Plasticity in Single Axon Glutamatergic Connection to GABAergic Interneurons Regulates Complex Events in the Human Neocortex

In the human neocortex, single excitatory pyramidal cells can elicit very large glutamatergic EPSPs (VLEs) in inhibitory GABAergic interneurons capable of triggering their firing with short (3–5 ms) delay. Similar strong excitatory connections between two individual neurons have not been found in nonhuman cortices, suggesting that these synapses are specific to human interneurons. The VLEs are crucial for generating neocortical complex events, observed as single pyramidal cell spike-evoked discharge of cell assemblies in the frontal and temporal cortices. However, long-term plasticity of the VLE connections and how the plasticity modulates neocortical complex events has not been studied. Using triple and dual whole-cell recordings from synaptically connected human neocortical layers 2–3 neurons, we show that VLEs in fast-spiking GABAergic interneurons exhibit robust activity-induced long-term depression (LTD). The LTD by single pyramidal cell 40 Hz spike bursts is specific to connections with VLEs, requires group I metabotropic glutamate receptors, and has a presynaptic mechanism. The LTD of VLE connections alters suprathreshold activation of interneurons in the complex events suppressing the discharge of fast-spiking GABAergic cells. The VLEs triggering the complex events may contribute to cognitive processes in the human neocortex, and their long-term plasticity can alter the discharging cortical cell assemblies by learning.


Introduction
Evolution has shaped the human neocortex producing microcircuit features that are specific to our species [1]. Neuronal density, ultrastructural features, and functional properties of neurons [2][3][4] reflect specific adaptations in the human neocortex to perform complex and fast signal processing [5][6][7][8][9][10][11][12][13]. A remarkable feature in the human neocortex is that single pyramidal cell (PC) action potentials (APs) are able to generate di-and polysynaptic GABAergic interneuron discharge known as complex events [10,11]. The events emerge from the activity of a small subset of excitatory connections forming very large glutamatergic excitatory postsynaptic potentials (VLEs), specifically to GABAergic interneurons in supragranular layers of the frontal, the temporal, and the prefrontal cortices [10,11]. Similar strong excitatory connections between individual neocortical neurons have not been found in nonhuman brains. Therefore, it has been proposed that the VLEs and the complex events participate in cortical information encoding in high order cognitive processes [10]. However, this would predict that these events are dynamically modulated by learning [14,15]. Yet, it is unknown whether the VLEs show usedependent long-term plasticity, and if their specific modulation indeed affects the complex events. We hypothesize that the immense strength of VLEs is generated and regulated by common activity-driven synaptic long-term plasticity processes, and that they may occur in various different inhibitory interneuron types [16,17]. Alternatively, these connections could be hardwired selectively in a specific, yet unknown, subset of postsynaptic GABAergic interneurons without prominent lasting plasticity in the adult neocortex [18,19].
We asked whether VLEs show activity-induced long-term plasticity, and if their selective modulation had impact on the local complex events. By performing triple and dual whole-cell recordings of synaptically connected identified neurons, we found that VLEs exhibit metabotropic glutamate receptor (mGluR)-dependent long-term depression (LTD) that converts them to common weak excitatory postsynaptic potential (EPSP) connections. In addition, this alters the neocortical complex events suppressing the cell assemblies activated by the PC. Thus, the VLEs occur in various interneuron types, and their occurrence is regulated by the synapse's activity history. To the best of our knowledge, this is the first study reporting synaptic plasticity in human neocortical interneurons.
with nonparametric distribution (failures excluded, evoked with 10 s interval at Em −68.5 ± 1.4 mV, n = 21, Shapiro-Wilk test). Despite the amplitude difference, the excitatory postsynaptic currents (EPSCs) in FSINs similarly exhibited fast time-to-peak kinetics (0.59 ± 0.04 ms, n = 18) (Fig 1B, S1 Data), indicating that the amplitude variability is unlikely to derive from different electrotonic filtering of the glutamatergic synaptic inputs. Likewise, distribution of the average EPSP amplitudes in PCs to non-fast-spiking interneuron (non-FSIN) pairs showed nonparametric distribution with a range from 0.7 mV to 6.9 mV (failures excluded, at Em −70.3 ± 1.5 mV, n = 9, Shapiro-Wilk test) (S1 Table). Thus, VLEs are not occurring solely in FSINs, but are exhibited in various types of GABAergic neurons including fast-and non-fastspiking cells. On the contrary, PC-PC connections showed parametric distribution of average EPSPs with small amplitude (2.01 ± 0.02 mV at Em −69.4 ± 1.8 mV, failures excluded, Shapiro-Wilk test, n = 16) (Fig 1B, S1 Table, S1 Data). The interneuron EPSPs were defined as VLEs when their average (failures excluded) was larger than mean + 2 x standard deviation (SD) of the EPSPs in PC-PC connections (4.21 mV, failures excluded) in baseline conditions (mean ± SD = 2.01 ± 1.10 mV, n = 480 in 16 cells). The postsynaptic interneurons in the triple and paired recordings were immunohistochemically confirmed positive for vesicular GABA transporter (vgat+) (n = 31). Cells that in addition were immunopositive for parvalbumin (pv +) showed rapid axon currents (spike inward current width [SW] of 0.43 ± 0.02 ms, n = 11) characteristic of the FSINs [10,11]. The pv+ cells, together with vgat+ interneurons showing similar fast spike kinetics (SW 0.49 ± 0.02 ms, n = 11), but with nonconclusive or untested pv reaction, were considered FSINs (n = 22). Ten FSINs were further identified as putative basket cells by their axon morphology [10]. The non-fast-spiking vgat+ interneurons and the PCs had significantly longer spike kinetics with SW of 0.96 ± 0.04 ms (n = 9) and 1.18 ± 0.06 ms (n = 16), respectively (p < 0.01 between all groups, ANOVA with Tukey's posthoc test) [13]. The non-FSINs with intact soma were immunohistochemically tested for somatostatin (sst) for further identification of the cells [20]. Detailed results on the EPSPs excluding and including failures, the EPSCs, and the immunohistochemical reaction analyses with cell type identification are shown in S1 Table. In all potential connections tested between neurons (n = 1,056), we found (including connections lost during baseline) a monosynaptic response in 11.0% of cases. Success rate for identified PC-FSIN pairs was 4.0%, PC-non-FSINs connections: 1.9%, PC-PC pairs: 1.1%, and FSIN-PC connections: 3.8%, showing similar or slightly lower connectivity rates than reported in the rodent neocortex L2-3 [21][22][23].

VLEs in FSINs Are Subject to Group I mGluR-Dependent LTD
We asked whether glutamatergic connections to interneurons showed long-term plasticity akin to that reported in the rodent cortex [24,25]. To test this, we performed experiments applying high frequency bursts of APs (5 APs at 40 Hz, x 40 with 0.5 s interval) in the presynaptic PC after a baseline of EPSPs (at least 5 min, but less than 10 min, analyzed including failures) [26,27]. Postsynaptic FSINs (SW 0.40 ± 0.02 ms) were held in resting membrane potential in current clamp (-67.6 ± 2.2 mV, n = 9) (Fig 2). First, we tested VLEs (average in baseline 5.85 ± 0.59 mV, did not show failures, n = 5) in control conditions and found that the afferent PC bursts firing generated an LTD of the EPSPs (amplitude to 0.52 ± 0.02 of baseline at 20-25 min after 40 Hz bursts, n = 5 cells, p < 0.01, Wilcoxon test) (Fig 2A, S1 Fig, S2 Data, S7 Data). The LTD was associated with a reduced paired-pulse EPSP ratio (1 st /2 nd EPSP amplitude with 50 ms interval) to 0.74 ± 0.06 of baseline (p < 0.05, Mann-Whitney test, baseline mean 1.41 ± 0.22) and a decrease in the EPSP amplitude CV −2 (1/squared coefficient of variation) value (to 0.44 ± 0.12% from baseline, p < 0.05, Mann-Whitney test) (S3 Data), indicating presynaptic site of depression [28]. The results on the paired-pulse ratio (PPR) and altered EPSP amplitude variation by LTD are summarized in histograms in Fig 2B. Likewise, corresponding experiments in occasional PC-non-FSIN pairs with VLEs (n = 2, averages in the baseline including failures 5.81 mV and 6.89 mV, failure rates 0% and 13%, respectively) showed that single fiber burst firing can also generate LTD in some non-FSINs (S2 Fig).

LTD by Single PC Bursting Is Specific to VLEs
In contrast to VLEs, the PC-FSIN pairs with small EPSPs (average in baseline with failures 1.89 ± 0.43 mV at Em −69.2 ± 3.5 mV, failure rate 11.2 ± 9.5%, SW 0.48 ± 0.04 ms, n = 5) failed to show lasting plasticity following the 40 Hz bursts (amplitude 1.07 ± 0.06 of baseline at 20-25 min) (Fig 3A, S1 Fig, S4 Data). Likewise, no lasting plasticity was seen in any (paired t-test) of the three vgat+ non-FSINs (SW 0.91 ± 0.07, n = 3) with small amplitude EPSP (2.08 ± 0.58 mV, n = 3) in similar experiments. Given that LTD in the VLEs was accompanied by stronger postsynaptic depolarization during the presynaptic spike bursts, we studied whether small amplitude EPSP connections would show plasticity if the postsynaptic FSIN was depolarized during the PC bursts. We reproduced experiments above with a separate set of PC-FSIN (SW 0.47 ± 0.03 ms, n = 5) pairs with small amplitude EPSP (average with failures 1.44 ± 0.22 mV, failure rate 11.0 ± 3.3%, n = 5), and paired presynaptic PC spike bursts with postsynaptic cell depolarization (20-30 mV, 250 ms steps from Em, see Methods for details) beyond the firing threshold ( Fig 3A). This protocol also failed to generate long-lasting change in EPSPs in the PC-FSIN pairs (0.94 ± 0.04 of baseline at 20-25 min, n = 5, Wilcoxon test) (Fig 3A, S4 Data). Interestingly, a small but significant LTD was observed with this configuration in two (0.71 ± 0.10 and 0.83 ± 0.05 at 20-25 min compared to baseline, p < 0.05 for both cells, paired t-test) of three individual PC-non-FSIN (vgat+) pairs tested. Finally, we studied the synaptic connections between L2-3 PCs applying presynaptic 40 Hz bursts while the postsynaptic cell was at resting membrane potential. The PC-PC pairs were connected with small amplitude EPSPs (average with failures 1.40 ± 0.30 mV at Em −65.9 ± 5.4 mV, failure rate 4.2 ± 3.2%, n = 4), and the 40 Hz bursts failed to generate lasting plasticity in the EPSP (1.01 ± 0.05 of baseline at 20-25 min, n = 4 cells, Wilcoxon test) (Fig 3B, S4 Data), possibly because LTP and LTD in human PCs require strong postsynaptic depolarization for either activation of glutamate NMDA receptors or L-type voltage-gated calcium channels [7]. Input resistance in the plasticity recordings showed small increase to 1.09 ± 0.01 (baseline-normalized) at 20-25 min from baseline (n = 26, p < 0.01, t-test) [33]. The results show that following just a single PC burst firing, LTD specifically occurs in large EPSPs between PCs and interneurons, and not in other investigated synaptic connections.

Co-activation of Many Weak Glutamatergic Synapses Evokes mGluR-Dependent LTD in Fast-Spiking Interneurons
Because studies in rodents have reported LTD in glutamatergic synapses to cortical interneurons either by chemical or strong synaptic activation of group I mGluRs, we studied whether FSINs with weak excitatory inputs showed the LTD when multiple glutamatergic fibers were simultaneously activated ( Fig 3C) [27,29,34]. Evoking compound EPSCs from many small glutamatergic inputs with extracellular electrical stimulation (see Methods), we applied 40 Hz bursts to the glutamatergic pathway as above after baseline (at least 5 min, but less than 10 min). Focusing the EPSC analysis in the FSINs on the monosynaptic component of the current (see Fig 3C) [35,36] we found that the 40 Hz burst stimulation resulted in LTD (EPSC 0.72 ± 0.02 from baseline at 20 min, n = 7 cells, p < 0.001, t-test). Blockers for glutamate Nmethyl-D-aspartatereceptor (NMDARs) (DL-2-Amino-5-phosphonopentanoic acid [DL-APV], 100 μM) and GABA A Rs (PiTX, 100 μM) were present in the experiments. The EPSC amplitude LTD was accompanied by decreased CV −2 (baseline-normalized to 0.72 ± 0.16%, n = 7, t-test). This LTD was blocked in experiments with group I mGluR antagonists LY367385 (100 μM) and MPEP (25 μM) (n = 7, p < 0.05, t-test) (Fig 3C, S5 Data). The FSINs in these extracellular stimulation experiments showed narrow SW 0.62 ± 0.04 ms, n = 14. Accordingly, we reproduced these experiments with rat glutamatergic fibers to FSINs in L2-3 (SW 0.64 ± 0.06 ms, n = 10) and confirmed LTD (EPSC 0.64 ± 0.04 from baseline at >15 min, n = 5, p < 0.01, Wilcoxon test) and its blockade with the group I mGluR antagonists (EPSC from baseline 0.95 ± 0.5, n = 5, Wilcoxon test) ( Fig 3C, S5 Data). The EPSC amplitude in LTD showed reduced CV −2 (baseline-normalized to 0.41 ± 0.09%, n = 5, Mann-Whitney test), but not when LTD was blocked with the mGluR antagonists (1.04 ± 0.24% of baseline at >15 min) (p < 0.05 between the groups at >15 min, Mann-Whitney test) (Fig 3C, S5 Data). We tested three of the recorded rat FSINs for pv immunoreaction and found them all positive. Thus, in both human and rat cortex, weak glutamatergic connections to L2-3 FSINs exhibit group I mGluR-dependent LTD if multiple glutamatergic inputs are activated simultaneously.

LTD of VLEs Is Sufficient to Modify Network Activity
Given that interneuron-PC connections with VLEs have been proposed to be essential in generation of the neocortical complex events, we studied whether the LTD in these connections would selectively modify network activity. First, we confirmed that single PC AP-evoked VLEs in the FSIN as well as in the non-FSIN elicited firing of these interneurons from the resting membrane potential [11]. We found that VLE-evoked postsynaptic spikes in a FSIN (Em −69 mV) typically followed with a short 3-5 ms delay ( Fig 4A)   recordings between identified PCs revealed disynaptic GABA A R-mediated inhibitory currents (dIPSCs) in complex events elicited by a single AP (interval 10 s) (Fig 4B) [10,11]. The dIPSCs occurred with short delay (6.23 ± 0.72 ms, n = 16 pairs) and high probability (0.70 ± 0.05, n = 16 pairs in baseline conditions) (S6 Data). The dIPSCs showed longer and more variable delay to the presynaptic spike than monosynaptic GABA A R-mediated inhibitory currents (monIPSCs) from FSINs (0.96 ± 0.10 ms, n = 9, p < 0.001, t-test) (SW 0.48 ± 0.03 ms, n = 9) (S3 Fig, S9 Data). In addition, dIPSCs were blocked by the glutamate AMPAR blocker GYKI53655 (25 μM) (n = 3, p < 0.001, Chi-square test) (Fig 4C, S4 Fig, S6 Data). The evoked dIPSC amplitudes (averages excluding failures in all plasticity recordings 32.1 ± 3.7 pA, n = 12) were similar to monIPSCs (35.2 ± 5.6 pA, n = 9, t-test) (S5 Fig, S10 Data), indicating that these early complex event inhibitory currents (IPSCs) were generated by a single FSIN. The dIPSCs and the monIPSCs were recorded at −55 mV. The dIPSCs were detected in 3.0% of all potential connections tested (n = 1,056).
To determine whether plasticity modified this network, we performed long recordings from PC pairs showing that the probability and delay of the dIPSCs were stable for at least 30 min (n = 3) in normal conditions (Fig 4D, 4E and S6 Data). However, if the 40 Hz burst firing was delivered in the presynaptic PC (similar to the VLE LTD experiments in Fig 2), the dIPSC occurrence (total in 25 ms from PC spike) rapidly and permanently attenuated after a baseline showing strong LTD (from 0.71 ± 0.04 in baseline to 0.14 ± 0.09 at 15-20 min, n = 3, p < 0.05, Chi-square test) (Fig 4F, 4G and S6 Data). Interleaved experiments in the presence of group I mGluR antagonists LY367385 (100 μM) and MPEP (25 μM) showed that LTD of dIPSCs was blocked in the presence of the group I mGluR antagonists (0.81 ± 0.10 in baseline and 0.88 ± 0.12 at 15-20 min, n = 3, Chi-square test) (Fig 4H, 4I and S6 Data). In conclusion, the results demonstrate that a single PC burst firing at 40 Hz elicits robust LTD in VLE connections to FSINs and causes suppression of phase-locked early dIPSCs between PCs. These two LTDs both require group I mGluRs. Thus, the results show that plasticity of VLEs in FSINs changes the activation pattern of the neurons discharging in supragranular layers during complex events.

Discussion
Almost a decade ago, Molnar et al. [10] first reported that a subset of excitatory PC synapses in the human neocortex form VLEs in local GABAergic interneurons in supragranular layers. These strong connections specifically from PCs to inhibitory interneurons have since been reported in the frontal, prefrontal, parietal, and temporal cortices, where the VLEs often are suprathreshold, driving assemblies of inhibitory interneurons to fire after a single PC spike [11,37]. Similar strong connections between two individual neocortical neurons have not been found in nonhuman species [37][38][39]. Analyses of large datasets from rodent visual and somatosensory cortices have revealed that neurons in local networks are not randomly connected, but specific local connectivity patterns exist between neuron types, and the strongest excitatory synapses control local network activity [40][41][42]. This suggests that there is a skeleton of strong connections in the network that dominates the activity [40]. Therefore, it has been proposed that the VLE connections in the human might be important in generating neocortical cell assemblies and be involved in higher cognitive functions [11]. However, until now it had remained unknown how neuronal plasticity regulates these connections and whether their selective modulation indeed alters discharge of neuronal assemblies in the human neocortical network activity.
Our result that the VLEs occur specifically in human interneurons, and not in between PCs, is consistent with previous studies [10,11]. In addition, we demonstrate that VLEs are generated in different interneuron subpopulations, including fast-spiking (such as basket cells) and non-fast-spiking supragranular vgat+ neurons. Furthermore, single GABAergic interneurons receive both VLEs and more common small EPSP connections from layer 2-3 PCs. Indeed, anatomically identified L2-3 basket cells show huge variability in strength of PC inputs. It is likely, although we do not directly demonstrate it here, that a single presynaptic L2-3 PC cell evokes VLEs and small EPSPs in different L2-3 postsynaptic interneurons showing synapse specificity. A comprehensive recent study in rodent demonstrated that certain connectivity patterns between neurons are repeated in the neocortex across different regions [39]. The strong VLE connection, occurring between PCs and L2-3 inhibitory interneurons in various neocortical regions is one such feature, and probably specific to the human neocortical microcircuits.
We show that strength of VLE connections is controlled by their activity history. A common form of mGluR-dependent LTD characterized in the extracellular stimulation experiments of glutamatergic synapses in the rodent cortex [24] converts the VLEs to small amplitude EPSPs. This LTD suppresses VLE synapses through a presynaptic mechanism most likely controlling the vesicular transmitter release as indicated by the changes in the EPSP amplitude PPR and the coefficient of variation in the depressed synapses [37]. The LTD of VLEs in FSINs, many of which are also pv+, involves the group I mGluRs. In the rodent cortex, group I mGluRs have been shown to play a central role in long-term plasticity processes including presynaptic LTD and LTP [24,[29][30][31]34,[43][44][45][46]. Our results from the group I mGluR-dependent LTD in FSINs in the human and the rodent neocortex indicate that this is an evolutionarily conserved plasticity mechanism for controlling the fast-spiking interneuron activity in the mammalian brain.
Interestingly, we report that in a non-fast-spiking cell, LTD was not blocked by the group I mGluR antagonists (see S2 Fig). In addition, we found significant LTD in two of three PCnon-FSIN connections with small EPSP when the postsynaptic cell was depolarized during PC bursts. These suggest that there may be various LTD forms in human cortical interneuron synapses [24,25], and that non-FSINs may exhibit different LTD mechanism than the FSINs, possibly depending on postsynaptic depolarization. Indeed, many synapse-specific properties including long-term plasticity have been reported in glutamatergic fibers in the rodent cortex [21,32,[47][48][49][50]. Correspondingly, synapses originating from the same PC in the human neocortex may exhibit distinct long-term plasticity, depending upon the postsynaptic target cell; and different activity patterns may be required for plasticity in the synapses [24,50]. In this study, we have used elevated extracellular calcium (3 mM) to increase the stability of disynaptic IPSCs in baseline conditions. However, compared to recordings with 2 mM extracellular calcium, this modification is unlikely to strongly affect the high-frequency firing-evoked synaptic release and the long-term plasticity in FSIN synapses: the probability of synaptic release in VLEs in FSINs is already very high at 2 mM Ca 2+ , and only slightly modulated by further increase of calcium [37]. Yet, as demonstrated in the recent study by Molnar et al. [37], the VLE release probability can markedly decrease when extracellular calcium is reduced from 3 mM to 1.5 mM, which is considered lower range of the cerebrospinal fluid total calcium level in physiological conditions [51]. Therefore, it is also possible that in calcium concentrations close to 1.5 mM, the PC firing pattern used in this study may not be sufficient for such a robust LTD as reported here in the FSINs. Importantly, some non-FSINs show VLEs with low presynaptic release probability even in 3 mM calcium, as indicated by the large EPSP amplitude coefficient of variation. In these synapses, extracellular calcium modulations may have even stronger effects on the short-and long-term plasticities than in the FSINs.
Strikingly, in the human neocortex, the activity of a single PC is sufficient to trigger mGluR-dependent LTD in the VLE connections, but not in weak glutamatergic pathways. The level of the postsynaptic depolarization does not explain the failure of LTD in small EPSP connections to fast-spiking cells, and a potential explanation is that there is insufficient activation of the postsynaptic group I mGluRs [34,43]. The connections with VLEs are likely to release more glutamate and activate the critical mGluRs [52][53][54]. The hypothesis on strong glutamate release is supported by our finding that the small EPSP connections were unable to generate LTD by single fiber activity, but they showed the mGluR-dependent plasticity when multiple fibers were activated simultaneously with local extracellular stimulation [55]. Indeed, a recent study revealed that human neocortical PC-FSIN synapses with VLEs have more transmitter vesicle release sites, although the glutamate release quantal size is similar compared to synapses in rat neocortex [37]. This indicates multivesicular release in synapses with VLEs, and it is interesting to speculate that the conversion to common small EPSPs via the presynaptic LTD might reflect their transformation from a multivesicular release site to a single vesicle-releasing synapse.
Although a link between very large excitatory synapses and human cortical complex events has been suggested earlier [10,11], a relation between their selective modulation and complex events had not been directly demonstrated until this study. The LTD triggered by a single PC firing in the current conditions is specific to connections with VLEs and therefore provides a useful tool to test the relation between VLEs and neocortical network activity. Results here show that the VLE connections indeed trigger the complex events, and the LTD changes their temporal structure. The activation of individual fast-spiking interneurons was commonly observed in dual PC recording as disynaptic GABAergic currents with timing corresponding to the APs in fast spiking interneurons [11]. LTD of the VLE in FSINs and the suppression of dIPSCs in complex events were both induced by single PC burst firing, and they both required group I mGluRs. Some non-FSINs also exhibit VLEs and can show LTD, although these cells were unlikely to contribute to the GABAergic disynaptic currents investigated here: the dIPSCs occurred with short delay and high precision, whereas the non-FSINs show long and variable delay in their response to fire APs (see S2 Fig). However, the results of non-FSINs are based on small sample sizes, and should therefore be interpreted with caution.
In conclusion, the human neocortex is unique in many aspects, since its microcircuits show differences at the molecular, ultrastructural, and physiological levels compared to other mammalian species [1,3,56]. The capacity of the human neocortex to perform extraordinary and highly complex tasks may at least partly result from these microcircuit level specializations. We propose that VLEs with robust activity-induced plasticity and their contribution to neocortical cell assemblies may be crucial for higher cognitive functions and abstract mental abilities of the human brain. In addition, evidence in animal models suggests the involvement of group I mGluR-mediated plasticity in neocortical learning processes, and perturbation of the mGluRdependent plasticity has been reported with mental decline [43]. Therefore, the human-specific microcircuit features may also be substrates for pathological processes resulting in cognitive decline and other neurological and neuropsychiatric dysfunctions that we as a species are vulnerable to [57][58][59][60][61].

Ethics Statement
All procedures were performed according to the Declaration of Helsinki with the approval of the University of Szeged Ethical Committee and Regional Human Investigation Review Board (ref. 75/2014).

Electrophysiology and Analysis
Human neocortical slices were derived from material that had to be removed to gain access to the surgical treatment of deep-brain tumors from the left and right frontal, temporal, and parietal regions with written informed consent of the patients prior to surgery. The patients were 10-85 y of age (mean ± SD = 50 ± 4 y), including 17 males and 14 females. The tissue obtained from underage patients was provided with agreement from a parent or guardian. The resected samples were cut from the frontal and temporal lobes of left or right hemisphere. Anesthesia was induced with intravenous midazolam and fentanyl (0.03 mg/kg, 1-2 lg/kg, respectively). A bolus dose of propofol (1-2 mg/kg) was administered intravenously. The patients received 0.5 mg/kg rocuronium to facilitate endotracheal intubation. After 2 min, the trachea was intubated and the patient was ventilated with O 2 /N 2 O mixture (a ratio of 1:2). Anesthesia was maintained with sevoflurane at monitored anesthesia care volume of 1.2-1.5. After surgical removal, the resected tissue blocks were immediately immersed in ice-cold standard solution containing (in mM): 130 NaCl, 3.5 KCl, 1 NaH 2 PO 4 , 24 NaHCO 3 , 1 CaCl 2 , 3 MgSO 4 , 10 D(+)-glucose, and saturated with 95% O 2 and 5% CO 2 . Slices were cut perpendicular to cortical layers at a thickness of 350 μm with a microtome (Microm HM 650 V) and were incubated at room temperature (20-24°C) for 1 h in the same solution. Rat neocortical slices were prepared as described before [62]. Male Wistar rats were anaesthetized using halothane, and following decapitation (320 μm thick), coronal slices were prepared from the somatosensory cortex. The solution used during electrophysiology experiments was identical to the slicing solution, except it contained 3 mM CaCl 2 and 1.5 mM MgSO 4 . Recordings were performed in a submerged chamber (perfused 8 ml/min) at approximately 36-37°C. Cells were patched using water-immersion 20× objective with additional zoom (up to 4x) and infrared differential interference contrast video microscopy. Micropipettes (5)(6)(7)(8) were filled with intracellular solution for whole-cell patch-clamp recording (in mM): 126 K-gluconate, 8 KCl, 4 ATP-Mg, 0.3 Na 2 -GTP, 10 HEPES, 10 phosphocreatine (pH 7.20; 300 mOsm) with 0.3% (w/v) biocytin. Current and voltage clamp recordings were performed with Mutliclamp 2B amplifier (Axon Instruments), lowpass filtered at 6 kHz (Bessel filter). Series resistance (Rs) and pipette capacitance were compensated in current clamp mode and pipette capacitance in voltage clamp mode. Cell capacitance compensation was not applied. Rs was monitored and recorded continuously during the experiments. The recording in voltage clamp mode was discarded if the Rs was higher than 25 OM or changed more than 20%.
In paired cell recordings, APs were generated in the presynaptic cell with brief (2-3 ms) suprathreshold depolarizing (60-70 mV) paired pulses (50 ms interval) in voltage clamp delivered every 10 s from −60 mV. Postsynaptic cells were at resting membrane potential in current clamp mode. In some cells with VLEs, the postsynaptic cell was hyperpolarized (up to −10 mV) with constant current to prevent the VLE from triggering an AP. The 40 Hz firing protocol was similarly applied in voltage clamp mode with series of 2-3 ms depolarizing pulses (5 pulses at 40 Hz, delivered every 0.5 sec 40 times), while the postsynaptic cell was held in current clamp resting membrane potential. In some experiments (Fig 3A, 4-5), the postsynaptic cell was depolarized in voltage clamp during presynaptic 40 Hz firing with a continuous step (20-30 mV, 250 ms). This elicited on average 2.2 postsynaptic spikes for 1 st presynaptic spike (in following 25 ms, n = 200 in 5 cells) of the 40 Hz train, and on average 0.80 postsynaptic spike probability for the 2 nd -5 th PC AP.
Extracellular stimulation was applied with a concentric bipolar electrode (125 μm tip diameter, FHC Inc., US) positioned on L2-3. Paired pulse stimuli (50 μs, with 50 ms interval, intensity range from 20 to 300 μA) were delivered every 15 s with current isolator stimulator (Model DS3, Digitimer, UK). Compound EPSCs in Fig 3 were confirmed by observing less than 100 pA increases in the evoked EPSC amplitude when gradually increasing stimulation intensity.

Data Analysis and Statistics
Data were acquired with Clampex software (Axon Instruments, US) at 20 kHz. EPSC/P, IPSC, action current duration, and the cell input resistance were analyzed off-line with p-Clamp software (Axon Instruments, US) and Spike2 (version 7.0, Cambridge Electronic Design, UK). Liquid junction potential was not corrected. EPSC amplitude and kinetics analysis in voltage clamp mode (time-to-peak from onset) was omitted when access resistance was higher than 25 MO. SW was calculated from the onset of inward action current till recovery to baseline holding level. Data for SWs were collected in the beginning of experiments when synaptic connections from all cells were briefly tested in voltage clamp mode. All data are presented as mean ± s.e.m. and when showing baseline-normalized EPSPs of many cells, the values were calculated from binned (30 s bin) data. In rare cases when cell-spiked and accurate EPSP amplitude data was not available, bin includes two instead of three data points. For statistical analysis, ANOVA with posthoc Tukey's test and t-test were used for data with normal distribution (Shapiro-Wilk test) and sample sizes larger than n = 6. Chi-square test was used for categorical variables (occurrence of dIPSC in 25 ms time window from PC spike). Otherwise, Mann-Whitney U-test (unpaired) and Wilcoxon Signed Rank Test (paired) were used. EPSP amplitude in individual plasticity experiments was tested with paired t-test comparing data points in 5 min baseline and an equal time window at 20-25 min following the presynaptic bursts, unless stated otherwise (in some shorter experiments in the last 5 min of the recording). Correlation was determined with Pearson's r-test. Differences were accepted as significant if p < 0.05. Failures were included in the EPSP mean values (in binned data) in plasticity analysis.

Cell Visualization and Image Reconstruction
After electrophysiological recording, slices were immediately fixed in a fixative containing 4% paraformaldehyde and 15% picric acid in 0.1 M phosphate buffer (PB; pH = 7.4) at 4°C for at least 12 h, then stored in 0.1 M PB with 0.05% sodium azide as a preservative at 4°C. Slices were embedded in 10% gelatin and further sectioned into slices of 50 μm thickness in cold PB using a vibratome VT1000S (Leica Microsystems, UK). After sectioning, the slices were rinsed in 0.1 M PB (3 x 10 min) and cryoprotected: first step in 10% (30 min), and later in 20% sucrose (1 h) dissolved in PB and then permeabilized using a freeze and thaw procedure. Some sections for cell structure illustrations were further incubated in a solution of conjugated avidin-biotin horseradish peroxidase (ABC; 1:300; Vector Labs, UK) in Tris-buffered saline (TBS, pH = 7.4) at 4°C overnight. The enzyme reaction was revealed by the glucose oxidase-DAB-nickel method using 3'3-diaminobenzidine tetrahydrochloride (0.05%) as chromogen and 0.01% H 2 O 2 as oxidant. Sections were postfixed with 1% OsO 4 in 0.1M PB. After several washes in distilled water, sections were stained in 1% uranyl acetate and dehydrated in ascending series of ethanol. Sections were infiltrated with epoxy resin (Durcupan) overnight and embedded on glass slices. Three-dimensional light microscopic reconstructions from sections were carried out using the Neurolucida system with 100 x objective (Olympus BX51, Olympus UPlanFI, Hungary). Images were collapsed in z-axis for illustration. Cells in Fig 1 were reconstructed from confocal microscope z-stack images of streptavidin fluorophore signal using Image-J software as described previously [32]. Vgat immunoreaction analysis was used in parallel to confirm the interneuron axon.

Immunohistochemistry
For immunohistological reactions, free-floating sections were washed 3 times in TBS-TX 0.3% (15 min) at 22-24°C, then moved in 20% blocking solution with horse serum in TBS-TX 0.3%. The sections were incubated in primary antibodies diluted in 1% serum in TBS-TX 0.3% over three nights at 4°C, then put in relevant fluorochrome-conjugated secondary antibodies in 1% of blocking serum in TBS-TX 0.3% overnight at 4°C. Sections were washed at first step in TBS-TX 0.3% (3 x 20 min) and later in 0.1 M PB (3 x 20 min) and mounted on glass slides with Vectashield mounting medium (Vector Lab.Inc., UK). The characterizations of antibodies: pv (goat anti-pv, 1:500, Swant, Switzerland, www.swant.com), sst (rat anti-sst, 1:50, Merck Millipore, Germany, www.merckmillipore.com) and vgat (rabbit anti-vgat, 1:500, Synaptic Systems, Germany, www.sysy.com). Fluorophore-labelled secondary antibodies were: DyLight 488 (Donkey anti goat, 1:400, Jackson ImmunoResearch Lab. Inc., www.jacksonimmuno.com, US), Alexa488 (Donkey anti rat, 1:400, Jackson ImmunoResearch Lab. Inc.) and Cy5 (Donkey anti rabbit, 1:500, Jackson ImmunoResearch Lab. Inc.). Labelling of neurons by neurobiotin and immunoreactions were evaluated using first epifluorescence (Leica DM 5000 B, UK) and then laser scanning confocal microscopy (Olympus FV1000, Hungary). Immunoreaction was considered to be negative when fluorescence was not detected in relevant neurobiotin-labelled cell, but immunopositivity was detected in the same area in unlabelled cells. Immunoreactions were studied in axon boutons (vgat and pv) and soma and dendrites (pv, sst).  (t -test). The plot shows in left the average amplitudes of monosynaptic IPSCs (monIPSCs) elicited in 9 FSIN-PC pairs (failures excluded) as reported in the results. Average amplitudes of the disynaptic IPSCs (dIPSCs, failures excluded) shown in right. This suggests the earliest complex event IPSCs with 6.23 ± 0.72 ms delay are generated by single fast-spiking interneurons. Plot shows mean ± s.e.m. of the amplitudes in baseline conditions (S10 Data). (TIF) S1 Table. Results on EPSP/EPSC analyses and immunohistochemical reactions of the three postsynaptic cell populations: fast-spiking interneurons (FSINs), non-fast-spiking interneurons (nonFSINs) and pyramidal cells (PCs). The table comprises EPSP data in baseline conditions for 31 postsynaptic vgat+ cells studied for long-term plasticity and 16 PCs in the triple-and paired recordings. Experiment code identifies the postsynaptic cell recorded. Ten fast-spiking putative basket cells were anatomically identified. EPSP average amplitude is shown without synaptic failures and including the failures. Experiments are listed in a descending order starting from cells with the largest average EPSPs. EPSP failure rate is a percentage of failures of the EPSP to the presynaptic spike. EPSP rise time (from 20% to 80% of maximum) was measured from averaged EPSPs. Paired-pulse ratio (PPR) is defined as 1 st EPSP / 2 nd EPSP amplitude. EPSCs represent an average of 5 responses recorded in first 2 minute period in the voltage-clamp after break-in to the whole cell mode from the giga-seal. Because the EPSCs were recorded immediately after break-in (when whole-cell properties are typically not yet stabilized), the measured values may represent underestimate of the current. Note that in 3 FSINs and in one PC, EPSCs were not measured. Immunohistochemical reactions for vesicular GABA transporter (vgat), parvalbumin (pv) and somatostatin (sst) are indicated by (+) for positive and (-) for negative. The vgat was analyzed in axon terminals, pv in axon terminals, soma or dendrites, and sst in soma and dendrites. Putative basket cells in the FSINs were identified by their axon forming collaterals around L2-3 neuron somatae. No axoaxonic cell types were identified. Pyramidal cells showed characteristic densely spiny dendrites with filopodial, stubby and mushroom-shaped spines, and main axon projecting towards infragranular layers. In the PCs table "a" and "b" indicate distinct postsynaptic cells in a mutually connected PC couple.