Animals

We used in-house-bred Wistar (14–19 days old; Ecole Polytechnique Fédérale de Lausanne) and Sprague–Dawley (4–6 weeks old; Center for Psychiatric Neuroscience) rats of both sexes. We found no differences in electrophysiological measurements between different ages, strains (see Figs. 2–4 versus Fig. 5) or sexes, so all animals were pooled together (see also Supplementary Note 1). Animals were housed at room temperature (~20 °C) and placed under a 12-h light/12-h dark cycle, with behavioral experiments performed during the light cycle. All animal handling procedures were approved by the Veterinary Service of the Canton of Vaud (authorizations VD2745 and VD3205).

Whole-cell recordings on acute brain slices

Animals were decapitated, and their brains were swiftly extracted and placed in chilled artificial cerebrospinal fluid (ACSF). The ACSF slicing solution was saturated with oxycarbon (95% O2 and 5% CO2) at pH 7.4 and contained 110 mM sucrose, 60 mM NaCl, 28 mM NaHCO3, 3 mM KCl, 1.25 mM NaH2PO4, 7 mM MgSO4, 0.5 mM CaCl2 and 5 mM d-glucose (Sigma-Aldrich). Acute, horizontal 400-μm-thick rat brain slices were cut between −8.6 mm and −7.6 mm depth from bregma39 using a vibratome (Compresstome VF-200, Precisionary Instruments); the presence of external capsule fibers and the beginning of the lateral ventricle were used as landmarks. After slicing, each of the usually obtained four slices was transferred on a nylon grid in a beaker filled with extracellular oxygenated ACSF solution (described below), with a recovery period of at least 1 h at room temperature before being transferred to the recording chamber. Under hyperexcitable conditions, to study epileptiform bursting activity, the KCl concentration was increased to 5 mM, and bicuculline-methiodide (20 µM; Sigma-Aldrich) was added to block GABAA receptors.

A semiautomated 12-patch-clamp setup40 (Fig. 1) was used to allow multiple-patch-clamp recordings. Cells were visualized by infrared differential interference contrast video microscopy using a VX55 camera (TILL Photonics) mounted on an upright BX51WI microscope equipped with an Olympus U-RFL-T lamp housing a 100-W mercury burner (Olympus Corporation). A group of up to 12 cells were selected for the electrophysiological recordings based on their morphology (pyramidal shaped) and, where applicable, their fluorescence. The identity of these cells was further confirmed following the injection of square pulses of hyperpolarizing and depolarizing 400-ms currents in 50-pA steps41,42 to assess accommodating and nonaccommodating subtypes or interneurons based on their high-frequency (that is, >30 Hz) firing rates. Besides the number of APs resulting from sustained current injection, accommodating and nonaccommodating neurons could be separated further by the delay required to observe an AP (that is, time-to-spike) following minimal stimulation (400 ms, 20-pA steps) with 111 ± 12 ms for nonaccommodating neurons and 82 ± 8 ms for accommodating neurons. In this manner, we classified nonaccommodating neurons with a time-to-spike of >100 ms. Interneurons were excluded from further experimentation and analyses (see also Supplementary Note 1 for more details).

Electrophysiological data were acquired with a Multiclamp 700B (Molecular Devices) in either current clamp or voltage clamp mode. Data acquisition was performed through an ITC-1600 board (Instrutech) connected to a PC running a custom-written routine (Pulse-Q) under IGOR Pro (Wavemetrics, version 7). Recordings were sampled at 10 kHz, and the recorded signal was filtered with a 5-kHz Bessel filter.

Recording pipettes of 4–10 MΩ were pulled from borosilicate capillary glass (Sutter Instrument; outer diameter: 1.5 mm; inner diameter: 0.86 mm; 7.5 cm length) by a P-97 Flame-Brown Micropipette Puller (Sutter Instrument). The pipettes were filled with an internal solution composed of 135 mM KMeSO4, 8 mM NaCl, 10 mM HEPES, 2 mM Mg2ATP, 0.3 mM Na3-GTP and 1 mg ml−1 biocytin (Sigma-Aldrich) with a pH of 7.3 and an osmolarity of 300 mOsm. The ACSF in the recording bath was composed of 118 mM NaCl, 25 mM NaHCO3, 10 mM d-glucose, 2.5 mM KCl, 1 mM MgCl2, 1.25 mM NaH2PO4 and 2 mM CaCl2 (Sigma-Aldrich) dissolved in deionized water of 18.2 MΩ cm resistivity. ACSF was supplemented with 1 mM l-glutamine (Sigma-Aldrich) to avoid homosynaptic depression43,44. Recorded neurons were considered stable in current clamp configuration if their membrane potential was lower than −55 mV and in voltage clamp configuration if less than 200 pA was required to maintain the membrane potential at −70 mV.

EPSP and EPSC analyses

Evoked EPSPs and EPSCs were analyzed using Mini Analysis software (Synaptosoft). Synaptic delay was measured as the time difference between the peak of the presynaptic AP and the onset of the postsynaptic response. Other criteria used for selecting EPSPs were 100 μV for minimal amplitude and 1 mV2 for the minimal area under the EPSP. A duration of 5 ms was used as a baseline, sampled 20 ms before the peak. For EPSCs, the minimal amplitude was 5 pA. Finally, the root mean square of the noise was measured over 0.5 ms at the beginning of each trial, outside of spontaneous or evoked responses, and was used for estimating the parameters of quantal analysis.

Assessment of connectivity

Once a patch clamp was obtained, a 3-ms, 1- to 4-nA square pulse was injected in each neuron to determine the AP firing threshold. Nine suprathreshold pulses were then delivered with eight pulses at 20 Hz, followed by a recovery pulse 550 ms later. This stimulation pattern was delivered successively to each of the patched neurons and repeated 15 times (Fig. 1c). Following this, an average of the responses to each neuronal activation was plotted, and the traces were assessed for time-locked EPSPs occurring within <5 ms with <2.5-ms jitter16,45. The presence of such EPSPs indicated a connection between neurons, which was further subjected to visual inspection that could readily and unequivocally confirm the actual presence of a connection (see examples in Extended Data Fig. 1b). Confirmed connections were then further subjected to quantal analysis or plasticity. LA network connectivity was calculated as the proportion of connections found in a given slice to all possible connections. To calculate the number of possible connections, while excluding autoconnections, we used n(n – 1), with n equal to the number of patched neurons for a given slice.

Quantal analysis experiments

The Ca2+:Mg2+ ratio was modified to isolate the synaptic quantum. Lower Ca2+ concentrations are characterized by a lower probability to observe an EPSC, thereby facilitating the extraction of the quantal size46. The following concentrations were used:

[Ca2+] = 2 mM and [Mg2+] = 1 mM (initial ACSF solution, as described above),

[Ca2+] = 1 mM and [Mg2+] = 2 mM and

[Ca2+] = 0.5 mM and [Mg2+] = 2.5 mM.

EPSCs were measured following the same stimulation pattern (for one trial, 8 + 1 APs delivered at 20 Hz), as described above. At least 30 trials were recorded at 2 mM Ca2+, and at least 100 trials were recorded at lower concentrations, and the first response of each trial was used for extracting the quantal parameters. The intertrial interval was at least 8 s. Stability of the response was tested by comparing the average response value for the first trial with the last ten trials. Connections that had over 30% variability were excluded from the analyses.

To estimate quantal size, we then identified EPSC and EPSP peaks from individual traces and fed these into a simple binomial model based on earlier observations that multiple release sites on cortical synapses share similar release probabilities47,48. Therefore, EPSC and EPSP amplitudes were drawn from a simple binomial distribution with a given number of release sites (n), release probability (p) and quantal size (q)49. The mean and standard deviation of this simple binomial distribution are given by

$$}\,}\,=q\sqrt\left(1-p\right)]}.$$

Extraction of quantal parameters was performed using a custom MATLAB script developed by Hardingham et al. to successfully describe the quantal content of the cortical layer 2/3 synapse using the above simple binomial model48,50. This model also uses the value for the recorded noise to better estimate n, p and q. In addition, the maximum likelihood method51 was used to fit the acquired data (fminsearch function from MATLAB’s Optimization Toolbox), starting with the lowest estimate for n and increasing until a maximum of arbitrarily defined 14 release sites was reached (if the model reached the maximum number of release sites, the resulting fit was discarded). For increased accuracy, the data were fitted against ten different starting points in the parameter space. Finally, the resulting fit was tested against simulated datasets sharing the same parameters using the Monte Carlo simulation and the χ2 test.

The binomial model has the following parameters:

v = amplitude of EPSC or EPSP;

v0 = offset, assumed to be added to all EPSCs or EPSPs;

σnoise = standard deviation of noise, assumed Gaussian;

n = number of release sites;

q = quantal size;

p = release probability at each release site;

pstim = probability that stimulation results in an AP that reaches the release sites (one in our case);

mσq = quantal variance, equals standard deviation on first peak in absence of noise; and

σm = variance affecting the mth peak, where m ranges from 0 to n.

For type I quantal variance, \(_}}^=_}}^+}_}}^\),

for ‘flat’ quantal variance, σm = σnoise for m = 0, and \(_}}^=_}}^+_}}^\) for m > 0.

The probability density function f(v) for estimating the EPSP or EPSC amplitude (v), as it was used in the original MATLAB script48,50, was the following:

$$\beginf\left(v\right)=_}}}\mathop\limits_^\frac^^\frac_\sqrt}\exp \left(-\frac_-,q)}^}_^}\right)\\\qquad\quad+\left(1-_}}}\right)\frac_}}}\sqrt}\exp \left(-\frac_\right)}^}_}}}^}\right).\end$$

Plasticity protocol

To trigger pre- and postsynaptic APs, a square pulse stimulus was used (1–2 nA, 3 ms) in a train of ten stimuli at 30 Hz repeated 15 times (intertrial interval of 10 s), with the presynaptic potential leading the postsynaptic potential by 5–10 ms (refs. 24,52). For testing whether the long-term potentiation-inducing protocol led to changes in EPSP amplitude, the same protocol was used as for assessing connectivity (eight pulses at 20 Hz, followed by a recovery pulse 550 ms later), which was applied every 30 s for up to 1 h, with potentiation typically lasting 20–30 min, that is, until deterioration of the multiple patched cells.

Surgery

All surgeries were performed under aseptic conditions with isoflurane anesthesia (5% initially and then 2% for maintenance) on a stereotaxic frame (David Kopf Instruments). Animals were kept on a heating pad throughout the duration and recovery from surgery. The animal’s scalp was opened to expose the skull, which was then cleaned with 3% H2O2. For both virus injections and in vivo electrophysiology, bilateral holes were drilled at −3.00 mm (anterior–posterior) and ±5.15 mm (medial–lateral) relative to bregma. If the bregma-to-lambda distance was less than 8.72 mm (reference for adult39), these coordinates were proportionally adjusted.

Viral vector and virus injection and infection

To fluorescently tag recently activated memory-participating neurons in the LA, we expressed dGFP (a fusion of Venus with an mODC PEST sequence with a half-life of 2 h (ref. 53); excitation peak: 515 nm; emission peak: 528 nm) under a modified minimal Arc promoter downstream of a synthetic E-SARE25.

Fluorescent tagging of recruited neurons ex vivo

We used an AAV 2/1 vector to express the E-SARE-driven dGFP. In addition to E-SARE–dGFP, the AAV contained a second cassette that expressed an RFP as an infection marker (RFP635; excitation peak: 588 nm; emission peak: 635 nm) under the constitutive enhanced phosphoglycerate kinase-1 (Pgk1) promoter, with a woodchuck hepatitis post-transcriptional regulatory element54.

Rats were infused bilaterally in the LA with the AAV at 4 weeks of age under isoflurane anesthesia and semisterile conditions. The bregma coordinates used were anterior–posterior −3.00 mm, medial–lateral ±5.15 mm and dorsal–ventral −7.8 mm. The injector consisted of a glass pipette containing 1.2 μl of AAV at a titer of 1.4 × 1013 genomic copies per ml. One microliter was lowered to a depth of 7.8 mm, and virus was infused over 10 min. The pipette was left in place for an additional 3 min to allow viral diffusion55.

Optogenetic tagging of recruited neurons in vivo

For activity-dependent optogenetic tagging following fear memory recall, we used a dual AAV system, in which the expression of double-floxed E-SARE-ChR2, delivered by one AAV, was controlled by tamoxifen-inducible recombinase ERT2CreERT2 delivered by another AAV56. As previously shown for tamoxifen-based gene activation25, to induce ChR2 expression, tamoxifen (10 mg) was administered by gavage 8 h before fear testing, as for mice25. This time window coincides with peak tamoxifen metabolism into 4-hydroxytamoxifen in rats57. Thus, 1 day after CFC, rats were administered tamoxifen and exposed only to the CS+ 8 h later to induce stable expression of ChR2 in recruited neurons following CFC recall. ChR2 expression was assessed by BL responses 1 week after CFC recall.

Microdrive implantation

Microdrives were built in-house and were implanted according to the same coordinates as used for the virus injections. They included eight tetrodes for a total of 32 channels. The tetrodes were assembled from nichrome wire of 25 μm in diameter (STABLOHM 675 California fine wire), which was insulated with heavy Formvar58. Tetrodes were mounted on the microdrive on a copper screw with a 270-μm step. For the optogenetic experiments, an optical fiber was mounted 200 μm away from the tetrode bundle.

To stably anchor the implantation, four fixation points surrounding the microdrive were created to each harbor a small bone screw, two of which were attached to a ground wire. A dental cement layer was used to secure the screws to the skull. The microdrive was lowered to a depth of 6.8 mm. The space between the electrodes and the skull was filled with softened paraffin. An additional layer of dental cement firmly attached the microdrive to the skull. Finally, a copper screen was fitted around the implanted microdrive as a partial Faraday cage to reduce noise during the recordings.

Electrophysiology in vivo

Electrodes were connected to a headstage (Plexon) containing 32 unity-gain operational amplifiers. The headstage was connected to a 32-channel computer-controlled preamplifier (with a gain of 1,000 and bandpass filter from 400 Hz to 7 kHz, Plexon). Neuronal activity was digitized at 40 kHz bandpass filtered from 250 Hz to 8 kHz and isolated by time–amplitude window discrimination and template matching using a multichannel acquisition processor system (Plexon).

During fear conditioning, spike waveforms and associated time stamps were recorded. For analysis, the artifact waveforms were removed, and the spike waveform minima were aligned using Offline Sorter 4.0 software (Plexon). Principal component scores were calculated for unsorted waveforms and plotted on three-dimensional principal component space, and clusters containing similar valid waveforms were manually defined (based on principal component and waveform feature spaces; Extended Data Fig. 10). A group of waveforms was considered to originate from a single neuron if it was defined as a discrete cluster in principal component space that was distinct from clusters for other units and if it displayed a clear refractory period (1.2 ms) in autocorrelograms59,60. Template waveforms were then calculated for well-separated clusters and stored for further analysis in MATLAB and to track neurons over time. To ensure that the same neuron was recorded over multiple sessions (6 h or more), we quantified the squared Mahalanobis distance, discarding neurons with unstable values. For further confirmation, we also measured cluster stability across recording sessions using J3 and Davies–Bouldin statistics59,60.

For each isolated unit and for each experiment, neuronal spikes were plotted as a raster plot of time stamps relative to stimulus exposure (CS+, CS– and BL; t = 0). Spike counts were binned in 50-ms bins (Fig. 6) or 0.5-s bins (Fig. 7) and normalized to a 500-ms (Fig. 6) or 5-s (Fig. 7) baseline average to obtain a z-score. Neurons were considered responsive to a stimulus if the z-score value of their activity crossed the significance level (3 s.d. compared to prestimulus baseline)59. In particular, for optogenetic experiments, neurons were considered BL sensitive if they responded with time-locked (<5-ms jitter) millisecond precision61 (Extended Data Fig. 9f).

Behavior: auditory CFC

To fear condition the rats to a tone (the CS), animals were placed in a fear conditioning box, which included a metal grid for scrambled shock delivery (0.45 mA over 2 s) to the feet and a clear Plexiglass top for camera recording. The delivery of the tone (12 kHz, 250-ms blips presented at 1 Hz over 20 s) and the shock as the unconditioned stimulus (US) were controlled by MedPC IV software. The stimuli (CS and/or US) were presented at random intervals (1–3 min). Auditory fear memory recall was tested in a different context, the testing cage, which was a hexagonal cardboard box with wooden bedding. The conditioning and test boxes were cleaned with a solution containing 70% ethanol after each session.

Animals were randomly assigned to three independent groups: homecage, CS/US paired and CS/US unpaired. The homecage group was not exposed to any aspect of the behavioral experiment (including surgery, habituation and handling). Following a 3-day recovery period from surgery, the rats from the unpaired and paired CS/US groups were handled and habituated to the conditioning cage and the testing cage in 5-min sessions once per day for 3 days. One day after the last habituation session, the animals were placed in the conditioning chamber. After a delay of 5 min, the paired group received three paired co-terminating presentations of the CS–US over 3–9 min. The unpaired control group received three consecutive US presentations, followed by three CS presentations. Twenty-four hours after fear memory acquisition, rats from the paired and unpaired groups received three presentations of the CS in the testing cage and were then returned to their home cages. After a delay of 90 min to achieve optimal dGFP expression, the animals were killed either by decapitation (for use for electrophysiology) or were deeply anesthetized with 4% isoflurane and perfused with phosphate buffer (PB; pH 7.4) and formaldehyde (4% in PB; fixation for confocal microscopy). Animals from the homecage group were killed following the same procedures. For the in vivo optogenetics experiments (Fig. 7b–d), a second tone (CS−; 5 kHz, continuous tone over 20 s) was presented that was never paired to the US, and we infused AAVs to make tamoxifen-inducible expression of ChR possible (see above). One week later, animals were subjected to CFC, and fear memory was recalled 24 h after learning in the presence of tamoxifen. After one additional week to allow ChR to be expressed, fear was recalled again, and neuronal responses to the CS+, CS– and BL were recorded. An additional group of animals (n = 4) was used for visualizing the fluorescent reporter ChR2 by confocal microscopy (Extended Data Fig. 8c). For the short-term memory in vivo experiments (Fig. 6), CS and US association and subsequent CS memory testing were separated by a 6-h interval21. We chose this shorter time interval to maximize the probability of recording from the same neurons reliably at the beginning and end of the experiment.

A FT200EMT optical fiber (Thorlabs) was mounted 200 μm adjacent to the tetrode bundle, with a laser (Dream Lasers) ensuring BL delivery at 473 nm, with a power at source of 50 mW and ~15 mW at the tip of the optical fiber. BL was delivered with a pulse width of 2 ms, either as a single pulse or at 20-Hz frequency trains.

Rats were considered as freezing to a stimulus (CS or BL) if no movement was detected for at least 2 s; the sum of the freezing bouts was then expressed as a percentage of the stimulus presentation. Freezing was assessed by an experimenter blind to the experimental conditions.

Immunohistochemistry and confocal image acquisition

For confocal microscopy experiments, following perfusion, brains were extracted and postfixed for 2 days in 4% formaldehyde and cryoprotected in 30% sucrose for an addition 2 days. Horizontal sections (50 µm thick) were cut on a MICROM HM 440E microtome (GMI).

To visualize inhibitory neurons, nonspecific binding sites on free-floating sections were blocked with 2% normal horse serum (Jackson Immuno Research Laboratories) in 0.1 M PB (pH 7.4) supplemented with 0.3% Triton X-100 (Sigma-Aldrich) and sodium azide (1 g l–1; Sigma-Aldrich) for 1 h at room temperature. Sections were then incubated with mouse primary anti-GAD67 (1:2,500; MAB5406, Merck Millipore) in blocking buffer for 48 h at 4 °C. To visualize antibody–antigen complexes, an AlexaFluor 405-conjugated goat anti-mouse antibody (1:300; A31553, Life Technologies) was applied in PB (with 0.3% Triton X-100 and sodium azide; Sigma-Aldrich) for 1 h at room temperature. Sections were then mounted with Vectashield (Vectorlabs) and stored at 4 °C until assessment by microscopy.

Fluorescence of both dGFP and mCherry following viral expression was sufficiently strong to be visualized directly.

Images were acquired using a Zeiss LSM 780 Quasar Confocal Microscope (Zeiss). Three lasers were used to excite at 405, 488 and 561 nm (diode, argon and diode-pumped solid-state lasers, respectively) at 2–3% power. Zen 2012 software was used to control the acquisition parameters of the LSM 780. Constant parameters included a pixel depth at 16 bit, filtering the average of two values and scanning unidirectionally and in ‘Line’ mode.

An ImageJ script was used to automatically detect fluorescence thresholds and count cell bodies. Cutoff parameters were used to minimize the inclusion of false positives. Exclusion criteria included cell size (minimal cutoff of 45 μm2 neuron body area) and fluorescence intensity (minimal cutoff of 15,000 average pixel value for a given channel out of a maximal 2 (ref. 16) or 65,536 for saturated pixels).

Statistics

All statistical analyses were conducted with GraphPad Prism 9 and R4.2 (ref. 62). Sample sizes were determined online based on mean difference and standard deviation (http://www.biomath.info; Center for Biomathematics, Department of Pediatrics at Columbia University Medical Center). When comparing populations, data were first tested for normality with either the Kolmogorov–Smirnov test or, when the sample data size was less than 50, with the Shapiro–Wilk test and for homogeneity of variance using Bartlett’s test. If the data did not deviate significantly from normal distribution and the variances were homogenous, independent samples t-tests, paired t-tests or ANOVAs were used. Results that were found to deviate significantly from the normal distributions and/or whose variance was not homogenous were analyzed with appropriate nonparametric tests (see below for a list of tests used). When multiple comparisons occurred, the tests were Bonferroni corrected.

When the Kolmogorov–Smirnov test was used to compare differences between two distributions, the cumulative frequency of each distribution was normalized to its maximal value.

For the RM-ANOVA, data were tested for sphericity using the Mauchly test, and to correct for departures from sphericity, the Greenhouse–Geisser and Huynh–Feldt corrections were applied63,64,65. Where applicable, statistical tests were two tailed. For cross-correlations, the corrplot function in MATLAB (MathWorks, 9.6) was used, with a Pearson test for linear correlations and a Spearman test for nonlinear correlations (resulting in their respective correlation coefficients, r and ρ).

The frequency of observed connectivity motifs was compared to their respective frequency in a simulated network (100,000 Monte Carlo simulations taking into account real cell positions and random distance-dependent connectivity based on Fig. 3a; see a detailed explanation in Supplementary Note 3).

The following statistical tests and variables (degrees of freedom are shown as underscore numbers) were used:

Student’s t-test uses the t variable.

ANOVA uses the F variable.

Wilcoxon signed-rank test uses the W variable.

Mann–Whitney U-test uses the U variable.

Kolmogorov–Smirnov test uses the D variable.

Monte Carlo simulation for network connectivity: n = 100,000 simulations.

For neuronal activity represented as z-scores, when assessing stimulus response, a poststimulus response was considered significant when its value was greater than 3 s.d. of the prestimulus baseline (see above).

Granger causality analysis

To determine connectivity (positive relationship with P < 0.001) and connection strength (Granger causality statistic) between two recorded neurons in vivo, we used Weiner–Granger vector autoregressive causality analysis as implemented in the multivariate Granger causality toolbox28,29. Time stamps recorded from each neuron were converted into a continuous signal by binning in 1-ms increments and convolving the resulting signal with a half-Gaussian filter (5-ms width). This analysis was performed on spike train data gathered over a period of 20 min outside of stimulus exposure. Stationarity was checked and confirmed for all models by determining whether the spectral radius of the estimated full model was less than 1 (ref. 29). Model order was derived by using Bayesian information criteria, and the vector autoregressive model parameters were determined accordingly. Subsequently, time domain-conditional Granger causality values were calculated for each neuron pair. Causal density was taken as the mean pairwise-conditional causality and was subsequently normalized to the entire dataset29. Because the current model assessed Granger causality between presynaptic and postsynaptic spikes, it could only infer excitatory but not inhibitory connections (Granger causality of presynaptic spike to postsynaptic silence).

Granger causality analysis was always performed on baseline data recorded from the animal at rest (either before or after fear conditioning) to prevent false-positive connections resulting from stimulus-evoked activity. Furthermore, we did not estimate connectivity (only connection strength) after fear conditioning because CFC increased the number of positive connections (P < 0.001). However, CFC did not affect connection strength (Granger causality statistics), as assessed in Fig. 6c where connection strength increased only in the recruited connections but not in the nonrecruited connections. Whether neurons were recruited or not was assessed after fear conditioning.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.