Association learning drives synaptic plasticity at feedforward synapses in somatosensory cortex

Abstract

Learning broadly alters neocortical synapses, although the input and target specificity for this plasticity has not been well-defined. Feedforward synapses into sensory cortex have early critical periods for plasticity after which they are resistant to experience-dependent changes. Whether these synapses are altered during learning has not been investigated, particularly in a setting where animals must identify causal relationships between sensory stimuli and rewards. Here, we examined whether these feedforward synapses can be altered by training mice in a freely-moving and whisker-dependent association task. Pathway-specific optogenetic stimulation and analysis of quantal excitatory postsynaptic currents in layer 2/3 (L2/3) pyramidal neurons from barrel cortex revealed a rapid and transient potentiation of layer 4 (L4) inputs at the onset of training, without any change in thalamocortical inputs onto L4 neurons. In contrast, pseudotraining—where stimuli and rewards were decoupled—drove depression of L4–L2/3 quantal excitatory postsynaptic currents. Because environmental enrichment did not influence quantal excitatory postsynaptic current amplitude, these data suggest that reward-prediction accuracy is a key driver of feedforward plasticity in primary sensory cortex.

Significance statement

Although it is well accepted that sensory learning can alter cortical synapses, the pathways that are modified and the specific cues that drive this synaptic change have not been systematically investigated. By manipulating stimulus–reward probabilities, we identified discrete and opposite changes in the strength of L4–L2/3 synapses depending on the predictive accuracy of the stimulus. These data suggest that feedforward sensory circuits are exquisitely sensitive to the predictive value of sensory input in a goal-directed task.

Introduction

Abundant experimental evidence links synaptic plasticity to learning, but the specific brain area and synapse types that are altered remains unclear. During sensory association learning, experiments have broadly focused on postsynaptic change in dendritic spines from pyramidal (Pyr) neurons, typically in primary sensory neocortex (Kuhlman et al. 2014; Li et al. 2025). However, there are multiple input sources that project into primary sensory cortex, and within the cortical column dozens of molecularly distinct neurons are connected to each other according to highly specific principles (Lefort et al. 2009; Pfeffer et al. 2013; DeNardo et al. 2015; Jiang et al. 2015; Barth et al. 2016). The sheer number of potential connections that could be modified has presented a daunting obstacle in understanding how and when learning reorganizes neural networks. This is a key gap in knowledge that precludes understanding how cortical circuits receive and transform behaviorally relevant sensory information.

Feedback pathways convey important state- and context-dependent information and are thought to be a critical locus of learning-related activity. These feedback pathways arise from upstream cortical areas or from higher-order (HO) thalamus where axons typically arborize in layer 1 (L1) to synapse onto the apical dendrites of Pyr neurons. Experiments in acute brain slices from naïve animals suggest that feedback inputs in L1 may be especially important for gating plasticity at synapses from layer 4 (L4), where sensory information first arrives at the cortex, onto L2/3 Pyr neurons (Williams and Holtmaat 2019; Pandey et al. 2023). Consistent with this, accumulating experimental evidence supports a role for feedback inputs in synaptic plasticity during learning, particularly at the onset of training (Makino and Komiyama 2015; Doron et al. 2020; Schroeder et al. 2023). In vivo, HO-thalamus activity is required for learning and perceptual discrimination (Pardi et al. 2020; La Terra et al. 2022; Qi et al. 2022) and synapses from HO-thalamus onto Pyr neurons are strengthened at early but not later stages of sensory learning (Audette et al. 2019; Ray et al. 2023). These data suggest that feedback inputs are important for early circuit modifications but may not encode the memory itself.

In contrast, much less is known about learning-related plasticity at neocortical feedforward synapses. In primary sensory cortex, these feedforward connections convey direct and short-latency information about sensory stimuli, from first-order (FO)-thalamic inputs or connections between layer 4 and layer 2/3 (L2/3) Pyr neurons. Thalamocortical synapses onto L4 neurons and intracortical synapses from L4 to L2/3 undergo an early critical period during development, after which sensory experience does not modify synaptic strength (Crair and Malenka 1995; Feldman et al. 1998; Wen and Barth 2011). Although in vitro studies indicate that feedforward synapses can be modified acutely (Feldman 2000; Banerjee et al. 2014; Williams and Holtmaat 2019) particularly with sensory deprivation (see for example Li et al. 2014), evidence for this plasticity during learning is lacking. Strengthening of feedforward inputs could selectively enhance relevant sensory representations critical for learning, but experimental evidence for this has been mixed (Yotsumoto et al. 2008; Jurjut et al. 2017; Audette et al. 2019; Gilad and Helmchen 2020; Ray et al. 2023; Zhu et al. 2024; Barth et al. 2025; Drieu et al. 2025; Li et al. 2025).

Some of this discrepancy may be due to the time window of analysis during training, either in newly expert animals or after prolonged training. Analysis of task-related, fast sensory responses can be difficult to ascribe to feedforward or feedback pathways, particularly using Ca⁺⁺-imaging data with slow response kinetics. In addition, plasticity in a small number of neurons may be obscured by population-level analyses (Poort et al. 2015; Doron et al. 2020; Li et al. 2025). Although several studies have identified anatomical plasticity onto the dendritic spines of L2/3 Pyr neurons during learning (Kuhlman et al. 2014; Li et al. 2025), these neurons receive both feedforward and feedback inputs, and the dendritic location of these changes has not been consistent across studies. Identification of the inputs that are potentiated during learning has important implications for where behaviorally relevant sensory representations are altered and where putative memory engrams might reside in the brain.

Here, we examine postsynaptic plasticity at two different types of feedforward synapses in primary somatosensory cortex of mice as they learn a sensory association task. Using molecular genetic tools for pathway-specific stimulation, we find that learning does not modify synapses from ventral posterior medial (VPM, also FO) thalamus but drives a rapid but transient increase in quantal excitatory postsynaptic currents (qEPSCs) at L4 to L2/3 Pyr synapses. This increase in synaptic strength is not concentrated in Pyr neurons expressing the activity-dependent transcription factor c-fos, suggesting that changes are not concentrated in a small group of “activated” neurons. Importantly, this synaptic potentiation is not phenocopied by housing animals in a sensory-enriched environment, indicating that stimulus–reward pairing may be critical for plasticity induction. Consistent with this, decoupling the stimulus from the reward initiates depression at L4 to L2/3 Pyr synapses.

Methods

Key resources table.

Reagent or resource	Source	Identifier
Bacterial and virus strains
AAV1-hSyn-hChR2(H134R)-eYFP	Addgene	Cat# 26973-AAV1
AAV9-Ef1a-DIO-hChR2(H134R)-eYFP-WPRE-HGHpA	Addgene	Cat# 20298-AAV9
pAAV9-EF1a-DIO-hChR2(H134R)-mCherry	Addgene	Cat# 20297-AAV9
Chemicals, peptides, and recombinant proteins
Alexa 568	Invitrogen	Cat# A10437
Alexa 488	Invitrogen	Cat# A10436
D-AP5	Hellobio	Cat# HB0225
Tetrodotoxin citrate	Hellobio	Cat# HB1035
Picrotoxin	Hellobio	Cat# HB0506
Lidocain N-ethyl bromide	Sigma-Aldrich	Cat# L5783
Isoflurane	Patterson Veterinary	Pivetal
Ketoprofen	Sigma-Aldrich	Cat# K1751
Experimental models: organisms/strains
C57BL/6J mice	Jackson Laboratory	RRID:IMSR_JAX:000664
Scnn1a-tg3-Cre mice	Jackson Laboratory	RRID:IMSR_JAX:009613
fosGFP (1–3 strain) mice	Carnegie Mellon University
Nelf-Cre mice	MMRRC	RRID:MMRRC_037424-UCD
Ai32 mice	Jackson Laboratory	RRID:IMSR_JAX:024109
Software and algorithms
Igor Pro 6.37	Wavemetrics	https://www.wavemetrics.com/products/igorpro
Origin Pro (2022b)	OriginLab	https://www.originlab.com/
MATLAB	Mathworks Inc.	https://www.mathworks.com/
Minianalysis 6	Synaptosoft
Custom codes for sensory association training	Audette et al. 2019	https://github.com/barthlab/Sensory-association-training-behavior
Custom code for regression discontinuity tree	This paper	https://github.com/barthlab/RegressionDiscontinuityTree
Other
IR beam-break sensor	Adafruit	Cat# 2167
Yun Shield v2.4	Dragino	Discontinued
Leonardo	Arduino	Cat# A000057
Relay shield for Arduino v2.1	DFRobot	Cat# DFR0144
Capacitive touch sensor	Adafruit	Cat# 1374
Solenoid valve	The Lee Company	Cat# LHDA1233115H
Gas regulator	Fisherbrand	Cat# 10-575-105

Experimental model and study participant details

Scnn1a-tg3-Cre mice were used to virally express channelrhodopsin-2 (ChR2) in excitatory neurons of L4. Nelf-Cre and offspring of Nelf-Cre x Ai32 mice were used to express ChR2 in excitatory neurons in VPm. FosGFP transgenic mice were used to fluorescently label c-fos-expressing neurons with a GFP reporter. C56BL6J mice were used to maintain backgrounds when breeding our cre-dependent mouse lines. For all experiments, both sexes were used (see Table S1). Mice were postnatal day (P) 23 to 41. All experiments were performed in accordance with the NIH guidelines and were approved by the Institutional Animal Care and Use Committee at Carnegie Mellon University (IACUC ID: IAMEND202400000121).

Viral injections

Scnn1a-tg3-Cre mice were mated to either C57b6 or fosGFP mice. Scnn1a-tg3-Cre litters between 1 and 4 postnatal days were neonatally injected in primary somatosensory cortex (X: −1.75, Y: +1.60, Z: −0.5) using a nanoject attached to a digital stereotax. Neonates were anesthetized for 10 to 15 min using an ice bath then placed into a custom-made mold for injection. S1 was targeted on the left hemisphere and injected with 100 to 400 nL of an AAV9-encapsulated cre-dependent channelrhodopsin virus tagged with a reporter. To study thalamic inputs onto L4 neurons, offsprings of Nelf-Cre mice crossed to Ai32 mice were used to express ChR2 in excitatory neurons in the ventral posteromedial nucleus of the thalamus (VPm). Alternatively, Nelf-Cre litter mates that were negative for Ai32 were stereotaxically injected with 200 to 400 nL of an AAV9-encapsulated Cre-dependent channelrhodopsin virus tagged with a reporter between 18 and 22 postnatal days in VPm using the coordinates X: −1.80, Y: −1.30, and Z: −3.40. Male and female mice were distributed across all experimental groups in this study (Table S1).

Automated home-cage training

Sensory-association training (SAT) was carried out in a custom-made home-cage training system (Bernhard et al. 2020). During training, animals were maintained on a 12-h dark–light cycle. Animals were placed in training cages between 11:00 am and 2:00 pm to carry out various training paradigms that provide sensory stimulus to the whiskers. Training cages were designed for trial self-initiation, where the mouse initiates a trial with a nose-poke that breaks an IR beam. This beam-break is followed by a random delay between 0.2 and 0.8 s to prevent mice from relating timing cues from the nose-poke with water delivery. The next trial could not be initiated until >2 s after airpuff onset. Anticipatory licking frequency during reward and blank trials was collected for 300 ms immediately prior to water delivery. Performance was calculated by subtracting anticipatory licking frequency for blank trials from stimulus trials (Lick_stimulus − Lick_blank).

For both pretraining and SAT, the probability of receiving a water reward was set to 80% of trials. Before sensory training, all mice were acclimated (ACC) to the training environment for at least 1 to 2 d prior to the stimulus–reward coupling. This enabled mice to learn to use the lick port, the only source of water in the cage. At the onset of SAT, the sensory stimulus (airpuff; 4 to 6 psi) was initiated at noon after 1 to 2 d of ACC. This sensory stimulus was presented to the right-side whiskers in a cone-like fashion from the ceiling of the training cage, ~ 4.5 cm above the animal as it was positioned at the lick port, and the stimulus was perfectly predictive of the water reward. Thus, 20% of trials had no stimulus and no reward, and 80% of trials had the airpuff stimulus followed by the water reward. For control, ACC-only mice were used in which the sensory stimulus was not present. Control mice were matched to trained animals based on total time in the training cage.

As an alternative to SAT, pseudotraining (PSE) was designed to decouple the sensory stimulus from the water reward. The PSE training paradigm uses the same trial structure as SAT, but in this paradigm the airpuff stimulus was no longer predictive of the water reward. In PSE, the probability of receiving a water trial is 50%, but like SAT, the stimulus was delivered on 80% of trials. During PSE, the initiation of a trial could trigger four possible outcomes including stimulus alone occurring on 40% of the trials initiated, stimulus and water delivery 40% of the trials initiated, water alone occurring on 10% of the trials initiated, and no water or stimulus occurring on 10% of the trials initiated. The design of this paradigm allowed investigation of whether the sensory stimulus delivery alone was sufficient to strengthen cortical pathways or if plasticity requires stimulus–reward coupling.

Environmental enrichment

To understand how environmental novelty impacts cortical synapses during early timepoints, we employed an enrichment condition. Enrichment was carried out by creating an environment with several new huts, a running wheel, and objects for investigation such as popsicle sticks and compacted paper for bedding. Mice were placed in the enrichment cage between 11:00 am and 2:00 pm. Mice spent 1 d in an enriched environment prior to recording. At the end of the environmental enrichment (EE), mice were removed from the training cage and immediately sacrificed for electrophysiological recordings.

General electrophysiology

At the end of all training conditions, mice were transferred to a new container, briefly anesthetized using isoflurane, and decapitated. Tissue was prepared between 11:30 am and 2:30 pm and cut in ice-cold standard artificial cerebral spinal fluid (ACSF) contained (in mM) 119 NaCl, 3.5 KCl, 1 NaH₂PO₄, 26.2 NaHCO₃, 11 glucose, 1.3 MgSO₄, and 2.5 CaCl₂ equilibrated with 95%/5% O₂/CO₂). Slices were allowed to equilibrate in ACSF at room temperature for at least 45 min prior to recording. We note that acute brain slices were prepared using a 45° rostro-lateral and 25° rostro-dorsal angle for cutting tissue, which may alter somatic distance from the pia compared to coronally sectioned tissue.

qEPSC measurement

To investigate the postsynaptic strength of L2/3 Pyr neurons, qEPSCs were measured (Clem et al. 2008; Wen and Barth 2011; Audette et al. 2019). Quantal events were detected through whole-cell patch-clamp recordings in ACSF containing 1 mM SrCl₂ in place of CaCl₂, except in VPM-evoked recordings where 2.5 mM SrCl₂ was used. ACSF was supplemented with D-AP5 at a final concentration of 50 μM to prevent NMDA receptor-mediated plasticity during recordings (Wen and Barth 2012) and 50 μM picrotoxin (PTX) to block GABA_A currents. To investigate the L4 to L2/3 synapses, 100 nM tetradotoxin was added to the ACSF to prevent cells from firing. Cesium-gluconate internal solution, used for all experiments, was composed of (in mM) 130 cesium gluconate, 10 HEPES, 0.5 EGTA, 8 NaCl, 10 tetraethylammonium chloride, 4 Mg-ATP, and 0.4 Na-GTP (pH 7.25 to 7.30, 280 to 290 mOsm) and contained 5 QX-314 (lidocaine N-ethyl bromide) and Alexa Fluor dye 488 or 568 to visualize recorded neurons. During recordings, cells were held at −70 mV and blue light through a 40× Olympus objective was used to stimulate axons at varying durations and light intensities. Cells with a series resistance <40 MΩ, input resistance >100 MΩ, and pyramidal cell morphology were included in the analysis. Events occurring within 500 ms post-stimulus, excluding events found on the primary response, were analyzed in MiniAnalysis 6 (Synaptosoft). The following detection parameters were used: amplitude threshold, 9 pA; local maximum period, 3.5 ms; peak for baseline period, 6 ms; decay time period, 10 ms; average baseline period, 4 ms; area threshold, 10 pA; averaging 3 values from the 10-kHz trace. A subset of data was analyzed blind to the training condition, and all the qEPSC analysis was done blind to laminar depth.

Paired recordings in fosGFP transgenic mice

To compare L4 EPSC amplitude across fosGFP+ and fosGFP− neurons, a glass stimulating electrode was placed in the center of a barrel in L4, directly below the targeted neurons (Benedetti et al. 2013). All recordings were performed in regular ACSF containing 50 μM D-APV and 50 μM PTX. Simultaneous recordings of fosGFP+ and fosGFP− neuron pairs were voltage clamped at −70 mV, and recordings were carried out using Cs-gluconate internal solution (see above) containing trace amounts of Alexa 568 for confirmation of Pyr cell identity. To facilitate a direct comparison of the evoked response, targeted neurons were <50 μm apart. Stimulus intensity and duration (1 to 10 ms) were varied to produce a well-isolated, monosynaptic response in fosGFP+ and fosGFP− neurons (Benedetti et al. 2013). To minimize the possibility of stimulus-related plasticity, only 1 pair was recorded from a single L4 barrel, and ~10 stimuli per pair were delivered. EPSC responses with an onset latency <12 ms were averaged for analysis, and averaged trials always included traces from both cells. Pairs where one or both cells had an access resistance >40 MΩ and resting membrane potential (mV) >−40 mV were excluded from analysis. Statistical analysis was carried out using a Wilcoxon paired test.

Laminar depth analysis

The laminar position of record neurons was determined by measuring the distance of the soma from the pia using ImageJ. A counting grid slide (Graticules Optics Cat# 02B00429) was used to calibrate the scale bar for images taken under 40× and 4× objectives. Cell somas were easily identified due to Alexa cell fills and the presence of the patch electrode. Vertical lines drawn from the soma center to the pia surface were measured and used as the depth value of the cell in the cortical column.

Regression discontinuity tree model

Regression discontinuity tree was used to identify the impact of a treatment (in this case, the effect of learning-induced plasticity) across a continuous variable (in this case, laminar position in superficial layers of the neocortex) where the probability of effect depends on the position in the continuous variable (Imbens and Lemieux 2008; Lansdell and Kording 2023). This design functions to identify a specific cutoff position that can be used to separate data into treatment-affected and unaffected groups. While this method is not traditionally applied to neurophysiological data, it has been widely used to identify genetic populations of neurons (He et al. 2021).

To apply this design to qEPSC-recording data, custom regression discontinuity tree analysis code was written in Python. Further details on the analysis code used are available on GitHub. The purpose of this code is to identify a specific cutoff point for a treatment (qEPSC amplitude) based on a running variable (laminar position) when there is a significant correlation between amplitude and depth. The following packages were used for implementation and visualization of the regression discontinuity analysis: sklearn, matplotlib, tkinter, scipy, and pandas (Pedregosa et al. 2011). All code written in Python were run using the spyder IDE software. Cutoff validity was assessed by the discontinuity statistics and by comparing the L2 and L3 groups per condition with a Mann–Whitney U test. Training groups where there was no significant regression produced cutoff values strongly skewed toward one end of the continuous variable. Therefore, regression discontinuity analysis was only performed on training groups in which a significant regression was identified.

Custom analysis code

All custom code used in this study will be freely available on GitHub.

Quantification and statistical analysis

All statistics were performed in Origin Pro 2022 or using custom code written in. All statistical tests ran with custom code were validated against values produced in Origin Pro 2022, except the regression discontinuity tree model. For comparing cell and animal averages across training conditions, a Mann–Whitney U t-test was used; a Kolmogorov–Smirnov (K-S) test was used for all cumulative percent comparisons. All behavioral comparisons where anticipatory licking frequency was assessed were compared using a Wilcoxon signed-rank sum test. All linear regressions were performed with a Pearson’s correlation to assess significance. Each figure reports the sample size as animal (N), cell (n), and event (q) number. All bar graphs represent the mean ± SEM. Licking frequency for stimulus and blank trials for the last 20% of trials completed on a given day were compared using a Mann–Whitney U t-test. Animals with a significant difference (P < 0.05) between lick rates for these two trial types were considered learners. Mice that did not significantly lick more during stimulus trials versus blank trials were considered non-learners (summarized in Table S2). For some mice, training data was not recovered due to file corruption, but assessment of water consumption indicated that task participation did occur. Therefore, the number of animals used for behavioral and electrophysiological measurements may sometimes differ.

Results

Sensory association training (SAT) that pairs a multiwhisker stimulus with a water reward drives rapid synaptic strengthening and increased response properties within barrel cortex, primarily in L2/3 and layer 5 (L5) (Audette et al. 2019; Ray et al. 2023; Zhu et al. 2024). Thalamocortical inputs from VPM thalamus to L4 are thought to be resistant to experience-dependent plasticity after an early critical period (Crair and Malenka 1995), and optogenetic VPM thalamic stimulation in acute brain stimulation does not reveal enhanced firing activity in L4 excitatory neurons at the onset of SAT (Audette et al. 2019). However, inhibitory neurons within L4 can undergo plasticity during—and are required for—some forms of association learning (Dobrzanski et al. 2022; Kanigowski and Urban-Ciecko 2024), suggesting that feedforward pathways may be a site of learning-related modifications. Synaptic plasticity at VPM synapses onto L4 neurons during SAT has not been investigated.

To test whether VPM synapses onto L4 stellate cells were strengthened at the onset of sensory learning, we expressed channelrhodopsin (ChR2) in VPM neurons using stereotaxic viral injections or in a Nelf-Cre transgenic mouse crossed to an Cre-dependent ChR2 (Fig. 1A). Animals were trained to associate a multiwhisker stimulus driven by a gentle airpuff (4 to 6 psi) with a delayed water reward in a freely moving, home-cage training paradigm (Audette et al. 2019; Bernhard et al. 2020; Ray et al. 2023; Zhu et al. 2024; Mosso et al. 2025; Park et al. 2025). In this SAT paradigm, mice typically increase anticipatory licking following the stimulus after 1 to 2 d of training, a finding that was also observed in this experimental cohort (Fig. 1B to F; cage acclimation control/ACC Lick_stimulus 5.90 ± 0.75 vs. Lick_blank 5.40 ± 0.81 Hz; SAT1 Lick_stimulus 7.62 ± 0.62 vs. Lick_blank 6.32 ± 0.90 Hz).

Figure 1.

VPM-evoked qEPSC amplitude in L4 is not altered after 1 d of association training. (A) Schematic of the experimental design. Transduction time for Nelf-Cre mice injected with a Cre-dependent ChR2 virus was between 7 and 11 d before recording. Mice underwent either ACC training or SAT training; all mice were allowed at least one ACC day prior to sensory stimulus being turned on. (B) Sensory association training trial structure. To initiate a trial, the mouse breaks an IR beam which triggers a varied delay (0.2 to 0.8 s) before beginning a trial. During SAT trials, a 6-psi airpuff is triggered for 0.5 s, indicated by the gray bar. Water is delivered 0.5 s after the end of the airpuff, indicated by the blue bar. Anticipatory licking data are collected 300 ms prior to water delivery to assess learning. (C) (left) Schematic of the trial structure for ACC, where 80% of trials initiated receive water. (right) Schematic of the trial structure for SAT, where 80% of trials initiated will receive a 0.5-s airpuff delivered to the right whiskers prior to water delivery. (D) Example image of a mouse approaching the lick port to initiate a trial. (E) (top) Anticipatory licking frequency for both rewarded (closed, green) and blank (open, red) trials plotted in 4-h time bins. Blue shading indicates trials with sensory stimulation turned on. (bottom) Performance values plotted in 4-h time bins. Blue shading indicates trials with sensory stimulation turned on. (F) Anticipatory licking frequency for stimulus (green, closed) and blank (red, open) for the last 20% of trials, compared using a Wilcoxon signed-rank sum test (ACC N6, P = 0.09; SAT1 N6, P = 0.04). (G) (left) A schematic of the experimental setup for recording excitatory neurons in layer 4. (right) Example image of VPM axons expressing ChR2 in barrel cortex (scale bar = 200 μm). (H) Global average of qEPSCs from control animals (black) or in animals trained for 1 d of SAT (blue). Scale bar 4 pA and 5 ms. (I) Cumulative percentage of VPM-evoked qEPSCs for ACC (black, dotted) and SAT1 (blue, solid) trained groups. Distributions are generated from 25 random events per cell and are compared using a Kolmogorov–Smirnov (K-S) test (ACC vs. SAT1 P = 0.31). Sample size is reported as animal (N), cell (n), and event (q) number. (J) Average qEPSC plotted by cell for ACC (black) and SAT1 (blue) trained animals. Groups were compared using a Mann–Whitney U (MW) test (ACC vs. SAT1 P = 0.95). (K) Same as (D) but plotted by animal averages (ACC vs. SAT1 P = 0.81). All bar graphs represent mean ± SEM. ^*P < 0.05.

Tissue from barrel cortex representing the stimulated whiskers was prepared from mice after 1 d of SAT and also from animals that had been acclimated to the training cage without directed whisker stimulation (ACC). Whole-cell patch-clamp recordings were made from L4 stellate neurons in acute brain slices to investigate changes in postsynaptic responses (Fig. 1G). Because expression levels of virally transduced ChR2 in VPM can vary across experimental animals, comparing light-evoked responses across conditions in L4 neurons is unreliable. Instead, we used a method for isolating postsynaptic response to a single quantal release events, replacing Ca⁺⁺ with Sr⁺⁺ in the bath solution (Clem et al. 2008; Audette et al. 2019).

Light-evoked qEPSCs were recorded in L4 neurons that were presumed to be spiny stellate cells, based on electrophysiological and anatomical characteristics, and compared across conditions (Fig. 1G). SAT did not drive an increase in light-evoked VPM qEPSC amplitude, analyzed either by a comparison of the population qEPSC distribution in a cumulative distribution histogram (Fig. 1H, I) or in the mean, averaged over cells or across animals (Fig. 1J, K; by cell ACC 12.8 ± 0.38 vs. SAT1 13.1 ± 0.58 pA; by animal ACC 12.7 ± 0.47 vs. SAT1 12.8 ± 0.49 pA). Thus, SAT is not sufficient to drive postsynaptic plasticity at VPM synapses onto L4 spiny stellate neurons. These results are consistent with analysis of VPM-evoked firing in L4 neurons, indicating no increase in activity after SAT (Audette et al. 2019).

Rapid strengthening of L4 inputs to L2/3 Pyr at onset of SAT

Layer 2/3 Pyr neurons undergo substantial plasticity during sensory learning (Kato et al. 2015; Makino and Komiyama 2015; Godenzini et al. 2022; Rabinovich et al. 2022; Zhu et al. 2024), and feedforward synapses onto L2/3 Pyr neurons can be acutely modified in vivo (Gambino et al. 2014) and in vitro (Williams and Holtmaat 2019), suggesting a potential substrate for changes in neural response properties. Our prior studies identified significant SAT-dependent strengthening of synapses from HO-thalamus and also intracortical synapses, presumably from L2/3 onto L2/3 Pyr neurons (Audette et al. 2019). However, some forms of experience-dependent plasticity at L4 to L2/3 excitatory synapses have a critical period around the second postnatal week (Wen and Barth 2011), and evidence to link excitatory synaptic plasticity to altered response properties of L2/3 neurons is lacking. Indeed, suppressed activity in L2/3 Pyr neurons during sensory learning has been observed in some studies (Gdalyahu et al. 2012; Makino and Komiyama 2015), suggesting that strengthening of feedforward pathways might be weak, present in only a minority of cells, or masked by enhanced inhibition.

We sought to test whether SAT would modify excitatory synapses from L4 onto L2/3 Pyr neurons using input-specific qEPSC analysis. Initially, we focused on plasticity after the first training day since this is when in vivo Ca⁺⁺ imaging studies have revealed a transient increase in L2/3 Pyr neuron activity during SAT (Zhu et al. 2024). Cre-dependent ChR2 was injected into Scnn1a-Cre transgenic mice for selective optogenetic stimulation of L4 excitatory neurons (Fig. 2A, B; Madisen et al. 2015).

Figure 2.

L4-evoked qEPSC amplitude in L2/3 Pyr is enhanced during early learning. (A) All animals used for this study were Scnn1a-cre transgenic mice that were neonatally injected with a Cre-dependent ChR2 virus. (B) Example image of excitatory neurons in L4 labeled with ChR2, scale bar = 200 μm. (C) (top) Anticipatory licking frequency for both rewarded (closed, green) and blank (open, red) trials plotted in 4-h time bins. Blue shading indicates trials with sensory stimulation turned on. (bottom) Performance values plotted in 4-h time bins. Blue shading indicates trials with sensory stimulation turned on. (D) Anticipatory licking frequency for stimulus (green, closed) and blank (red, open) for the last 20% of trials, compared using a Wilcoxon signed-rank sum test (ACC P = 0.71, SAT1 P = 0.40, SAT2 P = 0.014). (E) (left) Schematic of the experimental setup in S1 L2/3. (right) Example image of L4-expressing ChR2 for a single barrel column. (F) Cumulative percentage of L4-evoked qEPSCs for ACC (black, dotted), SAT1 (light blue, solid), and SAT2 (blue, solid) trained groups. Distributions are composed of 25 random events from each cell and are compared using a K-S test (ACC vs. SAT1 P = 2.6 × 10⁻⁶; ACC vs. SAT2 P = 2.5 × 10⁻⁷). Sample size is reported as animal (N), cell (n), and event (q) number. (G) Average qEPSC plotted by cell for ACC (black), SAT1 (light blue), and SAT2 (blue) trained animals. Groups were compared using a Mann–Whitney U test (ACC vs. SAT1 P = 0.0049; ACC vs. SAT2 P = 0.0046). (H) Same as (D) but plotted by animal averages (ACC vs. SAT1 P = 0.071; ACC vs. SAT2 P = 0.010). All bar graphs represent mean ± SEM. ^*P < 0.05.

Animals showed a similar learning trajectory as other mice that had undergone SAT (Audette et al. 2019; Bernhard et al. 2020; Ray et al. 2023) with a significant increase in anticipatory licking following the reward-predictive water stimulus on the second day of training (SAT2; Fig. 2C, D; ACC Lick_stimulus 5.24 ± 0.74 vs. Lick_blank 5.35 ± 0.77 Hz; SAT1 Lick_stimulus 6.42 ± 0.55 vs. Lick_blank 6.02 ± 0.52 Hz; SAT2 Lick_stimulus 6.96 ± 0.50 vs. Lick_blank 5.02 ± 0.40 Hz).

L2/3 Pyr neurons were targeted for recording light-evoked qEPSCs (Fig. 2E). A comparison of the population qEPSC distribution using a cumulative distribution histogram revealed a small but significant rightward shift in amplitude after 1 d of SAT, indicating an increase in qEPSC amplitude (Fig. 2F). These differences were significant when averaged over individual cells or animals (Fig. 2G, H; by cell ACC 14.3 ± 0.35, SAT1 15.18 ± 0.26, and SAT2 15.7 ± 0.39 pA; by animal ACC 14.5 ± 0.70, SAT1 15.01 ± 0.31, and SAT2 15.7 ± 0.42 pA). Thus, SAT initiates plasticity at feedforward intracortical synapses from L4 to L2/3 at the onset of training, even at a time when many animals do not demonstrate an increase in anticipatory licking to the stimulus.

L2 inputs are selectively strengthened at the onset of SAT

Because some studies suggest that Pyr neurons in L2 and L3 may be both anatomically and molecularly distinct (Gur and Snodderly 2008; Sorensen et al. 2015; van Aerde and Feldmeyer 2015; Deitcher et al. 2017) and, indeed, can participate in separable cortical microcircuits (Lefort et al. 2009; DeNardo et al. 2015), we asked whether the magnitude of synaptic strength might be different depending on the location of Pyr cell soma (Fig. 3A). Analysis of qEPSCs from control (ACC) animals suggested no difference in amplitude across the depth of L2/3 (Fig. 3B). However, after 1 d of SAT, qEPSC amplitudes showed a significant negative correlation with soma depth, where Pyr neurons at the top of L2 showed larger qEPSC amplitudes than those at the bottom of L3 (Fig. 3C).

Figure 3.

Intracortical qEPSC strengthening is enriched in L2 Pyr. (A) Example image of recorded cells filled with Alexa cell fill. Scale = 50 μm. (B) Correlation of cell amplitude and cell distance from the pia calculated using Pearson’s correlation for ACC (R = −0.16, P = 0.30). (C) Correlation of cell amplitude and cell distance from the pia calculated using Pearson’s correlation for SAT1 (R = −0.32, P = 0.027). Gray dotted line indicates the cutoff predicted by regression discontinuity tree (209.8 μm). (D) Correlation of cell amplitude and cell distance from the pia calculated using Pearson’s correlation for SAT2 (R = 0.010, P = 0.95). (E) Schematic of the experimental setup in L2. (F) L2 global average of qEPSCs from control animals (black), animals trained for 1 d of SAT (light blue), or animals trained for 2 d of SAT (blue). Scale bar 4 pA and 5 ms. (G) L2 cumulative percentage of L4-evoked qEPSCs for ACC (black, dotted), SAT1 (light blue, solid), and SAT2 (blue, solid) trained groups. Distributions are composed of 25 random events from each cell and are compared using a K-S test (ACC vs. SAT1 P = 1.9 × 10⁻⁵; ACC vs. SAT2 P = 0.0016). Sample size is reported as animal (N), cell (n), and event (q) number. (H) L2 average qEPSC plotted by cell for ACC (black), SAT1 (light blue), and SAT2 (blue) trained animals, compared using a Mann–Whitney U test (ACC vs. SAT1 P = 0.0014; ACC vs. SAT2 P = 0.014). (I) L2 average qEPSC plotted by animal for ACC (black), SAT1 (light blue), and SAT2 (blue) trained animals, compared using a Mann–Whitney U test (ACC vs. SAT1 P = 0.0026; ACC vs. SAT2 P = 0.011). (J) Schematic of the experimental setup in L3. (K) L3 global average of qEPSCs from ACC (black), SAT1 (light blue), or SAT2 (blue) trained animals. Scale bar 4 pA and 5 ms. (L) L3 cumulative percentage of L4-evoked qEPSCs for ACC (black, dotted), SAT1 (light blue, solid), and SAT2 (blue, solid) trained groups. Distributions are composed of 25 random events from each cell and are compared using a K-S test (ACC vs. SAT1 P = 0.18; ACC vs. SAT2 P = 0.014). (M) L3 average qEPSC plotted by cell for ACC (black), SAT1 (light blue), and SAT2 (blue) trained animals, compared using a Mann–Whitney U test (ACC vs. SAT1 P = 0.54; ACC vs. SAT2 P = 0.15). (N) L3 average qEPSC plotted by animal for ACC (black), SAT1 (light blue), and SAT2 (blue) trained animals, compared using a Mann–Whitney U test (ACC vs. SAT1 P = 1; ACC vs. SAT2 P = 0.15). All bar graphs represent mean ± SEM. ^*P < 0.05.

Using regression discontinuity tree analysis (Imbens and Lemieux 2008), a statistical method to determine whether a data distribution could be segregated across a continuous variable, we asked whether the effects of SAT were related to the sublaminar distribution of Pyr soma depth from the pial surface. This analysis suggested that there was a natural split in qEPSC amplitudes, where more superficial Pyr neurons showed larger (and also more variable) qEPSCs than those measured deeper in the column. This analysis enabled us to identify a discrete depth by which to separate superficial and deeper Pyr neurons, at 210 μm from the pial surface. Note that the angle at which acute brain slices were prepared to retain columnar connectivity might increase this value slightly compared to more conventional coronal sections.

Because there was no correlation for qEPSC values versus depth in the control (ACC) dataset (Fig. 3B), the regression discontinuity analysis was not applied. Although many studies have separated L2 from L3, the rationale and depth used for this split has not always been clearly articulated (Lefort et al. 2009; DeNardo et al. 2015; van Aerde and Feldmeyer 2015; Weiler et al. 2022). This depth aligns with prior studies that have segregated L2 and L3 Pyr neurons using molecular and electrophysiological properties (Sorensen et al. 2015; Cheng et al. 2022; Brandalise et al. 2025).

We used the value from this statistical analysis to separately average qEPSC values from L2 and L3. As suggested from the correlation across depths, population analysis using a cumulative distribution histogram to compare L4-evoked qEPSC amplitude from control (ACC) and SAT1 indicated a pronounced and highly significant rightward shift in qEPSC amplitude for L2 (Fig. 3E to G). This significant difference was maintained in comparing cell-averaged qEPSC amplitude between conditions for L2 (Fig. 3H, I; by cell ACC 14.1 ± 0.33, SAT1 15.8 ± 0.34, and SAT2 15.8 ± 0.49 pA; by animal ACC 14.0 ± 0.34, SAT1 15.7 ± 0.30, and SAT2 15.8 ± 0.53 pA). In L3, a population-level analysis of qEPSCs showed no increase at SAT1 and a modest enhancement at SAT2 (Fig. 3J to L) that was not significant at either the cell or animal average (Fig. 3M, N; by cell ACC 14.4 ± 0.63, SAT1 14.1 ± 0.23, and SAT2 15.4 ± 0.67 pA; by animal ACC 14.83 ± 0.817, SAT1 14.2 ± 0.28, and SAT2 15.7 ± 0.71 pA). To investigate if postsynaptic plasticity is correlated with learning, we split mice into learners and non-learners based on their ability to differentiate anticipatory licking between stimulus and blank trials (see Methods for details). Although we found significant changes in mean synaptic strength after 1 and 2 d of SAT, we did not detect a significant correlation with animal performance for neurons in either L2 or L3 (Supplemental Fig. 1, Table S2; note that the L2 qEPSC amplitude shows a nonsignificant, positive correlation with performance at SAT2 (P = 0.06)). Thus, these data suggest that there may be a functional difference between superficial and deeper L2/3 Pyr neurons, selectively revealed at the onset of SAT.

Fos-expressing neurons do not selectively capture SAT-evoked synaptic potentiation

The immediate-early gene (IEG) c-fos, associated with elevated neural firing, can be used to identify neurons activated by behavioral experience. Reactivation of fos-expressing neurons in some brain areas is sufficient to drive memory-related behaviors (Josselyn and Tonegawa 2020), suggesting that IEG expression might “tag” neural ensembles involved in memory. Because elevated postsynaptic activity can facilitate synaptic potentiation (spike-timing dependent plasticity/STDP; Bi and Poo 1998), it is reasonable to predict that more active neurons might have a greater likelihood of undergoing synaptic potentiation during learning. However, experimental support for selective synaptic plasticity on activated, IEG-expressing neurons after behavioral experience has been limited (Koya et al. 2012; Hwang et al. 2022).

Fos-expressing neurons in somatosensory cortex show elevated spontaneous and sensory-evoked activity under basal conditions (Yassin et al. 2010; Jouhanneau et al. 2014). Despite this, SAT does not increase the overall number of fos-tagged neurons in L2/3 of barrel cortex and thalamocortical synaptic strengthening is not concentrated in fos-expressing pyramidal neurons in superficial layers (Lee et al. 2021). However, recent studies suggest that other IEGs besides c-fos might label a subset of Pyr neurons with increased input strength after sensory learning (Li et al. 2025).

Even L2 Pyr neurons show substantial variability in qEPSC input strength during SAT, suggesting that some neurons might be selectively modified. To test the hypothesis that the increase in synaptic strength at L4 to L2/3 synapses was concentrated in fos-expressing Pyr neurons, we trained fosGFP transgenic mice in SAT (Fig. 4A) and compared input strength between fosGFP+ and fosGFP− Pyr neurons after 1 d of training. L4 was electrically stimulated and evoked EPSCs were simultaneously collected in paired whole-cell recordings from neighboring (<50 μm distance) fosGFP+ and fosGFP− neurons, centered in L2 (Fig. 4B, C). Responses were collected in the presence of GABA_A receptor antagonists to isolate the EPSC, and D-APV was present to prevent further stimulation-induced NMDA-receptor dependent plasticity (Wen and Barth 2012).

Figure 4.

L4 inputs onto fosGFP-expressing L2 Pyr are not potentiated during early learning. (A) (left) Schematic of the behavioral training; mice either underwent ACC training or SAT1. (B) Schematic for recording pairs of fosGFP+ and fosGFP− cells in L2 of S1. (C) Example image of fosGFP+ expression in a L2 Pyr cell. (middle) Example image of a recorded pair filled with cell fill. (right) Example image of fosGFP and cell fill overlayed. Scale bar = 20 μm. (D) Example of a single L4-evoked EPSC in fosGFP− (black) and fosGFP+ (green) for ACC. Scale bar is 50 pA and 10 ms. (E) Red dot indicates the mean ± SEM for L4-evoked EPSC (pA) of fosGFP+ and fosGFP− cells for ACC. Black dots, individual cell pairs (two-sample Wilcoxon signed-rank test, P = 0.84). Each figure reports the sample size as animal (N) and cell (n) number. (F) Same as (D) but for SAT1 animals. Scale bar is 50 pA and 10 ms. (G) Same as in (E) but for SAT1 animals (two-sample Wilcoxon signed-rank test, P = 0.26).

Surprisingly, L4 inputs onto fosGFP+ neurons were not significantly larger than onto fosGFP− neurons from trained animals (Fig. 4D to G; fosGFP+ 40.3 ± 16 vs. fosGFP− 62.3 ± 23 pA; 6 animals, n = 10 pairs, P = 0.26). Input strength was also not greater onto fosGFP+ neurons in control (ACC) mice (fosGFP+ 34.2 ± 11 vs. fosGFP− 35.7 ± 11 pA; 5 animals, n = 10 pairs, P = 0.84). Thus, expression of the IEG c-fos in fosGFP transgenic mice was insufficient to explain heterogeneity in L4 input strengthening, in primary somatosensory cortex.

Renormalization of L4–L2/3 excitatory inputs after prolonged training

After 5 d of training, performance on this task plateaus (Fig. 5A to C; ACC Lick_stimulus 6.39 ± 1.1 vs. Lick_blank 6.35 ± 1.0 Hz; SAT5 Lick_stimulus 6.06 ± 0.60 vs. Lick_blank 2.75 ± 0.61 Hz), and other neurophysiological measures of network plasticity have stabilized or returned to baseline values (Audette et al. 2019; Ray et al. 2023; Zhu et al. 2024; Park et al. 2025).

Figure 5.

Learning-induced qEPSC plasticity in L2/3 is transient. (A) Licking frequency for both rewarded (closed, green) and blank (open, red) trials plotted in 4-h time bins. Blue shading indicates trials with sensory stimulation turned on. (B) Performance values plotted in 4-h time bins. Blue shading indicates trials with sensory stimulation turned on. (C) Licking frequency for stimulus (green, closed) and blank (red, open) for the last 20% of trials. Licking frequency was compared using a Wilcoxon signed-rank sum test (ACC P = 0.86, SAT5 P = 0.014). (D) Schematic of the experimental setup in S1 L2/3. (E) Cumulative percentage of L4-evoked qEPSCs for ACC (black, dotted) and SAT5 (dark blue, solid) trained groups. Distributions are composed of 25 random events from each cell and are compared using a K-S test (ACC vs. SAT5 P = 0.58). Sample size is reported as animal (N), cell (n), and event (q) number. (F) Average qEPSC plotted by cell for ACC (black) and SAT5 (blue) trained animals. Groups were compared using a Mann–Whitney U test (ACC vs. SAT5 P = 0.75). (G) Same as (D) but plotted by animal averages (ACC vs. SAT5 P = 0.94). (H) Correlation of cell amplitude and cell distance from the pia calculated using Pearson’s correlation for ACC (R = −0.14, P = 0.52). (I) Correlation of cell amplitude and cell distance from the pia calculated using Pearson’s correlation for SAT5 (R = −0.58, P = 7.7 × 10⁻⁵). The gray dotted line denotes the predicted split based on RDD (207.9). (J) Schematic of the experimental setup in L2. (K) L2 global average of qEPSCs from ACC (black) or SAT5 (dark blue) trained animals. Scale bar 4 pA and 5 ms. (L) L2 cumulative percentage of L4-evoked qEPSCs for ACC (black, dotted) and SAT5 (dark blue, solid) trained groups. Distributions are composed of 25 random events from each cell and are compared using a K-S test (ACC vs. SAT5 P = 0.32). L2 average qEPSC plotted by cell for ACC (black) and SAT5 (blue) trained animals. Groups were compared using a Mann–Whitney U test (ACC vs. SAT5 P = 0.41). (M) L2 average qEPSC plotted by animal for ACC (black) and SAT5 (blue) trained animals. Groups were compared using a Mann–Whitney U test (ACC vs. SAT5 P = 0.34). (N) Schematic of the experimental setup in L3. (O) L3 global average of qEPSCs from ACC (black) or SAT5 (dark blue) trained animals. Scale bar 4 pA and 5 ms. (P) L3 cumulative percentage of L4-evoked qEPSCs for ACC (black, dotted) and SAT5 (dark blue, solid) trained groups. Distributions are composed of 25 random events from each cell and are compared using a K-S test (ACC vs. SAT5 P = 1.1 × 10⁻⁵). (Q) L3 average qEPSC plotted by cell for ACC (black) and SAT5 (blue) trained animals. Groups were compared using a Mann–Whitney U test (ACC vs. SAT5 P = 0.097). (R) L3 average qEPSC plotted by animal for ACC (black) and SAT5 (blue) trained animals. Groups were compared using a Mann–Whitney U test (ACC vs. SAT5 P = 0.16). All bar graphs represent mean ± SEM. ^*P < 0.05.

We thus examined whether qEPSC amplitudes in L2/3 Pyr neurons would maintain or undergo further changes in response amplitude once animals were expert in the task (Fig. 5D to G). At SAT5, the regression discontinuity analysis suggested a natural break in qEPSC measurements according to Pyr soma depth that localized to 208 μm, nearly identical to the depth identified at SAT1 (Fig. 3C). To control for possible effects of longer exposure to the training cages for mice trained for 5 d, control mice underwent a similar duration of ACC (ACC6). There was no correlation of qEPSC amplitude with soma depth in ACC mice (Fig. 5H).

Despite the correlation of qEPSC amplitude with laminar depth at SAT5, mean qEPSCs in L2 Pyr neurons were similar to control mice (Fig. 5L to N; by cell ACC6 15.0 ± 0.43 vs. SAT5 15.4 ± 0.27 pA; by animal ACC6 14.7 ± 0.64 vs. SAT5 15.3 ± 0.19 pA). In L3, qEPSCs showed a small though significant decrease in amplitude, likely the cause of the negative correlation observed (Fig. 5I). However, this decrease was not detected in cell or animal qEPSC averages (Fig. 5R, S; by cell ACC6 14.5 ± 0.40 vs. SAT5 13.6 ± 0.22 pA; by animal ACC6 14.6 ± 0.41 vs. SAT5 13.8 ± 0.31 pA).

To investigate if postsynaptic plasticity is correlated with learning at this timepoint, we again tried to separate animals into learners and non-learners. However, unlike early training timepoints, all mice had learned the task by 5 d of training so that learners and non-learners could not be compared. Performance, or the difference in anticipatory licking across trial types, was not correlated with the magnitude of synaptic strength in either L2 or L3 Pyr neurons (Supplemental Fig. 2, Table S2). These data indicate that postsynaptic excitatory strengthening at L4 to L2/3 inputs is not only layer dependent but also transient, most pronounced in L2 at early stages of association learning.

Decoupled stimulus–reward pairing initiated synaptic depression at L4–L2/3 synapses

SAT involves the consistent coupling of multiwhisker stimulation followed by reward. To determine whether the predictive value of the whisker stimulus was important in driving potentiation of L4 to L2 qEPSCs, animals were exposed to a pseudotraining procedure (PSE), where whisker stimulation was decoupled from reward outcome (Fig. 6A, B).

Figure 6.

L4-evoked qEPSC are depressed during pseudotraining. (A) Schematic of the experimental design. All animals used for this study were Scnn1a-cre transgenic mice that were neonatally injected with an AAV-DIO-ChR2 virus. Mice were trained in either ACC or PSE prior to whole-cell recording. (B) (left) Schematic of the trial structure for ACC, where 50% of trials initiated receive water. (right) Same as ACC, but on trials that receive water a 0.5-s airpuff will sometimes be coupled to water delivery. (C) (top) Anticipatory licking frequency for both stimulus (closed, green) and no stimulus (open, red) trials plotted in 4-h time bins. Blue shading indicates trials with sensory stimulation turned on. (bottom) Performance values plotted in 4-h time bins. Blue shading indicates trials with sensory stimulation turned on. (D) Schematic of the experimental setup in S1 L2/3. (E) Cumulative percentage of L4-evoked qEPSCs for ACC (black, dotted), PSE1 (red, solid), and PSE2 (maroon, solid) trained groups. Distributions are composed of 25 random events from each cell and are compared using a K-S test (ACC vs. PSE1 P = 1.2 × 10⁻⁵; ACC vs. PSE2 P = 0.063). Sample size is reported as animal (N), cell (n), and event (q) number. (F) Average qEPSC plotted by cell for ACC (black), PSE1 (red), and PSE2 (maroon) animals. Groups were compared using a Mann–Whitney U test (ACC vs. PSE1 P = 0.0017; ACC vs. PSE2 P = 0.25). (G) Same as (D) but plotted by animal averages (ACC vs. PSE1 P = 0.046; ACC vs. PSE2 P = 0.45). (H) Correlation of cell amplitude and cell distance from the pia calculated using Pearson’s correlation for ACC (R = 0.030, P = 0.86). (I) Correlation of cell amplitude and cell distance from the pia calculated using Pearson’s correlation for PSE1 (R = −0.095, P = 0.57). (J) Correlation of cell amplitude and cell distance from the pia calculated using a Pearson’s correlation for PSE2 (R = −0.52, P = 0.0037). Gray dotted line indicates the cutoff predicted by RDD (199.8 μm). (K) Schematic of the experimental setup in L2. (L) L2 global average of qEPSCs from ACC (black), PSE1 (red), or PSE2 (maroon) trained animals. Scale bar 4 pA and 5 ms. (M) L2 cumulative percentage of L4-evoked qEPSCs for ACC (black, dotted), PSE1 (red, solid), and PSE2 (maroon, solid) trained groups. Distributions are composed of 25 random events from each cell and are compared using a K-S test (ACC vs. PSE1 P = 1.8 × 10⁻⁵; ACC vs. PSE2 P = 0.81). (N) L2 average qEPSC plotted by cell for ACC (black), PSE1 (red), or PSE2 (maroon) trained animals, compared using a Mann–Whitney U test (ACC vs. PSE1 P = 0.16; ACC vs. PSE2 P = 0.55). (O) L2 average qEPSC plotted by cell for ACC (black), PSE1 (red), or PSE2 (maroon) trained animals, compared using a Mann–Whitney U test (ACC vs. PSE1 P = 0.41; ACC vs. PSE2 P = 0.27). (P) Schematic of the experimental setup in L3. (Q) L3 global average of qEPSCs from ACC (black), PSE1 (red), or PSE2 (maroon) trained animals. Scale bar 4 pA and 5 ms. (R) L3 cumulative percentage of L4-evoked qEPSCs for ACC (black), PSE1 (red), or PSE2 (maroon) trained groups. Distributions are composed of 25 random events from each cell and are compared using a K-S test (ACC vs. PSE1 P = 1.1 × 10⁻⁵; ACC vs. PSE2 P = 0.0019). (S) L3 average qEPSC plotted by cell for ACC (black), PSE1 (red), or PSE2 (maroon) trained animals, compared using a Mann–Whitney U test (ACC vs. PSE1 P = 0.0027; ACC vs. PSE2 P = 0.022). (T) L3 average qEPSC plotted by cell for ACC (black), PSE1 (red), or PSE2 (maroon) trained animals, compared using a Mann–Whitney U test (ACC vs. PSE1 MW P = 0.066; ACC vs. PSE2 P = 0.093). All bar graphs represent mean ± SEM. ^*P < 0.05.

Importantly, these animals received a similar proportion of whisker stimulus trials (80%), although only half of the stimulus trials were followed by reward, and sometimes a reward was delivered without any stimulus. As expected, animals did not develop a stimulus-associated anticipatory licking response during PSE (Fig. 6C).

Unexpectedly, we observed that qEPSCs from L4 onto L2/3 Pyr neurons were depressed after 1 d of PSE (Fig. 6E to G; by cell ACC 15.2 ± 0.31, PSE1 14.0 ± 0.20, and PSE2 14.8 ± 0.34 pA; by animal ACC 14.9 ± 0.41, PSE1 14.1 ± 0.26, and PSE2 14.7 ± 0.35 pA). The correlation between soma depth and qEPSC amplitude was only present after 2 d of PSE, revealing a discontinuity at 200 μm that separated L2 and L3 Pyr neurons, a depth similar to that detected at SAT1 and SAT5 (Fig. 6H to J). Accordingly, we compared qEPSC amplitude for L2 versus L3 between ACC and PSE animals. L2 Pyr neurons showed a more modest leftward shift after PSE in the cumulative distribution (Fig. 6M), but this difference was not significant when averaged across cells or animals (Fig. 6N, O; by cell ACC 15.0 ± 0.41, PSE1 14.2 ± 0.31, and PSE2 15.7 ± 0.50 pA; by animal ACC 14.7 ± 0.40, PSE1 14.2 ± 0.32, and PSE2 15.6 ± 0.51 pA) and disappeared after 2 d of PSE.

The effect of PSE in driving synaptic depression was most pronounced in L3 Pyr, where the population qEPSC showed a significant leftward shift as well as a significant decrease in mean qEPSC amplitude (Fig. 6R to T; by cell ACC 15.3 ± 0.46, PSE1 13.8 ± 0.25, and PSE2 13.9 ± 0.35 pA; by animal ACC 15.1 ± 0.60, PSE1 13.9 ± 0.34, and PSE2 13.9 ± 0.42 pA). Thus, PSE generates opposite effects on L4 excitatory synapses compared to SAT. Because PSE maintains the proportion of stimulus trials, it is likely that the contingency of the reward, given the stimulus, is a key variable in driving this change.

Environmental enrichment does not alter feedforward synapses onto L2/3 Pyr neurons

Excitatory synaptic plasticity can be driven not only by learning but sensory experience during environmental enrichment (Yang et al. 2009; Landers et al. 2011; Ohline and Abraham 2019). To determine whether L4 inputs onto L2/3 Pyr neurons might be sensitive to other conditions known to drive synaptic plasticity, we exposed mice to a novel sensory environment enriched with diverse objects as well as a running wheel. Animals were housed in this environment for a single day since 1 d of SAT was sufficient to initiate qEPSC strengthening.

In general, novel environmental stimuli promote whisker-dependent exploration (Deschênes et al. 2012), and both novelty and curiosity are intrinsically rewarding (Kang et al. 2009; Gruber et al. 2014). Thus, we hypothesized that environmental enrichment might couple whisker-dependent sensory input with reward signals, similar to stimulus–reward coupling in the more controlled SAT paradigm, that would be sufficient to drive plasticity at the L4 to L2/3 synapse.

Mice were housed in this EE, and then acute brain slices were prepared to assess qEPSC amplitude at L4 inputs to L2/3 Pyr neurons (Fig. 7A to D). Control animals were housed in standard vivarium cages and qEPSCs were recorded for comparison to the EE group. Enrichment failed to drive any plasticity at L4 to L2/3 synapses, either in the cumulative distribution of qEPSC values (Fig. 7E) or in cell or animal averages (Fig. 7F, G; by cell naïve 14.5 ± 0.37 vs. EE 14.0 ± 0.27 pA; by animal naïve 14.1 ± 0.39 vs. EE 13.6 ± 0.35 pA). There was no significant correlation of qEPSC amplitude with depth, and analysis of neither L2 nor L3 across conditions revealed any change with EE (L2 by cell naïve 14.9 ± 0.36 vs. EE 14.9 ± 0.52 pA, P = 0.79; L3 by cell naïve 14.1 ± 0.37 vs. EE 13.4 ± 0.48 pA, P = 0.09). Thus, exposing mice to an EE does not alter the postsynaptic strength of L4 synapses onto Pyr neurons in superficial layers. These data suggest that novelty and exploratory behavior is not sufficient to initiate synaptic strengthening, and that L2/3 synapses are selectively sensitive to the introduction of stimulus–reward outcomes.

Figure 7.

Enrichment does not alter L4 input strength to superficial layers after 1 d. (A) Example images of the home-cage environment (left) or the enriched environment (EE) (right). Enrichment contained several items to play with, multiple houses, and a running wheel. (B) Schematic of the experimental design. Mice were taken from their home cage or placed into an enrichment environment for 1 d prior to whole-cell recording. (C) Schematic of the experimental setup in S1 L2/3. (D) Global average of qEPSCs from naïve animals (black) or animals that underwent 1 d of enrichment (gray). Scale bar 4 pA and 5 ms. (E) Cumulative percentage of L4-evoked qEPSCs for naïve (black, dotted) and EE (dark blue, solid) groups. Distributions are composed of 25 random events from each cell and are compared using a Kolmogorov-Smirnov (K-S) test (naïve vs. EE K-S P = 0.40). Sample size is reported as animal (N), cell (n), and event (q) number. (F) Average qEPSC plotted by cell for naïve (black) and EE (blue) animals. Groups were compared using a Mann–Whitney U test (naïve vs. EE MW P = 0.13). (G) Same as (D) but plotted by animal averages (naïve vs. EE MW P = 0.20). All bar graphs represent mean ± SEM. ^*P < 0.05.

Discussion

Here, we identify input- and target-specific plasticity at feedforward excitatory synapses in primary somatosensory cortex that is regulated by stimulus–reward pairing during learning. Excitatory neurons in L2/3 but not L4 show a transient strengthening of glutamatergic inputs during SAT but depression during PSE (Fig. 8). Our data also suggest that these plasticity mechanisms may be differentially manifested in Pyr neurons based upon target-cell depth within L2/3. The absence of plasticity at the L4 to L2/3 pathway in animals under EE indicates that exploratory whisker activity and novelty are not sufficient to drive synaptic changes.

Figure 8.

Summary of training-related plasticity in L4–L2/3 synapses. (A) Schematic of the experimental setup in L2. (B) Comparison of the average of the percent difference from acclimation in L2 recorded cells for SAT1 (light blue, P = 0.0014) and SAT2 (blue, P = 0.014) using a Mann–Whitney U test. (C) Same as (B) but for SAT5 (dark blue, P = 0.41). (D) Same as (B) but for PSE1 (red, P = 0.16) and PSE2 (maroon, P = 0.55). (E) Same as (A) but for L3. (F) Same as (B) but for L3 SAT1 (light blue, P = 0.54) and SAT2 (blue, P = 0.15). (G) Same as (C) but for L3 SAT5 (dark blue, P = 0.097). (H) Same as (D) but for PSE1 (red, P = 0.0027) and PSE2 (maroon, P = 0.022). All bar graphs represent mean ± SEM. ^*P < 0.05.

Synaptic plasticity at feedforward intracortical synapses and network activity

Enhancing stimulus representations has long been considered a useful mechanism for encoding learning-related information, and even passive sensory experience can increase the response probability of neurons in somatosensory cortex (Benedetti et al. 2009; Glazewski and Barth 2015). Although the increase in qEPSC amplitude was small—on average, only 10%—it is important to note that this reflects detected changes at single inputs onto L2/3 Pyr neurons. Since L4 neurons are connected to individual Pyr neurons by multiple synapses, this small difference may be amplified across the target population. Such an increase may drive additional spiking in L2/3 Pyr neurons, increasing the sensitivity of animals to tactile stimuli broadly. Indeed, we have observed an increase in stimulus-evoked Ca⁺⁺ fluorescence restricted to the onset of SAT (Zhu et al. 2024), similar to the transient increase in qEPSC amplitude reported here. Other tactile learning studies have reported an increase in task-related Ca⁺⁺ activity of some L2/3 Pyr neurons during training (Chen et al. 2015; Gilad and Helmchen 2020; Rabinovich et al. 2022 but see Makino and Komiyama 2015; Le Merre et al. 2018; Pancholi et al. 2023). It is likely that Ca⁺⁺ imaging studies underestimate postsynaptic spiking due to the sensitivity of the indicator and imaging frequency, so that single spikes evoked by sensory activation may be undetected.

Although qEPSC measurements isolate postsynaptic changes in Pyr response properties, changes in the number of synapses or in presynaptic release properties could not be detected by Sr⁺⁺ desynchronized release analysis. It is also worth noting that replacement of Ca⁺⁺ with Sr⁺⁺ in the bath can reveal a different population of vesicles that are released during stimulation (Searl and Silinsky 2002; Bhalla et al. 2005), which might influence qEPSC measurements. In addition, the use of the Ca⁺⁺-permeable, light-activated ChR2 also makes it difficult to isolate changes in presynaptic Ca⁺⁺ dynamics that may occur during learning. Thus, it remains open whether presynaptic plasticity or a change in the number of synapses at the L4 to L2/3 excitatory synapse are initiated during SAT or PSE. The transient changes we observed do not preclude other changes in the local network that might underlie a long-lasting memory trace or homeostatic downscaling of synaptic weights during some consolidation period.

Notably, the differences reported in the present study were observed in acute brain slices, indicating that they are durable and not task related. These long-lasting synaptic changes might thus influence sensory processing outside of the task context. Increased activation of neurons in sensory neocortex may facilitate the detection of sensory stimuli, via improving signal flow to downstream areas (Quiquempoix et al. 2018). In contrast to SAT, depression of feedforward transmission during PSE might be a way for cortical circuits to ignore irrelevant sensory input. The critical downstream areas and synaptic targets that are selectively modified during association learning remain largely unknown.

Layer differences in the cortical column

Although L2 and 3 are frequently combined in the analysis of neocortical Pyr neurons, a molecular and anatomical separation between these neurons has been suggested by multiple studies (Lefort et al. 2009; Sorensen et al. 2015; van Aerde and Feldmeyer 2015; Cheng et al. 2022; Condylis et al. 2022; O’Toole et al. 2023). Although we were unable to identify putative Pyr subtypes in this study, we did observe a marked difference in the magnitude of plasticity induction between superficial and deep L2/3. Using regression discontinuity analysis, this border localized to ~210 um below the pial surface, a value that was similar across multiple experimental conditions. This depth shows remarkable correspondence to other studies that have segregated L2/3 Pyr neurons using cytoarchitecture features (DeNardo et al. 2015). Spatial transcriptomic analysis suggests that superficial L2/3 is dominated by Pyr neurons expressing Trpc6 and that these cells preferentially project to frontal areas (Sorensen et al. 2015). However, other studies suggest that molecularly distinct populations of L2/3 Pyr neurons are intermingled, without a strong laminar preference (Condylis et al. 2022; O’Toole et al. 2023). Recent studies using different genetic markers suggest a more robust laminar division; thus, this is an area of active research (Cheng et al. 2022; Butrus et al. 2025). Our data show that SAT can reveal layer-specific differences, where L2 exhibited faster and stronger qEPSC potentiation than in L3 after training. Furthermore, PSE initiated a significant and sustained depression of feedforward inputs onto L3 Pyr neurons, further supporting a distinction between L2 and L3. Future studies to relate the classification of L2/3 Pyr neurons with synaptic plasticity may be useful in testing the hypothesis that feedforward plasticity is target-specific among the larger population of L2/3 Pyr neurons.

Differences in feedforward plasticity might also be attributed to altered inhibition within these sublamina. However, using light-evoked SST or PV inhibition in SST-ChR2 or PV-ChR2-expressing mice, we did not observe depth-specific differences in the magnitude of SST-inhibitory postsynaptic currents (SST-IPSCs) or PV-IPSCs (Kuljis et al. 2020; Park et al. 2025, 2026) for control and ACC animals (Supplemental Fig. 3). After SAT, we observed a positive correlation between soma depth and PV-IPSC and SST-IPSC amplitude, with reduced inhibition from both sources in L2 neurons. Depending upon when this disinhibition is initiated (ie in the first few hours of SAT), it could conceivably facilitate the potentiation of L4–L2 excitatory synapses observed at SAT1.

Neuroanatomical studies of L2/3 Pyr neurons indicate that L2 Pyr have a larger and more elaborated apical dendrite in L1 (Oberlaender et al. 2012; Weiler et al. 2022), where they may receive different long-range inputs as well as layer-specific inhibition (Schuman et al. 2021). These inputs may act cooperatively with L4 inputs to facilitate synaptic potentiation, as has been observed in reduced preparations (Williams and Holtmaat 2019). Altered cortical inhibition might influence plasticity at multiple synapse types that were not explored. We note that VPM thalamocortical axons can synapse directly onto L2 and L3 Pyr neurons; plasticity at these inputs was not investigated in our study but could also be influenced by altered cortical inhibition.

fosGFP and engrams within sensory cortex

IEG expression can label ensembles of stimulus-driven neurons and IEG-expressing neurons can show enhanced response properties to specific sensory inputs (Barth et al. 2004; Wang et al. 2006; Jouhanneau et al. 2014). For this reason, it has been hypothesized that IEG expression in sensory cortex might tag a subset of neurons that have been activated and undergone stimulus-specific plasticity during learning (Li et al. 2025). Although L4 excitatory inputs underwent a broad-scale increase in synaptic strength across the population of supragranular neurons—particularly in L2—fosGFP+ neurons did not show enhanced drive from L4 after SAT compared to unlabeled, neighboring fosGFP− neurons. Thus, expression of the IEG c-fos was not predictive of synaptic potentiation at this pathway. This result was similar to the analysis of HO-thalamic input plasticity, where qEPSCs from HO-thalamus onto L2/3 Pyr neurons were larger after SAT but mean evoked EPSCs from optogenetic activation of HO-thalamus onto fosGFP+ neurons did not show selective potentiation (Lee et al. 2021). It remains possible that fosGFP+ neurons show increased input after SAT from other intracortical pathways that have yet to be identified.

The changes we observed during both early- and late-stage learning suggest that if there is a long-lasting memory trace in primary sensory cortex, it is not concentrated in fosGFP+ neurons and it does not require sustained enhancement of individual synaptic weights in this feedforward pathway (but see Li et al. 2025). Other changes in the local network that might underlie a long-lasting memory trace or homeostatic downscaling of synaptic weights during some consolidation period have not been investigated. Alternatively, it is possible that a long-lasting, learning-related engram is not present in superficial layers of primary sensory cortex.

Stimulus–reward contingencies during training and synaptic plasticity

SAT and PSE initiated distinct and opposite patterns of synaptic plasticity at L4–L2/3 synapses (summarized in Fig. 8). Because the stimulus probability was conserved between these two behavioral paradigms, we conclude that feedforward sensory input is not sufficient to initiate synaptic potentiation. Indeed, it is worth noting in this freely moving paradigm that the animals spend only a small fraction of their time engaged in training, <30 min out of a 24-h window, and that whisker-related activity is more or less constantly present. Other studies have employed passive sensory stimulation to contrast with learning-related plasticity and have shown a reduction in the population response imaged in vivo after multiple days of isolated (non-contingent) stimulus exposure in headfixed animals (Cooke et al. 2015; Kato et al. 2015; Makino and Komiyama 2015; Lopez-Ortega et al. 2024; Drieu et al. 2025). These data are consistent with the decrease in qEPSC amplitude we observed at L4–L2/3 synapses in PSE.

In contrast, passive sensory experience such as occurs during EE—where there is no temporally correlated reinforcement—lacks both positive and negative reward-prediction signals. Although EE can facilitate synaptic gain in the hippocampus (Ohline and Abraham 2019), analysis of barrel cortex in our study did not reveal EE-associated potentiation of L4–L2/3 inputs. Notably, prior studies from our laboratory indicate that PSE fails to drive synaptic plasticity at other cortical synapses that were affected by SAT, for example at HO-thalamocortical synapses onto L5 Pyr neurons or onto and from SST interneurons (Audette et al. 2019; Mosso et al. 2025; Park et al. 2025). Thus, feedforward excitatory synapses may be particularly susceptible to down-weighting during PSE and passive sensory stimulation.

Mechanisms for differential plasticity during SAT and pseudotraining

What are the key variables that distinguish SAT from PSE? Whisker stimuli delivered during PSE do not recapitulate passive sensory input since the structure of that training paradigm decorrelates the stimulus from reward, using both stimulus-only trials and reward-only trials. Animals may unsuccessfully attempt to extract an association between the stimulus and reward during PSE, leading to both correct predictions of rewards, given a stimulus but also some incorrect predictions, where the stimulus arrives without ensuing reward.

Because the stimulus probability is conserved but the outcome differs between SAT and PSE conditions, we hypothesize that the predictive relationship of the stimulus to the reward is a critical factor that regulates the direction of synaptic strength at this feedforward synapse. Because the water reward follows the stimulus at some delay, there may be a retroactive mechanism to assign value to the predictive sensory stimulus. Alternatively, sustained activity or synaptic activation in sensory cortex may create a trace for subsequent reward signals to interact with and potentiate cortical synapses (Chubykin et al. 2013; He et al. 2015). Since the unreliable prediction of stimulus to reward in PSE is not simply neutral but over time initiates depression, there may be a separate and distinct pathway conveying expectation violation or reward uncertainty to primary sensory cortex (Hong et al. 2022).

These data are consistent with two separate and opposing neuromodulatory pathways—one associated with a highly predictive stimulus and another associated with uncertainty. Such signals have been observed for monoamine neuromodulators like norepinephrine and serotonin (Aston-Jones et al. 1997; Grossman et al. 2022; Su and Cohen 2022; Jordan and Keller 2023) and signals from these pathways are shown to have profound effects on excitatory plasticity (Martins and Froemke 2015; Hong et al. 2022). Acetylcholine may also be involved in reward signaling and can influence activity of both inhibitory and excitatory neurons in the neocortex (Hangya et al. 2015; Zannone et al. 2018). The signaling pathways that convey reward information to cortical circuits, their relationship with local plasticity mechanisms, and their influence on sensory information processing are of great interest.

Author contributions

Joseph A. Christian (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Visualization, Writing—original draft, Writing—review & editing), Eunsol Park (Data curation, Formal analysis, Investigation, Writing—review & editing), and Alison L. Barth (Conceptualization, Data curation, Funding acquisition, Methodology, Project administration, Resources, Supervision, Visualization, Writing—original draft, Writing—review & editing). All behavior training and in vitro slice recordings for qEPSCs were performed by J.A.C. All paired recordings of fosGFP neurons were performed by E.P. Data analysis was done under the supervision of A.L.B. The manuscript was written by A.L.B. with help from J.A.C.

Funding

This work was supported by the National Institutes of Health (NIH R01 NS123711 (to A.L.B.) and NIH R21 NS127354 (to A.L.B.).

Conflicts of interest

A.L.B. serves on an advisory board for the NIH-funded Sensory Biology COBRE Center, at the University of Wyoming. She is also on the Scientific Advisory Board for the Cluster of Excellence NeuroCure, Charité-Universitätsmedizin Berlin Germany and for the program Inhibitory neurons: shaping the cortical code (IN-CODE), German Science Foundation, Freiburg Germany. A.L.B. also holds a patent for the fosGFP transgenic mice.