Learning binds new inputs into functional synaptic clusters via spinogenesis.

Data availability: Data and available upon request from corresponding author. Code availability: code used to analyze data and generate figures for this manuscript available upon request from corresponding author.

Animals

All animal procedures were performed in accordance with guidelines set forth and protocols approved by the UCSD Institutional Animal Care and Use Committee and the National Institutes of Health.

Mice

(C57BL/6 for all experiments described in this study) were grouped-housed in disposable cages with standard bedding in a temperature- and humidity-controlled room (~21°C and 42% humidity) with a reversed light cycle (10am-10pm: dark). All experiments were performed during the dark cycle. After surgeries, animals were singly housed. Males and females were randomly used for surgeries, with no selection criteria other than surgery outcome.

Surgery

Adult mice (6 weeks or older) were anaesthetized with isoflurane in an enclosed, ventilated chamber (5% isoflurane with a constant flow rate of 1L/min at 0.1 Bar and 21℃) until a deep plane of anaesthesia was reached, as indicated by low muscle tone and slowed breathing rate. Baytril (10mg/kg) and dexamethasone (2mg/kg) were injected subcutaneously to prevent infection and brain swelling, respectively. Skin and connective tissue over the dorsal surface of the skull was removed, the skull was slightly scored with a scalpel, and a custom stainless steel headplate was glued to the skull surface. A craniotomy (~3mm diameter) was performed, as previously described 9 , 10 , over the right caudal forelimb area around the central coordinate of ~300 μm anterior and ~1,500 μm lateral from bregma. Viruses (AAV1-CMV-PI-CRE and AAV1-Syn-FLEX-GCaMP6f from Addgene/UPenn Vector Core; AAV1-Syn-FLEX-SF-iGluSnFR-A184S, construct generously received from Dr. Loren Looger) were diluted to achieve sparse expression of iGluSnFR or GCaMP6f (1:1 mixture of iGluSnFR or GCaMP6f and 1:5000–10000 Cre in saline + 0.5% FastGreen for visualization of injections) and injected into the region of the caudal forelimb area of the exposed cortex using beveled glass pipettes (~12–25 μm inner diameter). Each injection consisted of a ~20nL volume at a depth of ~250 μm from the pial surface to target layer 2/3. Injection volumes were dispensed over the course of ~2 minutes. Multiple (3–5) injections were performed in each craniotomy, separated by at least 500 μm. Pipettes were left in the brain for 4 minutes after injection to avoid backflow. Chronic imaging windows consisting of a 3mm diameter plug glued to a larger (~5mm) glass base were then implanted in the craniotomy. The window was held in place with gentle pressure while the edges were affixed to the skull with small amounts of surgical glue (VetBond). Buprenorphine was injected subcutaneously at the end of surgery for pain management. Water restriction Animals were allowed to recover from surgery for ~10–14 days, after which they were progressively water-restricted (2mL/day for 3 days, 1.5mL for 3 days, then 1mL/day for the remainder) for ~14 days. Weight was constantly monitored to ensure loss of no more than 30% starting body weight.

Behavior

After water restriction, mice were trained in the lever-press task for 14 days. Simultaneous 2-photon imaging was performed on sessions 1–3 (“early sessions”), 6–8 (“middle sessions”), and 11–13 (“late sessions”). The lever comprised a piezoelectric flexible force transducer (LCL-113G, Omega Engineering) attached to a 1/14-mm-thick brass rod. The lever position was continuously recorded using a data acquisition device (LabJack) and software (Ephus, MATLAB, Mathworks) working with custom software running on LabVIEW (National Instruments) which monitored threshold crossing. The behavioral setup was controlled by MATLAB software (Dispatcher, Z. Mainen and C. Brody) communicating with a real-time system (RTLinux). A 6-kHz tone was presented to indicate a period during which a lever-press was rewarded with water (~8uL per trial) paired with a 500ms,12-kHz tone, followed by an intertrial interval of 8–12 sec. Successful lever-presses were defined as those crossing two thresholds ~1.5mm and ~3mm below the resting position) within 200ms. The 3mm threshold defined the target lever displacement, while the 1.5mm threshold ensured that the mouse did not hold the lever near the target threshold. Failure to perform a successful press during cue presentation triggered a white noise punishment signal and the start of the next inter-trial interval. “Non-cued” presses during the inter-trial interval were neither rewarded nor punished. Mice were exposed to 100 trials each day, or until the mouse became disengaged (no movements for 20+ trials) or satiated (no licking in response to water delivery). In vivo 2-photon imaging Imaging was performed using a commercial two-photon microscope (B-Scope, ThorLabs) equipped with a 16x/0.8-NA objective (Nikon) and a Ti-Sa laser (MaiTai, Newport) tuned to 925nm (or 810nm for control experiments). The laser power coming through the objective was controlled with a Pockel’s cell, and ranged from 10–40mW for these experiments. Image acquisition was controlled through Scanimage software. Imaging was always performed in awake animals. Images (256 × 512 pixels at either 8.5× or 12.1× zoom, corresponding to ~60 × 120 μm and 42 × 85 μm, respectively) were recorded at approximately 58.3Hz in 5-minute-on, 5-minute-off intervals for the duration of the behavioral session. Such interleaved imaging was performed so as to limit phototoxicity to the dendrites. The median number of trials imaged per day using this method was 60 (95% CI = [57, 62]). Further, imaging for a single field was performed in 5-day intervals, with three fields being selected for each animal, such that Field 1 was imaged on sessions 1, 6, and 11, Field 2 on sessions 2, 7, and 12, and Field 3 on session 3, 8, and 13. In our hands, such an imaging schedule preserved the health of most dendrites, preventing obvious dendritic blebbing and/or photo-bleaching. Separate fields were always at least 500 μm apart, taking advantage of the multiple injection sites. In 4 of the 45 iGluSnFR fields, imaging was performed only on early and late sessions. A small subset of imaging fields was processed for subsequent electron microscopy. In these cases, a high-resolution, low-zoom (1024 × 1024 pixels at 1× and 5× zoom, corresponding to ~1mm × 1mm and 250 × 250 μm fields, respectively) z-stacks of the target region were acquired from the pial surface to the target soma at 1μm intervals.

Identification of target dendrites

Fluorescent dendrites visible near the pial surface (presumed layer 1) were typically observed within the first month after viral injections. Dendrites in the superficial layers were targeted, with an average imaging depth of 35.1 ± 0.9 μm (mean ± SEM). As the injections were targeted to L2/3 of M1, expression of a reporter construct should be enriched in L2/3 excitatory neurons. Since expression in L5 is also possible, however, we took care to trace target dendrites in L1 back to their parent apical dendrites and follow the apical branch back to its parent soma to assess the laminar location of the cell whenever possible. In addition to their markedly different laminar depths, the apical dendrites of L2/3 excitatory neurons are also morphologically distinct from those of L5 excitatory neurons, with L2/3 typically presenting a more “shrub-like” appearance (shorter primary apical dendrite with long, tortuous higher-order branches) vs. the “tree-like” appearance of L5 neurons (long primary apical dendrite with more planar arborization/”tuft” in L1). Furthermore, a previous study 53 has shown that the spine density of L2/3 excitatory neurons (~0.73/μm) is higher than that of L5 excitatory neurons (~0.47/μm). Consistent with accurate targeting of L2/3 dendrites, the median spine density from the data shown in this manuscript is ~0.67/μm. Field identification across longitudinal imaging sessions On the first imaging session for each field (i.e. sessions 1,2, and 3), bright-field images of the surrounding vasculature were acquired, and the depth of FOV from the pial surface was recorded, allowing landmark-driven identification of the same field in future sessions. On subsequent sessions, the single-plane average projection image of the time series acquired on the first session was used as reference to frame the imaging field as similarly as possible. Brief (~10–20s) time series were first acquired, subjected to motion-correction (see description below), and projected so as to provide a comparison image against the previous session image. Care was taken to adjust the z-plane so as to maximally reproduce the appearance of the reference image prior to commencing the experiment.

Evaluation of dendritic health

Due to the risk of phototoxic effects from imaging, dendritic health was carefully monitored. The punctuated imaging schedule used in this study ( Extended Data Fig. 1c ) was designed through pilot experiments in which imaging power and duration were decreased until dendritic health was maintained through the end of the experiment. Dendritic health was evaluated based on: 1) spine density (analysis detailed below), with large decreases signaling poor cellular health, 2) dendritic morphology, with any “bleb”-like structure reflecting cellular death, 3) fluorescent event frequency (detailed below), with global decreases potentially indicating damage. In our data, median spine densities are not different between early (0.67 spines/μm) and late (0.61 spines/μm) sessions ( Extended Data Fig. 1e ). Event frequencies were largely stable, showing similar values for MRSs across early and late sessions, with a small decrease apparent in nonMRSs. Taken together, these data suggest that dendritic health was maintained in the data presented in this study.

Movement analysis

Movement analyses were performed as previously described 9 , 10 . Briefly, lever displacement traces (voltage recordings from the force transducer) were down-sampled from 10kHz to 1kHz, then filtered using a 4-pole 10Hz low-pass Butterworth filter, after which the velocity of the lever was determined by smoothing the difference of consecutive points with a moving average window of 5 ms. The envelope of the lever velocity was then extracted using a Hilbert transform, and movement bouts were defined by the envelope crossing a threshold of 4.9mm per second. Each movement bout was extended by 75 ms on either side. Bouts separated by less than 500 ms were considered continuous. Movement start and end times were defined as the points at which the lever exceeded or fell below the thresholds defined by rest periods before and after the movement bouts. Thresholds were defined as the resting position plus the 99 th percentile of the noise distribution, in turn defined as the difference between the Butterworth smoothed trace and the original trace. For reaction time analysis, the trials in which mice were moving the lever within 100ms before the cue start time were excluded from analysis. The median number of trials excluded for this purpose was 28.4 (95% CI = [23.8, 32.7]), corresponding to 35% (95% CI = [32.2, 37.9]) of trials. Movement correlations within sessions were calculated using the median of all pairwise correlations of rewarded movements that started after cue onset within a single session. Movement correlations across sessions were found using the median correlations of all possible pairs of movements between sessions (one movement taken from one session and the other movement taken from the other session). The movement correlation heatmap shown in Fig. 1d was computed by taking the mean within- and across-sessions correlation computed above for each animal. For correlations with the learned movement pattern, the learned movement pattern was defined as the average of rewarded movements that started after cue onset from the late (11–14) learning sessions. Movements coincident with new spine-MRS pair co-activity were defined as any movements executed during the session (i.e. not limited to rewarded movements) that overlapped with co-activity. Prolonged movements (lasting >3s) typically corresponded to repeated movements in succession, and were therefore excluded from this analysis.

Image analysis

Lateral motion of imaging time series was corrected using custom full-frame cross-correlation image alignment 54 . Motion within each frame was negligible due to the fast frame rate. To register fields across sessions, the average projection images of each of the time series corresponding to a particular imaging field were subjected to image alignment. To assist with spine categorization and new spine identification, duplicates of the session-aligned images were iteratively deconvolved (Diffraction PSF 3D; FIJI/ImageJ). Regions of interest (ROIs) were manually drawn using custom MATLAB software. For dendritic spines, elliptical ROIs were drawn around the center of the spine head beyond the edge of detectable fluorescence above background ( Extended Data Figure 4 ). Apparent spines along the z-axis of the dendritic shaft were included in analysis. Series of regularly-spaced elliptical ROIs were also drawn along the length of the dendrite, with the center of each ellipse serving as the point along a poly-line being used to calculate dendritic distance between spines. A single, large ROI was drawn in an empty region of the field to estimate background. For display purposes in figures, images were manually cropped around dendrites of interest for visual clarity. Care was taken (using fluorescence traces as reference) to ensure that no structures belonging to the dendrite of interest were removed in the process. Data exclusion We attempted to capture 3 imaging fields for each of the 23 animals used in this experiment. In a subset of animals, fewer than three fields were fully captured, resulting in a total field count of 61. Dendrites that presented any signatures of poor health (see “ Evaluation of dendritic health ” above) - such as blebbing, significant bleaching, and/or globally reduced/absent activity - at any point during the experiment were excluded from all analyses. Out of the 61 imaging fields acquired for this data set, 4 fields (6.6%) were excluded due to poor dendritic health. Fields that showed deteriorated optical quality (typically characterized by higher background and lower visibility of the dendrite, as well as cellular debris in the immediate environment of the target dendrite) or had newly visible structures blocking the target dendrite, were excluded if dendritic spines could not be confidently characterized. This resulted in the exclusion of an additional 12 fields, yielding 45 total fields used in this study.

Fluorescence analysis

Fluorescence time series were produced by averaging the pixels within each ROI for all imaging frames. The time-varying baseline (F 0 ) of a fluorescence trace was estimated by smoothing inactive portions of the trace, using a previously described iterative procedure 9 . Briefly, this process identified “active” and “inactive” portions of the trace, removing active portions and using the LOESS-smoothed inactive portions (interpolated across active periods) to estimate the time-varying baseline. The normalized ΔF/F 0 trace was then calculated, where ΔF was found by subtracting the baseline trace from the raw trace, and F 0 is the calculated time-varying baseline. Activity events were detected based on previously described methods 9 . Briefly, noise was estimated for each ΔF/F 0 trace as the standard deviation of negative fluorescence values mirrored about the origin. This noise estimate was then used to set two thresholds, one being 2× the noise to find active portions of the trace, and another being 1× the noise to define the baseline. Active portions of the trace were defined as when the 1-s LOESS-smoothed ΔF/F 0 trace crossed the active threshold and extended backwards to begin when the baseline threshold was crossed by the unsmoothed trace. Binarized traces with the value of 1 for active frames and zero otherwise were then produced for each region of interest. Such binarized traces were used for all co-activity analysis and event frequency calculations. Using this approach, we found that the median event rates for each spine type were as follows: MRSs, early: 5.3 events/min (95% CI = [5.1, 5.6]), late: 5.1 events/min (95% CI = [4.8, 5.4]) (late vs. early MRSs: p = 0.46); nonMRSs, early: 5.0 events/min (95% CI = [4.7, 5.3]), late: 4.4 events/min (95% CI = [4.2, 4.6]) (early vs. late nonMRS: p = 7e-7), and new spines: 4.9 events/min (95% CI = [4.4, 5.8]). For structural analysis, average projection images of the entire motion-corrected time series were produced. To estimate spine volume in each session, the integrated fluorescence intensity of pixels with intensity values above background (the average pixel intensity across the designated background ROI) over a given spine ROI was divided by the average fluorescence intensity of the nearby region of dendrite for normalization. Such normalization should account for global changes in the expression level of the sensor. Local dendritic fluorescence intensity was estimated by using the dendritic ROIs (described above) within 5μm of the base of the spine.

Activity onset analysis

To estimate the timing of the activity onset of individual spines during movements, the fluorescence traces of each spine were aligned to the onsets of movements overlapping with the activity of the spine. We only considered movements that did not have another movement within 1s prior to movement onset and 2s post movement onset to avoid contamination of activity related to other movements. Activity averaged across movements was then used for peak detection, as follows: first, peak activity was defined using the “findpeaks” function (Matlab) in a 2s window starting 1 sec before and ending 1 sec after movement onset, with a minimum distance between peaks set at 0.5s, and a minimum peak height corresponding to the median + standard deviation of the full 3s peri-movement period being inspected. If multiple peaks were found, each was given a score accounting for both the amplitude of the peak as well as the temporal proximity to movement onset ((1/abs(peak timing – movement onset timing) * peak amplitude), and the peak with the highest score was considered as the target peak activity. After peaks were identified, the velocity of the activity trace – defined as the first derivative of the robust-loess-smoothed activity trace (10-frame/~170ms window) – was used to identify changes in the slope of the activity from negative to positive, allowing the identification of the beginning of rising phases of activity. Onset timing was then defined by searching backwards in time from the end of the rising phase of the target peak (defined as 75% of the target peak amplitude) to one of two criteria: either 1) when the activity velocity trace fell below zero, corresponding to the start of the rising phase, or 2) when the smoothed activity fell below the median of the full 3s window of of peri-movement activity.

Spine structural classification

Average projections of time series from each session were registered with respect to the first imaging session for that field (described in “Image Analysis” section above) to allow for comparison across days. A duplicate set of these images were iteratively deconvolved with “Interative Deconvolve” plugin for ImageJ) as a guide for spine detection. We excluded spines that were too close to each other to be accurately separated for fluorescence trace extraction in the original 2-photon image series projection. Any visible dendritic protrusion emanating from the dendrite were considered putative spines. Bright, punctate regions of at least 0.5μm diameter overlapping with the dendrite in the imaging plane were also considered spines. Both assumptions were corroborated with subsequent electron microscopy reconstructions of a subset of imaged dendrites (e.g. Extended Data Fig. 4b , Extended Data 7g rows 2 and 3). Spines that appeared in the same location across sessions, or whose neck originated from the same dendritic region were considered to be the same spine. The spines that were represented across all sessions in this way were considered stable, “pre-existing” spines. Spines that were no longer visible in later imaging sessions (eliminated spines) were only considered for early session analyses, such as the analysis of early MRS density. Dendritic protrusions in later sessions that were not present in previous sessions were considered new spines. New spines that were present only for the “middle” learning sessions were considered “transient” new spines, and were used only for the analyses in Fig 3 , where specifically indicated. Spines that transiently disappeared in the middle session and then reappeared in the late session were rare, and were excluded from analysis. Using the above approach, we found that the overall spine density does not significantly change between early (0.67 spines/μm) and late (0.61 spines/ μm) sessions (p = 0.19, rank-sum test) ( Extended Data Fig. 1d ).

Movement-related classification

Spines were classified as movement-related on each individual session, as previously described 9 . Briefly, the dot product of binarized lever traces (movements vs. non-movements, as detailed above) and continuous ΔF/F 0 traces was calculated for each spine. This value was then compared to the dot products when shuffling the movement periods 10,000 times. The dot products of each of the shuffled traces with ΔF/F 0 traces were then compared to the values of the actual data. Actual values that were above the 97.5th percentile of the shuffled distribution were considered “movement-related”.

Distance analysis

In all analyses regarding the “distance from new spine”, a given spine’s distance value corresponded to the dendritic distance (as determined from single-plane 2p images) from the base of the spine to base of the nearest new spine. Note that this dendritic distance differs from Euclidian distance (direct distance between two points in the plane) in that the curvature of dendrites was considered. For spine density and volume change probability analyses, if multiple new spines were present on a dendrite, only the nearest new spine was considered. The territory of each new spine was bounded at the halfway point between the new spines. For analysis of the density of MRSs surrounding new spines, we focused our analysis on dendrites that showed formation of at least one new spine. For each spine, any MRS that fell within 10 μm from the spine in either direction along the dendrite were considered. The number of MRSs was then divided by the dendritic distance considered (20 μm in most cases, but occasionally less when a new spine is within 10 μm of the edge of a dendrite). Chance estimates of the number of nearby MRSs were performed by randomizing each new spine locations across all dendrites used for analysis in 3b (i.e. those that showed at least one new spine) 10,000 times. Randomized locations were assigned to dendrites with a spatial resolution of 0.5μm and using the full length of the imaged dendrite. For example, a 45μm-long dendrite could have a randomized new spine location at 0μm, 0.5μm, 1μm, 1.5μm, and so on, up to 45μm. As before, if a simulated new spine was within 10 μm of the edge of the simulated dendrite, then only the true total dendritic distance considered was used to calculate MRS density. For functional spine density measurements, the total number of either MRSs or non-MRSs on a single dendrite were counted with respect to the closest new spine, and divided by the total distance measured along the imaged dendrite. Boundaries for each new spine were defined as either the extent of the current distance bin being measured (successive 5μm steps in either direction; see below), the edge of a dendrite, or – in cases where multiple new spines were present on a single dendrite - the halfway point to the nearest, other new spine. Any bins that corresponded to less than 5μm total length (e.g. when multiple new spines were close to the edge of a dendrite) were excluded from analysis. All distance values were binned in 5μm increments to simplify visualization. Bins correspond to dBin n < d ≤ dBin n+1 , such that the 2.5μm bin represents inter-spine distance values from 0–5μm, the 7.5μm bin corresponds to 5 < d ≤ 10μm, the 12.5μm bin to 10 < d ≤ 15μm, and so on.

Co-activity analysis

Co-activity rates between all possible spine pairs in a single field were calculated using binarized event traces (binarization process defined above in “Fluorescence Analysis”) for the entire imaging session. All periods where activity events were present in both spines (i.e. frames for which both binarized activity traces are logical true) were considered co-active periods. A single co-activity event was defined as the entire duration that both spines were continuously co-active. The co-activity rates were calculated as the number of such co-active events per unit time. Co-activity rates were then normalized to the geometric mean of the activity frequencies of both spines. Since the geometric mean is highly correlated with the calculated co-activity rates ( Extended Data Fig. 3e ), normalization by this value allows better comparison between the relative co-activity between spine pairs showing different overall frequencies. All co-activity rates were calculated across the entire trace (i.e. in both movement and non-movement periods, as well as across all trial epochs, including inter-trial intervals and cue periods). Analyses for Fig. 4d , e were performed by removing co-active events from the traces of both constituent spines in each new spine-MRS pair, producing “new spine only” and “MRS only” activity traces. Such traces reflect periods when one spine is active while the other spine in the pair is silent, allowing the differentiation of coherent and desynchronized activity for a given spine pair. The fraction of events of each event “type” (i.e. new spine-only activity, MRS only activity, and co-activity) occurring during movement periods (defined in “Movement Analysis” above) thus represents the specificity of each signal to movement. Simulation of spine activity Spine activity was simulated to provide proof-of-principle data illustrating that the geometric mean of activity event frequencies scales better with co-activity rates than does the arithmetic mean ( Extended Data Fig. 3 ). Binarized event traces were simulated by assigning random activity blocks a value of ‘1’. The durations of these simulated events were assigned by randomly sampling from the durations of the real spine imaging data. Simulated event frequencies ranged from 0–25 events/min. The timing of simulated events was random, with the constraint that no activity block overlaps another. If such overlap occurs, the timing of events is randomized until the criterion was met. Co-activity of all possible simulated spine pairs was defined as elsewhere: whenever two spines’ binarized activity event traces were equal to 1, this was considered a co-activity period.

New spine density analysis

New spine density was calculated as the total number of new spines formed on a given dendrite normalized by the dendritic length. Normalizing the new spine number to the total number of pre-existing spines on the analyzed dendrites (which we found to be tightly correlated to the dendritic length; r = 0.83, p = 1e-35; Pearson’s correlation coefficient) yielded nearly identical results (not shown).

Correlated light and electron microscopy Sample preparation

After completion of the final imaging session, data was analyzed so as to identify dendrites that showed new spine formation. After selection of dendrites of interest, animals were retro-orbitally injected with fluorescent dextran (FITC-Dextran; Sigma-Aldrich 52194) and lectin (Tomato Lectin, DyLight 594; Vector Laboratories DL-1177) dyes to visualize vasculature, with lectin permanently marking the vessels following transcardial perfusion. Mice were then anesthetized with an intraperitoneal injection of ketamine/xylazine and transcardially perfused with a brief flush of Ringer’s solution containing heparin and xylocaine, followed by approximately 80 mL of 0.5% glutaraldehyde / 4% prills paraformaldehyde in 0.15M sodium cacodylate buffer containing 2mM calcium chloride (“caco”). The brain was removed from the cranium and post-fixed for approximately 1 hour on ice in the same fixative. The brain was then manually dissected with a razor blade to allow for horizontal sectioning of the cortex on a vibratome (Leica VT1000S). The vibratome blade was brought as close to the cortex as possible by eye and 150 μm-thick sections were then cut. The area under the imaging window was usually captured in one or two sections. Transmitted light images of the sections were collected with a dissecting scope at low magnification to reveal the vasculature and the images were aligned relative to each other in Photoshop (Adobe). By comparing this map with the images of the vasculature taken prior to perfusion, it was possible to accurately locate the area of interest within a particular vibratome slice. The slice was then stained for one hour on ice with DRAQ5 (Biostatus) diluted 1:1000 in caco. The slice was washed three times in caco. The area of interest was located and confocal volumes collected using 20× and 60× water objectives on an inverted microscope (FluoView, Olympus). The slice was post-fixed overnight at 4°C in caco containing 2.5% glutaraldehyde. The slice was then washed with solutions of caco and caco containing 100mM glycine. The slice was stained with the following series of solutions, thoroughly washing with distilled water after each step: 2% osmium tetroxide / 1.5% potassium ferrocyanide in caco for 1 hour at r.t., 0.5% aq. thiocarbohydrazide for 30 minutes at r.t., 2% aq. osmium tetroxide for 1 hour at r.t., 2% aq. uranyl acetate overnight at 4°C, Walton’s lead solution for 30 min at 60°C. The slice was then dehydrated with the following series of solutions, 10 min each step: 70% ethanol, 90% ethanol, 100% ethanol, 100% ethanol, dry acetone, dry acetone. The slice was placed into 50:50 acetone:Durcupan ACM (Sigma-Aldrich) overnight on a rotator. The Durcupan was made with 11.4g component A, 10g component B, 0.3g component C, and 0.1mL component D. The slice was placed into fresh 100% Durcupan in a vacuum chamber for two consecutive nights and then flat-embedded in Durcupan between two glass slides coated with liquid release agent (Electron Microscopy Sciences), using pieces of Aclar 33C as spacers to prevent crushing tissue. The Durcupan was cured at 60°C for 48–72 hours. The slice was mounted on the end of a small aluminum rod and a low-resolution microCT volume (~2.6 mm pixel size) was collected at 80 kV (Zeiss Versa 510 XRM). The vasculature pattern was used to locate the area of interest in the embedded section. The block was trimmed down to less than 1mm × 1mm in size, mounted to a Serial Block-face EM (SBEM) specimen rivet using conductive silver epoxy (Ted Pella), and left at 60°C overnight. Following trimming of the block using an ultramicrotome, a higher resolution microCT scan was collected (~1 mm pixel size) to allow for precise targeting of the SBEM stage to the ROI. An SBEM volume was collected on either a Zeiss Merlin SEM or Zeiss Gemini 300 SEM equipped with a Gatan 3View and OnPoint backscatter detector system. The SBEM volumes were collected at 2.5 kV EHT with 5 nm XY pixels, 50 nm Z steps, and 1 ms dwell time. Since volumes were often very near the surface of the brain, necessitating imaging of areas of empty resin, focal charge compensation with nitrogen gas was used to eliminate charging artifacts. The “tiltxcorr” program of IMOD 55 was used to generate the final SBEM volume by applying cross-correlation to eliminate any minor jitter between slices. The SBEM and confocal volumes were co-registered with the “Landmark Image Warp” module in Amira (versions 2019/2020), using nuclei and other distinctive features visible in both modalities as landmarks.

EM segmentation

Structures of interest in the EM volumes were manually segmented using IMOD software (version 4.9) 55 running on Cygwin Terminal (version 3.3). Segmentation was performed on binned (5–10 pixels in x-y) images to reduce computer memory load. Target dendrites were first identified using the overlaid confocal-EM images, and the accuracy of the overlaid image was then confirmed by tracing “landmark” spines and dendritic features that were particularly obvious in in vivo images (e.g. large, solitary spines, dendritic curves or bifurcations, etc.). Individual contours were drawn for each spine, dendrite, axon, and subcellular structure (e.g. spine apparati) using the sculpt tool. All visible protrusions from the target dendrite were segmented, irrespective of their visibility in in vivo images. The segmentation of each structure was evaluated by three individuals.

EM synapse classification

For a given spine, the presynaptic axon was identified based on the presence of an apparent synapse between the two structures. Synapses were defined based on 1) the presence of apparent post-synaptic density (PSD), i.e. a darkened band at the membranous edge of (primarily the head of) dendritic spines visible across at least 2–3 sections, 2) the presence of a vesicle-housing axonal bouton immediately opposed to the PSD, 3) the collection of vesicles within the bouton around the putative synaptic site, i.e., directly opposed to the PSD across at least 2–3 slices, and 4) a synaptic cleft, appearing as a small but distinct space between the pre- and post-synaptic membranes. In some instances, multiple axons formed apparent synapses on a single spine. In most of these cases, one of the two axons, when traced, connected primarily to the shaft of other dendrites, suggesting that these instances likely represent mostly inhibitory axons. These instances were also concomitant with less defined PSDs, consistent with inhibitory synapses. We did not have the spatial resolution in EM to categorize the vesicular shape of such inputs, and therefore cannot conclusively identify them as inhibitory. However, all instances of such axons were fully traced, and were considered as viable candidates in our “axon-sharing” analysis. However, we found no instances of such axons connecting with multiple spines on the same dendrite.

EM filopodia classification

Putative filopodia were categorized within EM images as dendritic protrusions that were at least 1μm in length and lacked an apparent synapse (as described above). Filopodia identified in this way were then compared to in vivo 2p image series (deconvolved, average projection images from early, middle, and late learning sessions) to see if they corresponded to spines identified in vivo that might have shrunken or been classified as “eliminated” by shrinking below a detectable threshold. In general, spines classified as eliminated were also absent in EM, and we only encountered one instance of a putative filopodia being in the location of a spine classified as eliminated. We excluded this case from analysis. EM distance calculations Dendritic distances were calculated in EM based on open objects drawn in the center of visible dendritic portions across and within slices. The resulting line segments were then visually inspected to ensure that they roughly corresponded to the center of mass of the dendritic branch along its length. Spine and filopodia locations were logged as the point at which the center of mass of the “neck” of the structure merged with the dendrite. EM spine volume calculation Spine volume calculations in EM ( Extended Data Fig. 4e ) were made in IMOD software by taking the area of contours drawn around individual dendritic spine heads multiplied by the z-step size (typically 4nm) to obtain volume.

Spine apparatus classification

Spine apparati were defined as densely stained (dark), typically laminar structures that invaded the spine head and/or the spine neck and spanned a majority of the slices covered by the spine. Such spine apparati were frequently observed to connect with apparent endoplasmic reticulum-like structures in the parent dendrite, which aided in classification. All spine apparatus classifications were made while blinded to the spine volume change conditions calculated in in vivo images.

Estimation of chance distances between new spines and filopodia

The estimation of the chance level of filopodia appearing nearby new spines was performed as follows: for a given number of filopodia on a particular dendrite, the location of each was randomized, finding the “shuffled” location’s distance to the nearest new spine. This was repeated for all the data 10,000 times, and the median new spine-filopodium distance was calculated for each shuffle. The p-value was defined as the number of such values that were less than or equal to the median value of the real data, as a fraction of the total shuffles. This p-value thus represents the fraction of simulated occasions that the null hypothesis was supported (i.e. that filopodia are not closer to new spines than chance).

Statistics and reproducibility

For all sets of experiments described here, pilot experiments were run to assess the intrinsic statistics of our readouts, as well as associated failure rates. Otherwise, no formal statistical method was used to pre-determine sample sizes. However, our sample sizes are similar to those reported in previous publications. Data were excluded only in cases of poor optical quality in imaging experiments or when cellular health was deemed poor (see “ Evaluation of dendritic health ”, Methods ). Animals in this study were not selected for allocation into experimental groups based on any other pre-requisite features other than general wellbeing. Control experiments presented in this study were performed on randomly selected mice and interleaved with learning-group mice. Investigators were blinded for all EM reconstructions (i.e. the properties of interest of spines were not known during their reconstruction) and ROI drawing for in vivo images (i.e. the properties of interest that might be affected by ROI drawing are not yet known at the time of drawing), but were otherwise not blinded to allocation during experiments and outcome assessment. Non-parametric statistics were used when possible to avoid assumptions of data normality. Multiple comparisons were corrected for using the false discovery rate (FDR) method. All tests performed were two-sided. Sample sizes (n) are as follows: total mice: 23 iGluSnFR; 30 GCaMP6f; 5 iGluSnFR (imaged at 810nm); 14 iGluSnFR (no-task controls); number of fields: 45 iGluSnFR, 66 GCaMP6f; 17 iGluSnFR (imaged at 810nm); 25 iGluSnFR (no-task controls) number of dendrites: 76 iGluSnFR, 137 GCaMP6f; 33 iGluSnFR (imaged at 810nm); 40 iGluSnFR (no-task controls); total number of unique imaged spines (iGluSnFR): 1915 (median 23 spines/dendrite); 484 iGluSnFR (imaged at 810nm); 1787 iGluSnFR (no-task controls); total number of early session spines (iGluSnFR): 1767; total number of middle session spines (iGluSnFR): 1582; total number of late session spines (iGluSnFR): 1656; number of early MRSs: 898; number of middle session MRSs: 836; number of late MRSs: 820. Number of mice showing new spines: 21/23 iGluSnFR, 25/30 GCaMP6f; number of dendrites showing new spines: 50/76 iGluSnFR, 62/137 GCaMP6f; total number of new spines: 118 iGluSnFR; 140 GCaMP6f; 51 iGluSnFR (no-task control) N value in figures regarding new spines only report animals/dendrites that show new spine formation. Summaries of spine numbers, including MRS and new spine counts, can be found in Supplementary Tables 1 (learning-group) and 2 (no-task control group).