsequent SV movements based on transition rates r_b, r_f, s_b, s_f (Fig. 2A; (Lin et al., 2022; Miki et al., 2018)) with appropriate time-dependent changes (Supplementary Fig. 3, Table 1, and Materials and Methods). This revealed a time window comprised between 10 and 200 ms when δ (red curve) was larger than its resting value of 0.6 (dotted red line, Fig. 3A), while ρ (blue curve) was lower than its resting value of 0.9 (dotted blue line, Fig. 3A). This time period corresponds with the facilitating part of the PPR vs. ISI plot (Fig 3B). Due to the low ρ value, SV entry into the RRP (IP to RS transition) was facilitated. We call this the ‘RS gate open’ configuration (green shade in Fig. 3A). By contrast, during the depressing part of the curve (Fig. 3B, ISI values comprised between 500 ms and 10 s), δ was smaller than its resting value while ρ recovered near its resting value, and the ‘RS gate’ was closed (an SV cannot move into the RS if it is already occupied; red shade in Fig. 3A). The predicted PPR plot (red curve in Fig. 3B) is proportional to the δ curve of Fig. 3A under the assumption of constant p_r. It displays a good fit to the data (blue dots), showing that the model can account for the facilitation/depression sequence of Fig. 1 based on changes in DS occupancy.

Figure 3:

Table 1

Modelling synaptic output in response to AP trains at high or low frequency

Next, we examined whether the same model could account for the results of Fig. 2 (LFD for 8-AP trains and following recovery). To this end, we made concatenations of the previous model over consecutive time segments (Materials and Methods). Data and simulations of 8-AP trains at high frequency (Fig. 4A), low frequency (Fig. 4B), and recovery trains at high frequency (Fig. 4C) illustrate the consequences of the time-dependent changes of DU occupancy shown in Fig. 3. Upper panels (Fig. 4A, B and C) show s_i plots together with simulation results, indicating that the model reproduces the main features of Fig. 2C data. Lower panels (Fig. 4A, B and C) show plots of δ and ρ as a function of AP number. At high stimulation frequency (Fig. 4A), δ increases and remains high throughout an 8-AP train, explaining facilitation (red curve). Meanwhile, ρ decreases below its resting level (blue curve). Because SV movement from the IP into the RRP requires a free RS, this ρ decrease facilitates RRP replenishment. In effect, the RS gate opens at AP #2 and remains open throughout the 8-AP train (Fig. 4A, bottom; Fig. 4D, green shade). At low stimulation frequency, a totally different pattern of changes is observed (Fig. 4B; Fig. 4D, blue shade). δ decreases at AP #2 and remains low throughout the train, explaining LFD. ρ remains close to its resting value of 0.9, so that the RS gate is closed before each AP of the train (Fig. 4B, bottom; note that this figure shows DU parameters just before each AP, and that it does not display the time-dependent DU changes during ISIs illustrated in Fig. 3). These results show that both at high and at low frequency, the proportions of DU states that are reached after the first ISI are kept without major changes throughout the 8-AP train.

Figure 4:

The simulation of the recovery train used the same transition parameters as the control train. The initial ρ and δ were obtained from simulation values at the end of LFD (Fig. 4B, bottom). As can be seen by comparing Fig. 4C, bottom with Fig. 4A, bottom, DUs behave identically to control from AP #2 to AP #8, in agreement with the data. Thus, our model predicts that a synapse where δ has decreased following LFD can return to its basal state within a single ISI of 10 ms when challenged with a high frequency stimulation (Fig. 4D, yellow shade). If a high frequency train is continued the synapse enters the HFD loop (Fig. 4D, green shade).

LFD does not require previous SV release events

The results of Fig. 2 and Fig. 3 suggest that LFD reflects a drop in δ, but they leave the question open as to whether the δ decrease is due to exocytosis, to SV return to RS (undocking), or to both. To better appreciate the share of each of these two mechanisms in LFD, we re-examined LFD experiments depending on the value of s₁. If s₁ is nil, the first theory predicts no synaptic depression in response to the 2^nd AP, because there was no prior loss of SVs by exocytosis. Conversely, a failure of release after AP #1 should not prevent synaptic depression according to the second theory. We set aside EPSC sequences during LFD where the 1^st AP resulted in release failure (s₁ = 0; labelled with a star in Fig. 5A). We pooled together results obtained with ISI values of 500, 800 and 1600 ms, as the PPR did not appear to depend on the ISI value in this range (PPR = 0.74 ± 0.05, 0.88 ± 0.04, and 0.75 ± 0.04 for ISIs of 500 ms, 800 ms and 1600 ms respectively). We calculated separately mean s₂ values for the failure traces (noted <s₂>Is₁=0) as well as overall mean s₂ values (noted <s_2,all>). We obtained two separate LFD values (LFD_fail and LFD_all) by dividing <s₂>Is₁=0 and <s_2,all> with the overall s₁ mean, <s_1,all>. Strikingly, we found that the two LFD values were similar (LFD_fail = 0.84 ± 0.05 vs LFD_all = 0.79 ± 0.03; p = 0.24, unpaired one-tailed Wilcoxon Rank test; Fig. 5B, left), indicating that the value of the LFD does not depend on whether the 1^st AP leads to a failure or to a response. Meanwhile, the two LFD values were significantly < 1 (p = 0.004 for LFD_fail and p < 0.0001 for LFD_all, one-tailed, one-sample Wilcoxon Rank test), indicating that LFD occurs independently of whether the 1^st AP leads to a failure or not. Next, we expanded the above analysis by sorting traces displaying successful EPSCs as a function of s₁. Experimental LFD values were < 1 in all cases: 0.79 ± 0.07 for s₁ = 1; 0.75 ± 0.10 for s₁ = 2; 0.73 ± 0.12 for s₁ > 2; respective one-tailed p values for LFD < 1: 0.002, 0.01 and 0.03 (Fig. 5B, right). We finally asked whether these results were compatible with our undocking model. Simulation results using the model of Fig. 3 gave: LFD_fail = 0.84; LFD_all = 0.73; LFD_{s1 = 1} = 0.77; LFD_{s1 = 2} = 0.70; LFD_{s1 > 2} = 0.59. These values are close to experimental data (Fig. 5B, black dots vs. bars). His agreement is remarkable since the model predicted LFD values without any parameter adjustment. Notably, the finding of identical experimental and simulated LFD_fail values (0.84 in both cases) confirms that the undocking mechanism of Fig. 3 accounts for LFD in the absence of any previous SV release.

Figure 5:

In conclusion, the above results indicate that LFD occurs independently of the amount of SV release following AP #1. They support the hypothesis that LFD reflects a shift of SV occupancy favoring RSs over DSs (undocking), rather than RRP depletion caused by exocytosis.

Long AP trains at low frequency reveal two phases of LFD occurring on widely different time scales

Earlier studies showed a slow onset LFD during long trains at low frequency (Abrahamsson et al., 2007; Doussau et al., 2010; Rudolph et al., 2011). We therefore searched for an additional slow component LFD at PF-MLI synapses. In each experiment, a series of control trains (8-APs @ 100 Hz) was presented first. Next, a long low frequency train was applied (200 APs @ 2 Hz in Fig. 6A), followed by a recovery train (50 or 100 APs @ 100 Hz; Fig. 6B; protocol in Fig. 6A, inset). Numbers of released SVs per AP (s_i) were normalized with respect to the s₁ value obtained for the control trains. During the long train, a plot of the normalized s_i (indicated as s_i/ctrl s₁) vs AP # displayed an initial decrease followed by a gradual depression (Fig. 6A). We call this initial drop and the subsequent linear decrease the first and second phase of LFD respectively. A linear fit to these data (red line in Fig. 6A) indicated a first phase depression (within the first 10 APs) to 75 ± 2 % of the control, consistent with the previous results obtained with paired stimulation and short train protocols (Figs. 1 and 2). The second phase of the depression exhibited a slope of -0.15 ± 0.02 % per AP, such that the value reached for the linear fit at the end of the 200-AP train was 45 % of the control. A similar LFD was seen at 1 Hz with a first phase drop to 51 ± 2 % of control and a slope of -0.05 ± 0.02 % reaching 40 % of the control at the end of the 200-AP train (n = 17 trials from 6 cells).

Figure 6:

We next investigated δ, ρ and IP pool size at the end of long low frequency trains with the previous s₁/ s₂/ s₅-s₈ approach. As expected, the first response of the recovery train was severely depressed compared to s₁ of control trains (recovery s₁ = 37.6 ± 20.2 % of control s₁), but s_i plots during control and recovery trains were indistinguishable for i values of 2 or more (s₂ ratio: 1.01 ± 0.20; s_5-8 ratio: 1.19 ± 0.23; Fig. 6C-D), indicating again a drop in δ without changes in either ρ or IP size.

The s₁/ s₂/ s₅-s₈ analysis assumes a constant p_r. To test this hypothesis, we determined the parameters of the RS/DS model by fitting responses to control 8-AP trains (Fig. 6E). Next, responses to the first 8 APs of the recovery train were predicted by keeping all parameters the same as control except for either δ or p_r. In two different simulations, δ and p_r were determined to fit the s₁ point of the recovery curve (Fig. 6F, red and black curves respectively). While the δ decrease prediction closely matches the data (Fig. 6F, yellow curve), the p_r decrease prediction fails to simulate the recovery train. Altogether, these results indicate that the slow phase of LFD illustrated in Fig. 6A, like the rapid phase shown in Fig. 2C, reflects a δ decrease without significant changes in p_r, in ρ or in IP size.

In these experiments the delay between the LFD train and the recovery train was variable, providing indications on the synapse recovery time. In sharp contrast with the almost instantaneous recovery of s_i for i > 1 when returning to 100 Hz frequency, depression was still obtained when presenting a new AP after rest periods of up to 1 min after the end of the LFD train (Supplementary Figure 4). These results suggest that SV re-equilibration within the RRP occurs very rapidly under elevated calcium concentration, as presumably occurs during the recovery trains, but very slowly under resting calcium concentration.

Finally, a full recovery train (50 or 100 APs @ 100 Hz) elicited a deep synaptic depression (Fig. 6B), similar in extent and in time course to the depression observed for stand-alone trains at high frequency. An exponential fit to these data (red curve in Fig. 6B) displayed a time constant of 412 ms (41.2x the ISI) and an asymptotic value of 13% of the control s₁ value. This type of depression, unlike LFD, was previously shown to involve a decrease in ρ and in IP size in addition to δ (Tran et al., 2022).

LFD is not due to stimulation failures

One tacit assumption of our analysis so far is that LFD reflects a decreased synaptic response to presynaptic stimulations. We next asked whether LFD could instead originate in erratic stimulation failures. This possibility was plausible since the failure rate increased markedly during the second phase of LFD (Supplementary Figure 5A). To test it, we compared the changes in the distributions of released SV numbers predicted from two LFD models. In the first model, during LFD, each DU releases less SVs when stimulated, presumably due to a reduction in δ. In the second model, a fraction of the stimulations fails to elicit a presynaptic AP during LFD, but the probability of release for each stimulated synapse remains constant (Materials and Methods). Both models were constrained to reproduce the observed mean number of failures during LFD. Observed distributions of released SV numbers during LFD were compatible with the decreased δ model but not with the stimulation failure model (Supplementary Figure 5). We conclude that stimulation failures cannot account for LFD, consistent with the above interpretation that LFD reflects a drop in δ, as well as with previous results in other preparations (Saviane et al., 2002; Silverman-Gavrila et al., 2005)

LFD is not blocked by postsynaptic BAPTA

At synapses between PFs and Purkinje cells, low frequency stimulation leads to a gradual synaptic depression that has been attributed to a transsynaptic mechanism (Casado et al., 2000). Since the simple synapse method is based on SV counts rather than EPSC amplitude, there is no doubt that the expression of LFD at PF-MLI synapses is presynaptic. Nevertheless, as AMPAR activation elevates the postsynaptic Ca²⁺ concentration (Soler-Llavina and Sabatini, 2006), we included BAPTA (10 mM) in the MLI recording solution to test the possibility of a retrograde LFD mechanism that would be driven by postsynaptic Ca²⁺. We applied a long 2 Hz train (100 APs @ 2 Hz) followed by a short 8-AP recovery train (experimental protocol in Supplementary Figure 6, insert). The result is very similar to Fig. 6 data (compare Supplementary Figure 6 to Fig. 6A and C) with a first phase drop to 65 ± 3 % of the control and a second phase slope of -0.2 ± 0.05 % per AP. In this case recovery trains were applied immediately following the LFD train (whereas waiting times of 10 s of s were included in the experiments of Fig. 6), and the recovery is identical to that observed in Fig. 6C, suggesting likewise a reduction in δ with no associated change in ρ or IP size.

The similarity of results with and without a postsynaptic calcium buffer supports a purely presynaptic mechanism for LFD.

LFD is not due to decreased Ca²⁺ transients

Our recovery results and simulations suggest that LFD is due to a decrease of δ, not of p_r (Fig. 6F). In line with this notion, earlier work indicated no change in Ca²⁺ transients at PF-MLI synapses for paired stimulations at short ISIs (Brenowitz and Regehr, 2007; Malagon et al., 2020; Miki et al., 2016). Nevertheless, it seemed possible that a different situation would prevail with multiple stimulations and/or at longer ISIs.

To explore potential changes of presynaptic Ca²⁺ signaling during AP trains, we preloaded individual granule cells with the calcium dye OGB6F (1-2 mM in pipette solution; duration of preloading: 90 s; (Rebola et al., 2019)). After allowing diffusion of the calcium dye in the axon, we imaged single PF varicosities using two-photon laser microscopy (Fig. 7A, and Supplementary Information), and we measured Ca²⁺-dependent changes in fluorescence during 50-AP train stimulations at 1 Hz, or 200-AP train stimulations at 2 Hz, using stimulation conditions similar to those of the simple synapse experiments of Fig. 6. To avoid photodamage and photobleaching, imaging was interrupted after the 5th AP in a train, and it was resumed for the last 5 APs of the train. Traces were averaged from 3-7 trials in each varicosity to improve the signal-to-noise ratio (Fig. 7A2). Averaged traces across varicosities revealed little changes in Ca²⁺-dependent fluorescence transients during trains either at 1 Hz (Fig. 7B) or at 2 Hz (Fig. 7C). ΔF/F₀ values for 10 varicosities from 8 granule cells at 1 Hz were: 19.1 ± 3.9 % for the 1st AP; 20.4 ± 2.8 % for the 2nd AP; 19.3 ± 2.8 % for the last AP; for 9 varicosities from 6 granule cells at 2 Hz: 19.4 ± 2.9 % for the 1st AP; 17.9 ± 1.6 % for the 2nd AP; 16.4 ± 1.9 % for the last AP; p > 0.05 in 1 vs. 2 and 1 vs. last paired comparisons at both frequencies, one-tailed Wilcoxon Rank test (Fig. 7D). These data indicate that Ca²⁺-dependent fluorescence transients were the same in response to the first, the second, or the last AP in a train. In the same data set, we found no difference between basal fluorescence levels at the end vs. the beginning of the trains at 1 Hz (respectively, 107 ± 12 a. u. vs. 103 ± 10 a. u.; p > 0.05, one-tailed Wilcoxon Rank test; black markers in Fig. 7E1), but we found an augmentation of basal fluorescence at the end of the trains compared to the pre-stimulation value at 2 Hz (118 ± 10 a. u. vs. 111 ± 9 a. u.; p = 0.026, one-tailed Wilcoxon Rank test; blue markers in Fig. 7E1). At 2 Hz, after the end of the train, the fluorescence trace relaxed to near its pre-stimulation value with a time constant of 750 ms, confirming that the basal Ca²⁺ concentration had increased during the train (Fig. 7E2).

Figure 7:

In conclusion, we find no evidence of decreased Ca²⁺ transients during trains, either at 1 Hz or at 2 Hz. This is in line with our earlier conclusion that p_r changes do not contribute strongly to LFD. We find a gradual increase of pre-stimulation Ca²⁺ concentration with 200 APs at 2 Hz, but not with 50 APs at 1 Hz. This effect may contribute to the second phase of LFD at 2 Hz.

Doublet stimulations inhibit LFD

Our model predicts a high ρ value during LFD that inhibits DU replenishment by preventing the transition from the IP to the RS (RS gate closed). We next tested the prediction that during a low frequency train, ρ remained elevated while δ decreased (Fig. 4B), by giving two successive stimuli with a 10 ms interval instead of single stimuli (doublet stimulations; Fig. 8A-B). According to the s₁/ s₂/ s_5-8 analysis, while the time course of the first response, s₁, should give the evolution of δ, the time course of the second response, s₂, should give the evolution of ρ (Tran et al., 2022). With the same experiments, we intended to test whether doublet stimulations would open transiently the RS gate and thus interfere with the time course of LFD, as suggested by earlier work at the PF-Purkinje cell synapse (Doussau et al., 2017).

Figure 8:

We performed these experiments with a longer than usual ISI (5 s) to provide ample time for SVs to equilibrate after each stimulus, either singlet or doublet. Control single AP train stimulation experiments at 0.2 Hz revealed LFD (exponential fit of first 12 normalized s_i points had an asymptotic value of 0.77 ± 0.14 and a time constant of 3.2 s). We then examined the extent of LFD in response to the first 20 doublet stimulations at low frequency. We found that s₁ values were reduced compared to control (normalized s₁ value during 20 first trials at 0.2 Hz: 0.80 ± 0.10; p = 0.008, two-tailed, one-sample t-test; Fig. 8C). These results indicate that doublet stimulations at 0.2 Hz produced significant LFD when considering the response to the 1^st stimulus. By contrast, s₂ values were not affected (s₂ values normalized to control s₁: 1.18 ± 0.11 for control 100 Hz trains; 1.11 ± 0.10 during 20 first doublet trials; p > 0.05, two-tailed, one-sample t-test; Fig. 8C). These results confirm our previous conclusion that LFD involves a decrease in δ without significant changes in ρ.

Next, we compared the plots of normalized released SV numbers as a function of AP number (i) for singlet stimulations (s_is; light blue), for the 1^st AP of doublet stimulations (s_id1; dark blue), and for the 2^nd AP of doublet stimulations (s_id2; orange; Fig. 8 D-E). In each case, the curves could be approximated to linear decay curves as a function of i (y = α – βi), but the line parameters differed for the 3 curves. Values of α were similar for s_id1 and for s_is (0.79 ± 0.07 and 0.88 ± 0.11 respectively; p = 0.53, two-tailed, two-sample t-test). This indicates that the first phase of LFD occurs with a similar amplitude for singlets and for doublets. On the other hand, values of β were markedly smaller for s_id1 than for s_is (β_d1 = 0.0004 ± 0.0014 vs. β_s = 0.0063 ± 0.0018; p = 0.011, two-tailed, two-sample t-test; Fig. 8D and Fig. 8E, top). These results indicate that the second phase of LFD is less pronounced for doublet stimulations than for singlet stimulations, with a slope ratio near 15-fold. The results for d₂ indicated a return of ρ values close to control at the end of each 5 s ISI throughout the train (orange curve in Fig. 8E, bottom; α_d2 = 0.97 ± 0.04; β_d2 = 0.001 ± 0.001), which suggests that the RS gate was closed at the end of each ISI after transitorily opening during the preceding doublet.

Overall, the responses to the first stimulations in doublets display less LFD than singlets, and the responses to the second stimulations in doublets display no LFD at all. The results are consistent with the notion that LFD involves a closing of the RS gate, which is transiently alleviated by doublet stimulations.

Discussion

In the present work, we use the SV counting method at simple PF-MLI synapses (Malagon et al., 2016) to investigate cellular mechanisms underlying LFD. Our results reveal similarities, but also significant differences, between these mechanisms and those previously described for moderate or high frequency synaptic depression (Lin et al., 2022; Tran et al., 2022). Based on our results, we propose that LFD reflects a decrease in δ following SV undocking. We further expand this model to explain how the mechanisms of synaptic depression vary as a function of stimulation frequency.

PPR as a function of ISI

A traditional view of synaptic physiology has been that synapses may be either facilitating or depressing. This view is clearly an oversimplification, however, as many synapses display a combination of facilitating and depressing properties (Dittman et al., 2000). Here we show that PF-MLI synapses exhibit PPR values > 1 for ISI values comprised between 10 and 200 ms, and PPR values < 1 for ISI values comprised between ∼500 ms and ∼1500 ms (Fig. 1). Importantly, the shift from facilitation to depression is apparent not only in group results (Fig. 1D), but also in individual experiments (Fig. 1C). A similar mix of facilitation at short ISI and depression at long ISI has been described in a number of central synapses (Chiu and Carter, 2024; Doussau et al., 2017). Following our analysis of simple synapse recordings, and in view of zap-and-freeze results in hippocampal cultures, we propose that facilitation and LFD are linked together in a docking-undocking sequence following AP stimulation (Kusick et al., 2020).

SV movements during HFD and LFD

Based on the present study and on earlier work (Kusick et al., 2020; Miki et al., 2016; Tran et al., 2022), we propose a global model for HFD and LFD as outlined in Fig. 4D. If an AP train is presented with short ISIs ( ∼ 10 ms), a rapid sequence of release/docking/replenishment occurs (open RS gate). The arrival of a new AP induces movement of SVs forward (from DS to exocytosis, from RS to DS, and from IP to RS). If a docked SV fails to be released, it does not undock, because a new SV rapidly occupies the corresponding RS. These events lead to facilitation at first. IP depletion eventually reduces SV influx into the RRP, such that the IP size becomes rate limiting at the end of the train. This accounts for HFD (HFD loop in Fig. 4D, green shade). If an AP train is presented with longer ISIs (0.5 s to 5 s), by contrast, release and replenishment are reduced since many DUs have their DS empty and their RS occupied at the time of the AP (closed RS gate). During the closed RS gate period, ρ is ∼0.9, so that the entry rate of SVs from the IP into the RRP, which is proportional to (1-ρ), is reduced 10-fold by RS occupation. By comparison, ρ is ∼0.5 when the RS gate is open. The larger ρ value results in a net 5-fold slower entry rate when the RS gate is closed compared to when it is open. With longer ISIs, SVs move from the RS to the DS after each AP, but they have the time to move back to the RS before the next AP. Therefore, DUs tend to undergo an idle docking/undocking loop without release, leading to LFD (LFD loop in Fig. 4D, blue shade).

AP induced docking and undocking happen in overlapping time frames. In Fig. 1 the highest amount of docking is seen at ∼10 ms ISI (it is also present at 5 ms: (Miki et al., 2016; Silva et al., 2024)). Undocking on the other hand is maximal more than 1 s after the AP. For ISIs in between there is a mix of docking and undocking. The PPR recovery observed at ISIs of ∼200 ms is only apparent because in this period undocking is still ongoing.

When stimulation is switched to high frequency after LFD, SV release recovers within 10 ms to the same level as during high frequency control trains (Fig. 2D; Fig. 6C). This happens because the SVs that are in the RS dock with each AP and are still in the DS when a second AP is presented rapidly (HF recovery in Fig. 4D, yellow shade). This explains why the s₂ of a recovery train is equivalent to the control value. If high frequency stimulation is continued the synapse will behave identically to a naïve synapse for the remainder of the train (Fig. 2D; Fig. 6C) and it will eventually undergo HFD (Fig. 6B; HFD loop in Fig. 4D, green shade).

It has long been recognized that synaptic activity enhances RRP replenishment (Neher and Sakaba, 2008). While a Ca²⁺-dependent rate of SV entry into the RRP has been invoked as underlying mechanism, other explanations cannot be ruled out (Eshra et al., 2021; Miki et al., 2020; Ritzau-Jost et al., 2018; Ritzau-Jost et al., 2014). Transient opening of the RS gate, as suggested here, offers a novel mechanism explaining activity-dependent RRP replenishment.

LFD involves a specific δ decrease rather than across the board RRP decrease

Our conclusion that LFD results from a δ decrease is largely based in our analysis of high frequency recovery trains. While this approach is model-dependent, as it assumes the RS/DS model of Fig. 2A, our conclusion is supported by two additional lines of evidence. First, we find that LFD occurs even after failures, and that the extent of LFD does not depend significantly on s₁ (Fig. 5). These findings argue against the standard RRP depletion mechanism. By contrast, the docking-undocking sequence of Fig. 4D (main sequence and LFD loop) accounts for our results as it predicts that the DU state at the end of a long ISI depends little on the response to the previous stimulation (simulation in Fig. 5B). Secondly, we find that LFD is alleviated when using doublet rather than singlet stimulations (Fig. 8). Again, this result is inconsistent with an across the board RRP depletion model, but it can be readily explained by the RS/DS model. During LFD with singlet stimulations, the RS/DS model predicts a gradual δ decrease accompanied by a high ρ (Fig. 3). Doublet stimulations change this pattern in two ways. Firstly, they transiently lift the RS block that prevents RRP refilling with singlet stimulations, leading to enhanced RRP replenishment. The second AP of the doublet arrives in the period where SVs are still present in the DS, releasing a fraction of them, and promoting SV movement from the RS to the DS. This means that SVs coming from the IP are allowed, in turn, to bind the now-free RS. Secondly, doublet stimulations prevent SV undocking. This happens as the incoming SV at the RS locks the SV that is bound to the DS in its docked position. The combination of these two effects leads to an inhibition of the slow phase of LFD during doublet stimulations. Of note, this complex RS/DS interaction is only possible because the RS/DS model allows simultaneous occupation of both RS and DS by distinct SVs (Silva et al., 2024).

Alternative presynaptic depression mechanisms

Simple synapse recording and deconvolution analysis give us a direct estimate of the number of SVs released, bypassing any possible postsynaptic effects. Expression of LFD as described in the present work is therefore presynaptic. Furthermore, our BAPTA experiments (Supplementary Fig. 6) argue against a participation of the postsynaptic compartment to LFD induction, so that both induction and expression of LFD are likely presynaptic. We consider here other presynaptic mechanisms than undocking that have been proposed to underlie LFD. A p_r reduction has been suggested (Wölfel et al., 2007; Wu and Borst, 1999), either resulting from a direct reduction in Ca²⁺ current (Xu and Wu, 2005) or from the buffering effect of calretinin (Bolshakov et al., 2019). The rapid reversal obtained during recovery trains in the present work is not consistent with either of these hypotheses, as both Ca²⁺ channel inhibition and a buffering effect would require hundreds of ms to seconds to recover. Furthermore, the results and simulations shown in Fig. 6F suggest a change of δ but not of p_r during LFD. Finally, our Ca²⁺ measurements in presynaptic varicosities do not suggest a difference in Ca²⁺ entry following repetitive low-frequency stimulation although there is a change in basal calcium during long trains that could be linked to the second, slow phase of LFD (Fig. 7). Of note, Ca²⁺ measurements have also been performed in relation to LFD in neocortical L4-L2/3 synapses with synaptic depression and no change in Ca²⁺ signal for paired stimuli was reported at ISIs of 500 ms (Chiu and Carter, 2024). In Aplysia, silencing of docking sites was suggested as the reason behind synaptic depression (Doussau et al., 2010), however a decrease in N would not predict the rapid return of high frequency release when recovery trains are applied after LFD.

In conclusion, it is unlikely that a p_r or N reduction underlies LFD. Overall, a δ reduction best explains LFD and its fast recovery.

Onset and offset kinetics of LFD

In some studies LFD has been reported to develop almost immediately, after one or a few stimulations (Chiu and Carter, 2024; Rudolph et al., 2011) whereas in others it develops much more slowly as a function of stimulation number (Abrahamsson et al., 2007; Doussau et al., 2017). In the present work we observe an immediate reduction in synaptic output (Fig. 2) followed by a very slow phase which does not reach a steady state after 200 APs at 2Hz (Fig. 6A) and 80 APs at 0.2 Hz (Fig. 8D). After a short train, in which only the initial phase is observed, and after a long train, where both phases are present, we find a reduction in DS occupancy but no changes in RS or IP occupancy (Fig. 2E and 6D). This leads us to propose that LFD comprises two phases with different onset kinetics.

We find that LFD fails to recover significantly following a rest period of up to 1 minute (Supplementary Fig. 4). These results are consistent with earlier studies showing that for sparse stimulations, LFD recovery occurs on a time scale of minutes (hippocampal synapses (Abrahamsson et al., 2007) and Aplysia (Doussau et al., 2010)). If, however, a high frequency train is presented, the synapse recovers from LFD almost instantaneously, within one single ISI (range 5-20 ms (Doussau et al., 2017, 2010); present work, Fig. 2, 6 and 8). As seen in Fig. 4D, the high frequency recovery loop after LFD is identical to the HFD loop: presenting new APs at short ISI does not leave enough time for undocking and immediately brings the synapse back to the HFD mode. Altogether, LFD places the synapse in a depressed state that is stable if stimulation is absent or infrequent, but that is immediately reversed if a high frequency train is presented. In our model, this dichotomy occurs because the SV-RS association is stable at rest, when presynaptic Ca²⁺ concentrations are low (RS gate closed) but is unstable during high frequency AP trains, when presynaptic Ca²⁺ concentrations are elevated (RS gate open). In contrast to LFD recovery, HFD recovers within seconds following the end of a high frequency train, reflecting the gradual replenishment of the IP (Tran et al., 2022).

Frequency-independent synaptic depression

In the present work, we observe similar extents of LFD at the end of the first phase for stimulation frequencies of 0.2 Hz (0.77 ± 0.14), 1 Hz (0.78 ± 0.02), and 2 Hz (0.71 ± 0.03). These results indicate that at PF-MLI synapses, the first phase of LFD results in a largely frequency-independent steady state synaptic output in the frequency range 0.2 Hz to 2 Hz. It has been likewise noted in earlier studies that the extent of synaptic depression at steady state is largely independent of stimulation frequency in a certain frequency range, albeit in a higher range than in the present work (10-100 Hz at Purkinje cell to deep cerebellar nuclei synapses (Turecek et al., 2016); 5-50 Hz at the calyx of Held (Lin et al., 2022)). These results have been interpreted either by a compensation between an increased facilitating component and an increased depressing component at higher frequencies (Turecek et al., 2017, 2016), or by a time limitation on calcium-dependent movement of SVs from a primed to pre-primed location, so that this movement is regulated by AP number rather than by time (Lin et al., 2022).

Molecular mechanisms of docking and undocking

The molecular mechanisms of docking and undocking are under intense scrutiny. Consistent with a role in AP-dependent docking, synaptotagmin 7 (syt7) is associated to facilitation in different preparations (Huson and Regehr, 2020; Jackman et al., 2016; MacDougall et al., 2018). Using zap-and-freeze electron microcopy in hippocampal cultures revealed that syt7 KO showed no rapid docking ∼10 ms after an AP, which was translated into a depressing phenotype when trains of APs are applied (Wu et al., 2024). In neocortical synapses, a comparison of PPR curves in WT and syt7 KO showed no facilitation at short ISIs and a deeper depression at longer ISIs in the KO, with the same recovery time frame (5 s) (Chiu and Carter, 2024), supporting a role of syt7 in AP-dependent docking, but not in undocking or baseline docking dynamics. Recently, intersectin-1 and endophilin A1 were reported to stabilize SVs in the replacement site (RS) (Ogunmowo et al., 2025), indicating a role in docking/undocking and possibly in LFD. Munc-13 is another likely candidate for docking (Kusick et al., 2022), as its deletion reduces the number of docked SVs and changes their relative distance to the plasma membrane (Imig et al., 2014; Siksou et al., 2009). In the calyx of Held, RRP recovery after high frequency facilitation and depression depends on both Munc-13 and synaptotagmin 3 (syt3)(Weingarten et al., 2022). Modeling suggested that both proteins are required to explain AP-dependent docking.

In conclusion, multiple proteins have been proposed as candidates for AP-dependent docking; in our preparation, Munc-13, syt3 and 7 are likely candidates. By contrast, the molecular mechanisms behind undocking remain largely unknown. It is likely that some of the STP heterogeneity arises from variations in protein expression and regulation amongst synapses, so that the docking/undocking kinetics are not exactly the same. However, LFD is present in multiple preparations, and may be a general mechanism.

Functional implications

In vivo MLI recordings show very infrequent EPSCs in the absence of PF input (< 1 Hz), while sensory stimulation results in short EPSC trains at high frequency (Chadderton et al., 2004; Jörntell and Ekerot, 2003). LFD as described here likely contributes to lower the frequency of background EPSCs, as it inhibits the responses to infrequent presynaptic APs. However, in view of the rapidity of the post-LFD recovery, full EPSC responses are likely restored as soon as a high frequency train of presynaptic APs is presented. This should increase the signal-to-noise ratio in a burst vs. background signaling mode and should improve the sensitivity of the cerebellar circuitry to external inputs. As many brain neurons display low background firing rates (< 1 Hz) in vivo (Buzsáki and Mizuseki, 2014; Margrie et al., 2002), a similar signal-to-noise improvement could occur in various brain regions following LFD.

Materials and Methods

Experimental model

C57BL/6 mice (12–16 days old; either sex) were used for preparation of acute brain slices. Animals were purchased from Janvier Laboratories (RRID:MGI:2670020). They were housed and cared for in accordance with guidelines of Université Paris Cité (approval no. D 75-06-07; animal rearing service of the BioMedTech Facilities, INSERM US36, CNRS UMS2009).

Slice preparation

200-μm thick sagittal slices were prepared from the cerebellar vermis as follows. Mice were decapitated under anesthesia and the cerebellar vermis was carefully removed. Dissections were performed in ice-cold artificial cerebrospinal fluid (ACSF) which contained the following (in mM): 130 NaCl, 2.5 KCl, 26 NaHCO₃, 1.3 NaH₂PO₄, 10 glucose, 1.5 CaCl₂, and 1 MgCl₂. Slices were cut using a vibratome (VT1200S; Leica) and incubated in ACSF at 34°C for at least 60 min before being used for experiments.

Synaptic electrophysiology

Whole-cell patch-clamp recordings were obtained from MLIs (comprising both basket and stellate cells). The extracellular solution contained (in mM): 130 NaCl, 2.5 KCl, 26 NaHCO₃, 1.3 NaH₂PO₄, 10 glucose, 3 CaCl₂ (1.5 in the experiments where it is stated), and 1 MgCl₂ (pH set to 7.4 with 95% O₂ and 5% CO₂). The internal recording solution contained (in mM): 144 K-gluconate, 6 KCl, 4.6 MgCl₂, 1 EGTA (ethylene glycol-bis tetraacetic acid), 0.1 CaCl₂, 10 HEPES (4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid), 4 ATP-Na, 0.4 GTP-Na (pH 7.3, 300 mosm/l). Gabazine (3 μM) was included in all experiments to block GABA_A receptors. Recording temperature was 32–34°C. To establish a single PF–MLI connection, puffs using a glass pipette filled with the high K⁺ internal solution were applied in the granule cell layer to identify a potential presynaptic granule cell. The same pipette was then used for extracellular electrical stimulation of the granule cell, with minimal stimulation voltage (stimulus voltage was higher than threshold by ∼5 V; range: 10–40 V; 150 μs duration). If the resulting EPSCs, particularly those that occurred after a short stimulation train, were homogeneous, then it was likely that only one presynaptic cell was stimulated. Recordings were only accepted as a simple synapse recording after analysis if the following three criteria were satisfied (Malagon et al., 2016): (1) the EPSC amplitude of the second release event in a pair was smaller than that of the first, reflecting activation of a common set of receptors belonging to one postsynaptic density; (2) all the EPSC amplitudes followed a Gaussian distribution with a coefficient of variation < 0.5; and (3) the number of release events during the baseline recording was stable.

Quantification and statistical analysis

Offline analysis of electrophysiological data and statistical analysis was performed using Igor Pro (WaveMetrics). Data are expressed as mean ± standard error of the mean (SEM). Error bars in all graphs indicate SEM. Details of statistical tests are described in the text or figure legends. Student’s t-test was used when the distribution was Gaussian; Wilcoxon Rank test was used otherwise. One-tailed tests were used when the sign of the effect was predicted before performing the experiments; two-tailed tests were used otherwise. Statistical significance was accepted when p < 0.05. The number of trials and cells (some cells have multiple trials) for each experiment is indicated by a lowercase n. Cells were derived from different animals. Each experiment started with control recordings of 8-AP trains @ 100 Hz, with 10 s inter-train intervals. At least 25 trials were performed for this initial control run before proceeding to specific stimulation protocols. For PPR experiments (Fig. 1) 4-AP trains were applied with varying ISIs (20, 100, 200, 400, 800, 1600 and 3200 ms) in a scrambled order, with a rest period of 10 s between trains (Fig. 1A, right); only synapses with minimum 20 repetitions of this protocol were considered for analysis. For Fig. 2, 8-AP trains at alternating frequencies (100 Hz and 2 Hz) were applied. A 100 Hz recovery train was applied immediately after the end of each 2 Hz train (see Fig. 2B); a minimum of 20 repetitions of the protocol were obtained for each synapse. For these two protocols the repetitions of each synapse were averaged together and synapse results were averaged for group results. For Fig. 6 a 200-AP train was applied at low-frequency (2 Hz), followed by a 50-AP or 100-AP high-frequency (100 Hz) recovery train (Fig. 6A). The time interval between the last AP in the LFD train and the 1st AP in the awakening train varied in the experiments (Supplementary Fig. 4). Synapses were considered with minimum one full trial (low frequency + recovery), the maximum number of trials performed in one synapse was 4. Trials were averaged for group results. For Fig. 8, 80-AP trains were applied at 0.2 Hz with either a single or a double stimulus (Fig. 8A), trials were averaged together for group results. In 5 cells singlets and doublets were applied sequentially. For all protocols a new control run was performed when possible, at the end of the experiment to verify functional stability of the synapse. For each set of experiments the numbers of repetitions are stated in the figure legend.

Decomposition of EPSCs

The time of occurrence and the amplitude of individual release events were determined based on deconvolution analysis, as described previously (Malagon et al., 2016). In brief, for each synapse, an mEPSC (miniature excitatory postsynaptic current) template was obtained from delayed release events recorded after the end of AP trains. This template was fitted with a triple-exponential function with five free parameters (rise time, peak amplitude, fast decay time constant, slow decay time constant, and amplitude fraction of slow decay). These five parameters were then used for deconvolution of the template and of individual data traces, producing a narrow spike (called spike template) and sequences of spikes, respectively. Next, each deconvolved trace was fitted with a sum of scaled versions of the spike template, yielding the timing and amplitude of each release event. The amplitude was further corrected for receptor saturation and desensitization, using the exponential relationship between individual amplitudes and the time interval since the preceding release events. Events that were at least 1.7 times larger than the average mEPSC were split into two (Miki et al., 2017). s₁ was determined as the number of SVs released within 5 ms after the first AP of a train. s₂ was the number of SVs released within 5 ms after the second AP and s_5-8 was the cumulative number of SVs released 5ms after each stimulus for AP# 5 to 8. As explained in Tran et al. (2022) (Tran et al., 2022), these 3 parameters are proxies respectively for the occupancy of DS (δ), the occupancy of RS (ρ), and the IP size.

Calcium measurements

To perform Ca²⁺ imaging of single PF varicosities we pre-loaded granule cells with a calcium indicator. A whole-cell recording was made with a K⁺ gluconate-based solution containing 166 K gluconate, 4.1 MgCl₂, 9.9 HEPES-K, 0.36 Na-GTP and 3.6 Na-ATP, 1-2 mM of the calcium probe Oregon Green BAPTA 6F. The neurons were held in I-clamp mode with currents of -5 to -30 pA injected to keep the average resting potential close to -90 mV for 1.5 to 2 minutes after which the pipette was removed. Obtaining an outside-out patch with > 5 GΩ resistance indicated successful pipette removal, and the granule cell was left for approximately 15 minutes to allow for dye diffusion into the axonal compartment. After this waiting time, a 2-photon laser scanning microscope (Tan et al., 1999) was used to identify an axonal varicosity. A glass pipette with tip diameter around 1 μm, filled with extracellular saline (in mM: 145 NaCl, 2.5 KCl, 10 K-HEPES, 2 CaCl₂, 1 MgCl₂), was introduced into the slice at the depth of the varicosity and placed 10-20 μm from the axon. The stimulation protocol consisted of 1 train of 8 pulses (100 μs duration, 10 to 40 V) delivered at 100 Hz, followed by either 50 stimulations at 1 Hz or 200 at 2 Hz and a new 8 pulse train at 100 Hz applied 30 seconds after the end of the long train. This cycle was performed 3 to 7 times. Fluorescence was acquired at dwell time of 5 ms/frame through raster scans of 5 by 5 μm fields encompassing the axonal varicosity. For each long stimulation train that was applied, we recorded two fluorescence traces. One trace covered the first 5 APs of the train (trace duration: 5 s at 1 Hz stimulation frequency, and 2.5 s at 2 Hz stimulation frequency). We next interrupted laser illumination and scanning (during 40 s at 1 Hz, and during 95 s at 2 Hz) to minimize photobleaching and photodamage. We resumed two-photon imaging near the end of the train, collecting another trace covering the last 5 APs. Time dependent fluorescence signals were analyzed in the pixels encompassing the varicosity and quantified in terms of changes relative to pre-stimulus values (ΔF/F₀ expressed in %) with software written in the IGOR-Pro programming environment (Wavemetric, Lake Oswego, OR, USA). To avoid possible errors due to tissue movement, focusing was performed twice: just before the first trace, and again before the second trace. To estimate changes in baseline Ca²⁺ concentration, we compared basal fluorescence averages (calculated just before the 1st AP in each trace) for early and late traces across trials for all experiments. This revealed an apparent (non-significant) increase by 4 % of the basal fluorescence at 1 Hz stimulation frequency, and a significant 7 % increase at 2 Hz stimulation frequency (Fig. 7E1). In view of this baseline shift (F_Inc), the second ΔF/F₀ trace needed a correction. In this trace, the basal fluorescence used to calculate ΔF/F₀ was corrected according to F_0corr = F_0app – F_Inc, leading to the relation ΔF/F_0corr = ΔF/F_0app * (1 + Finc /F₀) where ΔF/F_0app and ΔF/F_0corr are the apparent and corrected values of ΔF/F₀ for the late trace, and the ratio F_Inc/F₀ is 4 % at 1 Hz and 7 % at 2 Hz. For each train of stimuli, ΔF/F₀ values were averaged over 5 to 7 repetitions at each varicosity and subsequently data from all varicosities were pooled together to generate the average signal as a function of time from which mean peak values were extracted.

Model and simulation of SV docking and release

Monte Carlo simulation of two-step SV docking and release (Fig. 3) was performed using Igor Pro as described previously (Miki et al., 2016; Silva et al., 2024; Tran et al., 2022). The model is created by assuming a fixed number (N) of docking units (DUs) with 2 binding sites for each DU (RS and DS; Fig. 2A), each of which can be empty or occupied by an SV. The probabilities of initial occupancy for DS and RS are δ and ρ, respectively. The probability of SV release after one AP of an occupied DS is p_r. SV transitions within the DU are characterized by transition probabilities s_f, s_b, r_f, r_b.

For each series of experiments, control runs (usually 30 trials) were initially performed with stimulations of 8 APs at 100 Hz, interspersed by 10 s-long rest periods. We used previously developed procedures to determine from these data the parameters N, δ (the mean DS occupancy before the 1^st AP), ρ (the mean RS occupancy before the 1^st AP), as well as the probabilities s_f, r_f, and p_r, all of which were assumed constant throughout the AP train. In this part of the simulation, the back transitions s_b and r_b were neglected. The simulation yielded the numbers of released SVs (s_i) as a function of AP#. These numbers were calculated as s_i = N δ_i p_r where δ_i is the DS occupancy just before AP# i (Scheuss and Neher, 2001). Of note, there were some variations from one set of experiments to another regarding these parameters, reflecting differences between synapses (Table 1). Similar observations have been made at calyx of Held synapses (Lin et al., 2025).

DS and RS replenishment are not independent of each other. A DS can only be filled if an SV is present in the RS, and a downward transition from the RS can only take place if the DS is empty. Likewise, a transition from DS to RS will only happen if the DS is occupied and the RS empty. These considerations are taken into account in the Monte Carlo model. To better understand how the DUs as a whole changed after an AP and during trains, we considered four DU ‘states’ and we followed them throughout the simulations: DUs were “full” when both RS and DS were occupied; “up” when RS was occupied while DS was empty; “down” when RS was empty while DS was occupied; “empty” when both RS and DS were empty (Supplementary Fig. 3B). The initial δ and ρ values were used to estimate the initial DU states, keeping in mind the following relations: 1) δ = %down + %full, 2) ρ = %up +%full, and 3) %down + %up + %full + %empty = 1. The DU state was tracked alongside δ and ρ for each simulated AP.

The first AP was then simulated by releasing from the DS that were occupied (DU states full and down). Release resulted in an increase of % up and a simultaneous fall of % full (Supplementary Fig. 3B, grey shade).

The first 10 ms after an AP were simulated with short time intervals (50 μs) while following ρ, δ and the DU states. This timeframe only takes into account forward transitions (r_f and s_f), which were determined from fitting the 100 Hz control data of each dataset, as described before (Miki et al., 2016; Silva et al., 2024). This simulation was then used to establish the overall DU changes in the first 10 ms following an AP. For example, EE was defined as the probability that a DU that was empty just after release would be again empty 10 ms after the AP; EU the probability that a DU that was empty just after release would be up 10 ms after the AP, and so forth (Table 1). These probabilities were used to simulate the initial 10 ms of SV replenishment during an ISI for low frequency train simulations. The first 10 ms after release are dominated by the calcium-dependent movement of SVs from the RS to the DS (r_f, ‘transient docking’) which results in an increase in % down and a decrease in % up (see Supplementary Fig. 3B).

Next, to calculate DU state changes at times > 10 ms after the AP, time intervals for the calculations were changed from 50 μs to 10 ms. This change of time interval was justified because SV movements slow down markedly as a function of time since the AP. It resulted in a marked reduction in calculation time. Whereas transition probabilities were assumed to remain constant during the first 10 ms, they were assumed to change for each subsequent 10 ms time interval as specified in Supplementary Figure 3A. In this part of the simulation all transitions were considered, including the back transitions s_b and r_b. Note that the rate r_b has a delay before the start of the time-dependent exponential curve and that s_f remains constant throughout the ISI. Between 10 and 100 ms after the AP, up and down state proportions were stable, with a low % empty and up DUs and a high % down and full DUs (Supplementary Figure 3B; note time scale change at 10 ms). Next, SVs moved from DS to RS (undocking, with rate r_b), resulting in a gradual increase in % up and a decrease in % down (Fig. 3B). As described in the main text, this movement causes a temporary block of replenishment as ρ is high (RS gate closed). DUs gradually returned to their basal state on a time scale on the order of 10 s.

To model DU states during AP trains at low frequency, high time resolution modeling at times 0-10 ms after each AP (time domain 1) was followed by low time resolution modeling starting at 10 ms after the AP and ending at the end of each ISI (time domain 2). Models for consecutive APs were concatenated. For each binding site, the simulation started by attributing a site status to DS and RS. This was done at the beginning of each simulation trial by obtaining a pseudo-random value from the binomial distribution with a mean probability equal to the starting value of δ or ρ, for DS and RS respectively. DU states, δ and ρ were then followed step by step during the entire AP train.

Supplementary Fig. 3C depicts the changes in DU state during the simulations of high and low frequency 8-AP trains (Fig. 2 and Fig. 4). The proportion of DUs in each of the full/up/down/empty states is shown alongside ρ and δ just before each AP. During high frequency trains (left) an increase in % down (purple continuous curve) at the expense of a decrease in % up (green continuous curve) and % full (orange continuous curve) drives the rise in δ. During low frequency trains (center) the trend is the opposite, with undocking causing an increase in the % up (green), mirrored by a reduction in δ. This trend is mainly driven by a drop of % full (orange). DU changes during recovery trains (right) reproduce those obtained during control (left).

The kinetic parameters were adjusted by hand to find a good fit and to achieve a gradual return to the basal state for long ISIs. The simulation parameter values for the second domain were optimized according to the average release observed in the LFD trains as well as to the awakening train with the final δ and ρ values of the LFD train.

The states and occupancies were stored for each simulation trial and averages were obtained at the end of the simulation. For the simulations shown, calculations were done 10,000 times for each ISI and resulting averaged values are displayed.

Failure analysis

For this analysis we focused on the first component of LFD and more specifically, on the value of s₂. Results were taken from experiments with ISIs of 500 ms (Fig. 2), 800 ms and 1.6 s (Fig. 1). The trains were sorted according to whether the 1st AP failed or was successful (Fig. 5A; failures in purple, successes in black). The ratio of the released SV number after AP# 2 (s₂) and the average of the total released SV number after AP# 1 (<s₁>) was calculated for the s₁ failures and for the s₁ successes separately. One synapse had no failures; it was excluded from this analysis. Both traces with successes and failures show depression, as evidenced by a s₂/<s₁> (PPR) lower than 1, indicating that loss of SVs is not the main mechanism of LFD (Fig. 5B, left).

This analysis was extended by sorting the trains according to the number of SVs released for s₁. Though the data shows a trend of deeper depression with increased release during the 1st AP, the difference is not significant (Fig. 5B, right), indicating again that while release may play a role it is not the main mechanism. Next, we evaluated if the simulations replicated this characteristic of the data. The simulation trials were sorted the same way as the data and the PPRs calculated for each case (Fig. 5B; see black dots). The simulation also shows a significant depression for successful trials, with a somewhat stronger correlation between the amount of depression and the initial s₁ value than experimentally observed. Overall, the simulation reproduces well the results of the experiments.

Binomial analysis and predictions

According to the binomial distribution, the probability that s_i, the number of SVs released after AP number i, equals k is p(k) = P_i^k (1 − P_i)^N−k N!/(k!(N − k)!), where P_i is the release probability per docking site (P_i = δ_i p_r) and N is the number of docking sites (Malagon et al., 2016). The values of N and P were obtained by minimizing the summed squared deviations between the binomial prediction and data. This fitting was carried out for the control s₁ data of the synapses in Fig 7, giving the distribution p_ctrl(k). LFD was evaluated in a similar manner, considering each of the stimuli in the train as a separate event. The distributions of released SV numbers for 2 examples of control s₁ data is shown in black lines in Supplementary Fig. 5A and red lines show the LFD data distributions, p_LFD(k).

We tested the case of the depression in LFD being the result of an increase in stimulus failures, with a constant probability of release. In this case the probability of failures during LFD, p_LFD(0), would be larger than the control probability of failures, p_ctl(0), but the rest of the distribution would retain the same shape for the LFD and control distributions:

p_LFD(k) / p_ctl(k) = constant for k = 1 to N.

Considering that for each distribution, the sum of p(k) from 0 to N is 1, we obtain:

p_LFD(k) = p_ctl(k) x [1-( p_LFD(0)-p_ctl(0))/(1-p_ctl(0))] for k = 1 to N.

The dotted blue line in the examples of Supplementary Fig. 5A show the result of this prediction.

If, instead, LFD was the result of a decrease in probability of release then the p(k) of the LFD data would follow the distribution: p_LFD(k) = (P_LFD)^k x (1 - P_LFD)^N-k N!/(k!(N − k)!)

where N is the number of DS (assumed to be 5 in the example shown) and P_LFD is the P calculated from the LFD distribution: P_LFD = 1 - p_LFD(0)^1/N. The dotted purple lines in the examples of Supplementary Fig. 5A show the result of this prediction. The LFD data (red line) was then compared to both predictions. In both examples the decreased P prediction (purple) is much closer to the data than the prediction for increased stimulus failure (blue). To quantify this difference residuals were calculated for each point in the plot for both predictions. The group data for the residuals of each prediction is shown in Supplementary Fig. 5B, left. The right panel depicts the average of the residuals. The prediction of a decrease in P is much closer to the data, compared to the prediction for an increase in the number of stimulus failures.

This analysis shows that an increase in stimulus failures does not account for the distribution of events in the LFD trials and that a decrease in P explains this phenomenon better.

Supplementary Material

Supplementary Figure 1:

Supplementary Figure 2:

Supplementary Figure 3:

Supplementary Figure 4:

Supplementary Figure 5:

Supplementary Figure 6:

Data availability

The Data that supports the finding of this study will be made available in Dryad.

Acknowledgements

We thank Takafumi Miki for his help in the modelling. This work was supported by CNRS (UMR 8003), by the European Research Council (Advanced Grant SinSyn to A. M.), and by Agence Nationale pour la Recherche (Grant PLASTICAZ to A. M.).

Additional information

Funding

Agence Nationale de la Recherche (ANR) (ANR22 CE91 0013 01)

• Alain Marty