Tuft Dendrite Spikes Are Accompanied by Selective Input From Distinct Functional Networks

Abstract

The tuft dendrites of layer 5 (L5) neurons can support regenerative events—dendritic spikes—that have been proposed to coordinate context-specific engagement and plasticity within cortical networks. However, it remains unclear whether tuft spikes are accompanied by input activity with dynamics that could support these network-level functions. To address this, glutamatergic synapses and postsynaptic calcium signals were simultaneously imaged in the tuft dendrites of L5 extratelencephalic neurons within the premotor cortex of mice of either sex performing a cued directional licking task. Trial-to-trial, the generation of tuft spikes was associated with a multiphasic elevation in synaptic activity spanning hundreds of milliseconds. This activity was highly specific to the dendrite in which a spike was detected, suggesting the concurrent activation of select subnetworks. Synapses that were strongly coupled to the overall population were the most synchronized with tuft spikes and preferentially encoded the transition between the preparation and action epochs of the task. Even among these strongly coupled synapses, increases in activity were largely specific to synapses located on the spiking dendrite. Surprisingly, among synapses with the poorest population coupling, a second population of coactive synapses was discovered that was also associated with tuft spikes and functionally selective for task-outcome. These results suggest that tuft spikes may be particularly driven by inputs from neurons that are both embedded in sparse subnetworks and synchronized through coupling to larger-scale functional networks.

Significance Statement

Flexible behavior and learning may depend on interactions between activity in different brain networks and dendritic spikes generated within the neurons that make up the output layer of the neocortex. The results of this study indicate that at the moment of spike generation, spikes in different dendrites are associated with the activation of very specific networks. Yet, across time, the inputs most associated with dendritic spikes are broadly coactive and share selectivity for similar features of behavior. This suggests that dendritic spikes may be particularly driven by the activation of “hub” neurons that coordinate communication between large-scale and small-scale functional networks.

Introduction

Neurons do not simply compute the sum of all their inputs. This is especially apparent for layer 5 (L5) extratelencephalic (ET) cortical neurons, the primary output channel of the cortex. Their apical tuft dendrites elaborate in layer 1, far from the cell soma and basal dendrites. The apical tuft dendrites contain active conductances that can support powerful regenerative events—dendritic spikes—triggered by nonlinear integration of synaptic inputs with or without back-propagating action potentials (bAPs; Larkum et al., 1999; Stuart and Häusser, 2001; Larkum et al., 2009; Stuart and Spruston, 2015). These tuft spikes can trigger large calcium influx, prolonged somatic depolarization, and burst firing (Larkum et al., 1999). Evidence suggests that tuft spikes condition the output of L5 neurons on long-range contextual signals (Xu et al., 2012; Takahashi et al., 2016; Ranganathan et al., 2018; Fişek et al., 2023), and rapidly induce long-lasting functional changes (Bittner et al., 2017; Xiao et al., 2025; Yaeger et al., 2025).

To understand the patterns of synaptic inputs that trigger these sparse events, tuft dendrites have been extensively studied in ex vivo brain tissue. Short-timescale (≈10 ms) co-activation of spatially clustered inputs can be especially effective at engaging dendritic nonlinearities (Polsky et al., 2004; Losonczy and Magee, 2006; Larkum et al., 2009; Branco et al., 2010; Weber et al., 2016), consistent with the idea that small segments of dendrite could operate as independent computational subunits (Mel, 1994). However, in vivo imaging indicates that fully isolated spiking activity within small tuft segments is rare (Svoboda et al., 1997; Hill et al., 2013; Francioni et al., 2019; Kerlin et al., 2019; Otor et al., 2022), suggesting that dispersed input coactive over longer timescales may play a prominent role in tuft computations.

Recent models have focused on a potential role for dendritic spikes in more network-level operations (Bono and Clopath, 2017; Guerguiev et al., 2017; Antic et al., 2018; Poirazi and Papoutsi, 2020; Milstein et al., 2021; Payeur et al., 2021). For example, tuft spikes have been proposed to trigger recurrent activity that coordinates learning across subnetworks that span the cortical hierarchy (Payeur et al., 2021). Other models suggest tuft spikes may enable context-dependent computation by gating cortical hypernetworks (Larkum, 2013; Rao, 2024) or by facilitating synchronous firing of different functional subnetworks engaged in the same task (Antic et al., 2018). All of these models predict that spikes in distal dendrites should be accompanied by subnetwork-specific input activity to that dendrite at both long and short timescales. Yet, little is known about the dynamics of glutamatergic input to the tuft that accompanies dendritic spikes in vivo or its relationship to the functional subnetworks engaged during behavior.

Here, we simultaneously imaged presynaptic and postsynaptic activity in the tuft dendrites of L5 ET neurons in premotor cortex as mice performed a cued directional licking task. We defined synapses as belonging to different networks based on the degree to which their activity was coupled to the net activity of the population (Okun et al., 2015). In local cortical networks, a subset of neurons belong to a strongly coupled (SC) rich-club with dense intrinsic connectivity (Yassin et al., 2010; Gal et al., 2017; Tu et al., 2025; Finkelstein et al., 2026), which may play a unique role in coordinating network dynamics (Griffa and Van Den Heuvel, 2018; Hafizi et al., 2021). For the first time, we identified a similar topology in the activity of the diverse synaptic inputs to tuft dendrites. We found that tuft spikes were embedded in input episodes with short- and long-timescale components that were highly specific to the dendrite that generated a tuft spike. Tuft spikes were particularly associated with the activity of inputs belonging to the SC network or a newly discovered nonpositive coupled network, networks we found encoded the preparation-to-action transition and the task-outcome, respectively. Together, our results suggest that tuft spikes are associated with activation of distinct subnetworks, which are themselves embedded within a larger-scale cortical network topology.

Materials and Methods

Mice and surgeries

All animal procedures were carried out in accordance with protocols approved by the University of Minnesota Animal Care and Use Committee. Twenty adult C57BL/6J mice (Jackson Laboratory; 000664) were used for all experiments (11 male and 9 female). Nineteen mice were task trained, and 1 mouse was used for imaging during spontaneous behavior. Buprenorphine (0.1 mg/kg, IP injection) and Rimadyl (10 mg/kg, subcutaneous) were injected 30 min prior to surgery for analgesia. Surgical procedures were performed under isoflurane anesthesia (3% for induction, 1–1.5% during surgery). A burr hole was drilled above the left ventromedial thalamus (VM; coordinates: −1.5 mm posterior, 0.9 mm lateral relative to bregma). A retrograde adeno-associated viral vector encoding Cre recombinase (pENN-AAVrg-hSyn-Cre-WPRE-hGH 105553; Addgene) was injected into the VM thalamus. The virus was diluted to a final concentration of 1.2 × 10¹² vg/ml, and 200 nl was delivered at a depth of 4,000 μm. A 3 mm diameter craniotomy was made above the left hemisphere of anterior lateral motor (ALM; 2.5 mm anterior and 1.5 mm lateral to bregma). Cre-dependent adeno-associated viral vectors (pAAV9-hSyn-FLEX-iGluSnFR3-v857-SGZ 175182, pAAV1-Syn-Flex-NES-jRGECO1a-WPRE-SV40 100853; Addgene) were injected into ALM. Two mice (n = 12 sessions) were only injected with iGluSnFR3 and were included in some analyses. Three 100 nl injections were delivered in a triangular pattern, with injection sites spaced approximately 150 μm apart, at a depth of 700 μm. Viral vectors were diluted to a final concentration of 1.2 × 10¹³ vg/ml. A window (triple #1 coverglass 2.5/2.5/3.5 mm diameter; Potomac Photonics) was fixed to the skull using dental adhesive (C and B Metabond; Parkell). A metal bar for head fixation was implanted posterior to the window with a metal loop surrounding the window. Expression was monitored for at least 2.5 weeks before imaging. Mice were gradually water restricted and then received water according to task performance (minimum: 0.6 ml daily) or 1 ml on days without training (Guo et al., 2014).

Auditory decision task

Head-restrained mice were trained to perform an auditory-cued directional licking task (Fig. 1A; Guo et al., 2014; Inagaki et al., 2018). The behavioral apparatus was controlled using BPOD (Sanworks). Each trial started with the sample period, during which one of two auditory cues (3 or 18 kHz; 5 tones at 4 Hz) was presented. The sample period was followed by a delay epoch of 1.25 s. Following the delay epoch, an auditory “GO” cue (carrier frequency 6 kHz with 360 Hz frequency modulation, 100 ms duration) was presented, which instructed mice to lick one of two water ports centered just below the mouth (2 mm between ports). Licking was detected with a custom electrical lick detector. Licking the correct water port within 2 s (as instructed by the sample cue) dispensed a small amount of water (3 μl) from that port, whereas licking the incorrect port triggered a 15 s timeout period. Licking during the sample or delay period reset the corresponding period timer. Daily sessions generally lasted 1–1.5 h, with 220–300 randomized trials. Mice required 3–6 weeks to achieve performance greater than 70%. Imaging sessions began after mice maintained performance above 70% for at least 3 consecutive days.

Figure 1.

Simultaneous imaging of synaptic input and dendritic events during motor behavior. A, Illustration of cued directional licking task. Yellow line: period of 2P imaging. B, Illustration of labeling and imaging strategy. C, Example FOV showing simultaneously recorded iGluSnFR3 (green; top) and jRGECO1a (magenta; bottom) expression along with synaptic ROIs (S1, S2) and dendritic ROIs (D1, D2). D, Activity for ROIs in (C) for one example trial. Top: jRGECO1a ΔF/F₀ timecourses of dendrite segments. *: inferred tuft spike onset. Middle: iGluSnFR3 ΔF/F₀ timecourses of synapses (green). Horizontal dashed line shows threshold for Θ (ΔF/F₀) (black). Bottom: iGluSnFR3 dF/dt timecourses (dark green). Horizontal dashed line shows threshold for Θ (dF/dt) (black). E, Example GO active synapse. Top: Θ (ΔF/F₀) traces for individual trials for each trial-type. Bottom: trial-averaged timecourses for each trial condition. Error shading: trial-level bootstrap SEM. F, Same as (E), except a “response” selective synapse. G, Same as (E), except an outcome selective synapse.

In vivo imaging

Imaging experiments were conducted using a custom-built two-photon microscope utilizing a 25× water-immersion objective (NA 1.05, Olympus XLPLN25XWMP2) and a tunable femtosecond laser (1,000 nm; 100 fs; 80 MHz; InsightX3, Spectra Physics). Lateral scanning was performed by a resonant mirror (CRS 12 kHz, Novanta Photonics) conjugated to a set of galvanometer mirrors (6215H 5 mm, Novanta Photonics). For simultaneous two-color imaging (dichroic: T565lpxr, Chroma), iGluSnFR3 emission photons (FF01-531/46, Semrock) were collected using a photomultiplier tube (PMT; PMT2101, Thorlabs) and jRGECO1a emission photons (FF01-625/90, Semrock) were collected with a multi-pixel photon counter (C14455-0276, Hamamatsu). Frames were acquired at ∼175 Hz over a 140 μm × 35 μm imaging area, digitized (NI-5771/NI-7975R/PXIe-1092, National Instruments), and recorded with ScanImage (MBF Biosciences). For spontaneous activity imaging (Fig. S3E), two sessions were collected at 325 Hz with a 140 μm × 17.5 μm field of view (FOV), and a third session was collected at ∼90 Hz with a 280 μm × 140 μm FOV. Imaging was restricted to apical dendrites located <0.1 mm below the pial surface. To mitigate rapid photobleaching, imaging was limited to a 2.5 s window per trial, beginning at the start of the delay period and ending 1.25 s into the response period. A new FOV was imaged each day.

Rigid registration

Image registration methods were as previously described (Scheib et al., 2026). In brief, all image series were subjected to an initial rigid registration (phase cross-correlation; no upsampling; python scikit-image). Offsets were estimated from spatially filtered (Gaussian σ ≤ 1 μm) frames, adaptively filtered to suppress registration noise, and then used to transform unfiltered frames.

Intra-trial bleach correction

To correct for the fast trial-by-trial photobleaching component (Fig. S1A,B), the whole-frame fluorescence time series of each emission channel was averaged across trials. A double-exponential decay was then fit to the mean time series. For some sessions, frames (0–0.4 s) after the GO cue were excluded from the fit to mitigate influence from large transients of activity that occurred at that time. The fit was then reciprocally normalized by the last frame in the fit as:

$$F_{\mathrm{recip}\mathrm{\_}\mathrm{norm}}(t)=\frac{F_{\mathrm{last}}}{F(t)},(1)$$

where t denotes time (or frame index), F(t) is the fitted double-exponential photobleaching curve evaluated at time t, F_last is the value of the fitted photobleaching curve at the final frame included in the fit, and $F_{\mathrm{recip}\mathrm{\_}\mathrm{norm}}(t)$ is the reciprocal normalization factor used to correct fluorescence signals for photobleaching. Then it was multiplied element-wise by the activity traces of each trial (Fig. S1B).

Denoising

Following rigid registration, image series were denoised using DeepInterpolation (Lecoq et al., 2021), modified to minimize L2 (rather than L1) loss. Denoising was performed separately on glutamate and Ca²⁺ activity channels. For interpolation of jRGECO1a frames, the 3 preceding and 1 following frames were provided for training and inference. For interpolation of iGluSnFR3 frames, 6 mean projections of frames collected from non-overlapping exponential (base 1.26; exponents: 0–5) frame windows both preceding and following the interpolated frame were provided for training and inference. Following denoising, the first and last 12 frames from each trial were excluded to remove artifacts of interpolation near the edges of the 2.5 s imaging windows.

Non-rigid registration

Small within-session instabilities (e.g., resonant mirror temperature changes and vessel dilation) resulted in image warp that was significant for dendritic imaging. Image warp was corrected as previously described (Scheib et al., 2026). In brief, warp calculations were performed on contrast-enhanced frames using symmetric diffeomorphic registration (python DIPY; Avants et al., 2008; Garyfallidis et al., 2014), adjusting the spatial scale of the warp correction on a session-by-session basis. To mitigate the noise of frame-to-frame warp calculations, x- and y-warp fields were adaptively filtered in time.

Activity extraction

Dendrite regions of interest (ROIs) were manually segmented from the mean projection image of each session. Active synapses were manually segmented from the variance projection image of each session and assigned to dendrite segments based on the mean projection image. Only dendrites and synapses within regions of the imaging FOV where labeling was sufficiently sparse to permit clear assignment of synapses to dendrite segments were segmented. Mean fluorescence within each ROI was divided by a rolling baseline (F₀; averaging window: ±7 trials; synapse activity exclusion: z-score >2; dendrite activity exclusion: z-score >3) and subtracted by 1 to obtain ΔF/F₀. Synapse signal-to-noise ratio (SNR) was estimated as the value of the 99th percentile of positive ΔF/F₀ deflections divided by the 99th percentile of the absolute value of negative ΔF/F₀ deflections. Synapses with SNR <1.5 were excluded.

Two complementary transformations of iGluSnFR3 activity were used to assess synaptic activity. One was derived from the conventional ΔF/F₀ and the other was derived from the smoothed dF/dt (Gaussian filter σ = 2 with a radius of 4 frames). Both measures were binarized by a shifted Heaviside step function $\mathrm{\Theta }(x-x_{\mathrm{neg}\mathrm{\%}}$), where x is the continuous measure (ΔF/F₀ or dF/dt) and $x_{\mathrm{neg}\mathrm{\%}}$ is the 99th (for ΔF/F₀ ) or 92nd percentile (for dF/dt) of the absolute value of negative deflections x. To obtain the trial-specific activity (TSA) of a synapse, the trial-averaged activity for each trial-type (CL, CR, IR, and IL) was subtracted from activity on corresponding trials (Fig. 2G).

Figure 2.

Tuft spikes are associated with temporally extended, dendrite-specific increases in input. A, Trial-averaged spike dendrite (SD) Ca²⁺ transients aligned to inferred spike onset for individual dendrites. B, Mean population SD Ca²⁺ transient for dendrites in (A) (solid black; n = 313 dendrites) or other dendrite (OD) Ca²⁺ transient (dashed gray; n = 509 dendrites). C, Pairwise dendritic Ca²⁺ correlations (n = 5,078 pairs) with gray line at 0.4 correlation. D, Trial-averaged SD Θ (ΔF/F₀) of individual synapses. E, Population mean Θ (ΔF/F₀) of SD synaptic activity (black solid line; n = 826 synapses) and OD synaptic activity (dashed gray; n = 1,922 synapses) aligned to tuft spikes. Green line: fit of SD to exponential rise and decay (R² = 0.92). F, Mean of pre-onset (−575 to 0 ms) and post-onset (0 to 575 ms) activity in (E). G, Illustration of the calculation of trial-specific activity (TSA, see Methods). H–J, Same as (D–F), but for Θ (ΔF/F₀)_TSA. K, Mean normalized autocorrelation of Θ (ΔF/F₀)_TSA across synapses (black; n = 2,072 synapses) and exponential decay fit (green; R² = 0.98). L–O, Same as (H–J) for Θ (dF/dt)_TSA activity. Fit in (M): R² = 0.81. Fit in (O): R² = 0.98. Error shading and error bars are session-level hierarchical bootstrap SEM (see statistics in Table 1).

Tuft spikes were detected in the jRGECO1a dendrite segment ΔF/F₀ timecourses using a custom algorithm. Transients above the 99.9th percentile of negative ΔF/F₀ deflections for 15 consecutive frames were identified as potential spikes. Subsequent to the first transient identified in a trial, transients were only considered as potential spikes if activity dropped back below threshold for at least 25 frames prior to the transient. To identify the onset time for each potential spike, the derivative of a 13-frame moving average of ΔF/F₀ was convolved with an odd-symmetric slope kernel [2, 2, 2, 1, 1, 1, 0, − 1, − 1, − 1, − 2, − 2, − 2] and then smoothed again with a 13-frame moving average. Within the resulting timecourse, the maximum value within the 30 frames preceding the first frame of the transient was defined as the onset time. Only spikes with a relative ΔF/F₀ increase of at least 0.3, measured as the difference between the onset frame ΔF/F₀ and the average ΔF/F₀ across frames 12–25 after onset, were retained.

Pairwise correlation and autocorrelation analyses

Pairwise correlation of dendrite segment Ca²⁺ activity was determined using floored ΔF/F₀ timecourses (Fig. 2C). Values below the 99.9th percentile of the absolute value of negative ΔF/F₀ deflections were zeroed, and values above were left unchanged. Floored traces were used to ensure that small motion artifacts did not dominate the correlation estimate, as most dendrites produced few transients per session with a unimodal amplitude distribution (Fig. S1C,D). Pairwise correlation of synaptic activity was based on session-wide Θ (dF/dt)_TSA timecourses for all analyses (Fig. 3D). Synapse pairs within 2.5 μm in Euclidean distance were excluded to avoid contamination of input signals by potential glutamate spillover. Synapse pairs with a correlation >0.5 were also excluded to remove synapses potentially originating from the same axon. In all cases, Pearson’s correlation coefficient was used as the correlation metric.

Figure 3.

Strongly coupled synapses are particularly active at tuft spike onset. A, Illustration of the synapse population used to calculate the coupling of one synapse (bright green). B, Example snippet of Θ (dF/dt)_TSA from a synapse and corresponding synapse population sum Θ (dF/dt)_TSA. C, Distribution of population coupling values across all synapses (black; n = 2,072 synapses) and distribution after shuffling time (gray). D, Mean pairwise correlation (ρ) calculated from Θ (dF/dt)_TSA of synapse pairs within the same bin (black circles; n = 9,561 synapse pairs) or with synapses in other bins (gray triangles; n = 71,212 synapse pairs) binned by population coupling rank (statistics in Table 3). E, Θ (dF/dt)_TSA synaptic SD activity (solid lines) and OD activity (dashed lines) aligned to tuft spikes for synapses grouped by coupling strength (n: SC_SD = 144; SC_OD = 382; WC_SD = 522; WC_OD = 1,156; NC_SD = 160; NC_OD = 384). F, Mean activity in gray windows (−50 to 150 ms from spike onset) shown in (E). Significance tests: two-tailed hierarchical difference test (statistics in Table 6). Error shading and error bars: session-level hierarchical bootstrap SEM. All panels include only data with dual color imaging.

The autocorrelation of individual synapses was calculated using Pearson’s correlation coefficient at increasing time lags of Θ (dF/dt)_TSA or Θ (ΔF/F₀)_TSA, as noted (Fig. 2K,O). The autocorrelation decay constant was calculated from a single-exponential fit (with intercept) to the mean of the autocorrelation traces of all synapses. It was estimated by iteratively fitting the model to resampled group-level means (see Statistics section for details on bootstrapping procedures). The zero-lag autocorrelation value was excluded from this fit to mitigate distortion of the estimate by photon shot noise.

Characterization of synaptic activity associated with tuft spikes

Synaptic activity was aligned to the inferred onset of tuft spikes and averaged for each synapse before calculating group-level summary statistics. Only synapses on dendrite segments where at least 4 tuft spikes were detected across the imaging session were included in this analysis. For each spike, synapses located on the dendrite in which the tuft spike was detected were considered part of the SD group. The OD group consisted of simultaneously imaged synapses detected on other (defined by session-wide calcium activity correlation <0.4) dendrite segments (Fig. 2C). Putative multi-branch tuft spikes were defined as events where another dendritic segment in the FOV had a session-wide calcium activity correlation >0.4 and a same-trial correlation >0.7 with the SD (Fig. S1K,L).

Synaptic dynamics associated with tuft spikes were quantified using a composite model comprising a single-exponential rise and a single-exponential decay with a joint intercept. The fit of the composite model was estimated by iteratively fitting the model to resampled group-level means (see Statistics section for details on bootstrapping procedures). For each bootstrap sample, the transition point between the rise and decay phases was defined as the time of the maximum value within 0–400 ms of tuft spike onset.

Population coupling

Population coupling estimation used Θ (dF/dt)_TSA timecourses, with the exception of estimation from spontaneous activity (Fig. S3E), which used ΔF/F₀. Population coupling was quantified by computing the Pearson correlation coefficient between a synapse and the summed activity of all other simultaneously imaged synapses, excluding those located within 10 μm on the same branch (Fig. 3A,B). Synapses located nearby were excluded to avoid functional sampling biases. For the calculation of population coupling rank, all synapses across all sessions were pooled.

Activity modes and coding directions

The GO mode is a dimension in the space of synaptic activity that best captures the lick-direction-invariant change in activity pattern following the GO cue, which contributes to transitioning population activity from a motor-preparatory to an action-ready state (Inagaki et al., 2022b). The Response Direction is a direction in activity space that maximally distinguishes between activity on CR and CL trials (Economo et al., 2018; Inagaki et al., 2022b). The Outcome Direction is a direction in activity space that maximally distinguishes between activity on correct and incorrect trials (Yang et al., 2022; Chen et al., 2024). A weight for each synapse for each mode or direction was calculated as follows: GO mode:

$$\frac{{\overline{\mathrm{CR}}}_{0-0.25\mathrm{s}}+{\overline{\mathrm{CL}}}_{0-0.25\mathrm{s}}}{2}-\frac{{\overline{\mathrm{CR}}}_{-0.25-0\mathrm{s}}+{\overline{\mathrm{CL}}}_{-0.25-0\mathrm{s}}}{2}.(2)$$

Response Direction:

$${\overline{\mathrm{CR}}}_{0-0.6\text{ s}}-{\overline{\mathrm{CL}}}_{0-0.6\text{ s}}.(3)$$

Outcome Direction:

$$\frac{{\overline{\mathrm{CR}}}_{0.6-1.2\mathrm{s}}+{\overline{\mathrm{CL}}}_{0.6-1.2\mathrm{s}}}{2}-\frac{{\overline{\mathrm{IR}}}_{0.6-1.2\mathrm{s}}+{\overline{\mathrm{IL}}}_{0.6-1.2\mathrm{s}}}{2},(4)$$

where ${\overline{X}}_{a-b\mathrm{s}}$ denotes the trial-averaged synaptic activity of trial-type X over the time interval from a to b seconds relative to the GO cue. The GO mode was orthogonalized to the Response Direction using a Gram–Schmidt process to isolate condition-invariant activity. The Outcome Direction was orthogonalized to the GO mode and then to the Response Direction using a Gram–Schmidt process.

Analysis of synapse clustering by function

Pairwise noise correlation as a function of distance analysis was calculated using Θ (dF/dt)_TSA timecourses (Fig. S2A). Signal correlation was calculated using trial-averaged Θ (dF/dt) time courses of each trial-type concatenated together (Fig. S2B). Distance was measured along the centerline of dendritic segments for pairs located on the same dendrite. Euclidean distance was used for synapses located on different dendrites. The spatial structure of pairwise correlations was quantified by fitting a single-exponential decay model. The model was iteratively fit to resampled group-level means (see Statistics section for details on bootstrapping procedures).

To determine whether SC synapses were preferentially clustered on the same dendritic branch (Fig. S2C), for each SC synapse, we evaluated pairwise relationships with other synapses located either on the same dendrite or on different dendrites. Pairs in which both synapses were classified as strong couplers were assigned a value of 1, whereas pairs containing a weak or nonpositive coupler were assigned a value of 0. These values were then averaged separately across same dendrite and different dendrite pairs for each reference strong coupler synapse. To avoid biases arising from unequal sampling, we balanced the numbers of same dendrite and different dendrite pairs for each reference synapse prior to averaging.

Secondary analysis of electrophysiology dataset

We reanalyzed publicly available electrophysiology DANDI data that measured responses of single neurons in mouse ALM during a similar behavioral task (Fig. S4; Li et al., 2015a, 2015b). Detailed methods are included with the data. In brief, mice engaged in a similar delayed directional licking task, except that the sensory cue was tactile. Neural activity in ALM cortex was recorded with NeuroNexus silicon probes and spike sorted.

Our analysis included eight Sim1_KJ18-Cre mice and three Tlx_PL56-Cre mice from this study. Only high-quality units, as labeled in the original dataset, were included. To obtain more reliable coupling estimates, we also included only sessions that contained at least 18 units with a session-wide firing rate >0.5 Hz. Trials with photostimulation were excluded. Spike trains were downsampled by counting spikes in 5 ms bins and smoothed with a Gaussian filter (σ = 2 bins; 20 ms radius).

Population coupling for neurons was computed using procedures closely matching those described for the synaptic data (Fig. S4B). TSA was calculated as described for synaptic imaging data. The population coupling of each neuron was estimated session-wide from activity recorded between the tactile stimulus onset and 2 s after the GO signal.

Brain motion exclusion analysis

For some control analyses (Figs. S1H,I, S3D), we blanked inferred large motion periods to establish that dynamics we observed were not explained by motion artifacts. Large motion periods were inferred from the x and y registration corrections since animal movements are expected to be correlated in multiple directions. For each frame, the absolute x and y registration corrections were summed. To account for slow drift, a rolling movement metric (ΔM/M₀) was computed by averaging motion from the current trial together with the seven preceding and seven subsequent trials. For edge trials, the maximum number of available neighboring trials in the constrained direction was used. The resulting ΔM/M₀ trace was smoothed with a Gaussian filter (σ = 2, radius =5 frames) prior to computing the first derivative. The absolute value of this derivative served as the final motion estimate. Motion traces from all experimental sessions were concatenated to establish a uniform threshold (>85th percentile) based on the highest-motion periods across the full dataset (Fig. S1H). This prevented over-penalization of low-motion sessions and under-penalization of high-motion sessions.

Statistics

Group-level summary statistics were computed using nonparametric bootstrap procedures to avoid assumptions of normality. Group means, error bars, and mean differences were estimated using hierarchical bootstrapping. For each bootstrap resample, imaging sessions were sampled with replacement until the original number of observations was drawn. Observations were subsequently averaged at the lower hierarchical level specified in the corresponding figure or text. Unless otherwise noted, 10,000 valid (i.e., non-NaN) resamples were conducted for all analyses. For activity timecourses, the mean of each resample was lightly smoothed using a Gaussian filter (σ = 2, radius = 4 frames) or binned by averaging every four frames (for coupling group timecourses).

For primary pairwise coupling and correlation analyses (Fig. 3D), synapse pairs were generated by resampling individual synapses 100× without replacement (drawing unique, non-overlapping subsets so each synapse contributed to at most one pair within a resample). These estimates were averaged to obtain a single value per bootstrap sample and this procedure was repeated for 1,000 bootstrap iterations that resampled imaging sessions with replacement.

For mean difference tests, group comparisons were performed directly on the bootstrap difference distributions. For each comparison between groups i and j, bootstrap means were paired by iteration and the difference distribution was computed as d_k = a_k − b_k, where a_k and b_k denote the bootstrap estimates from iteration k for the two groups. To avoid p-values of exactly zero due to finite bootstrap sampling, each tail probability was computed with the add-one correction (Davison and Hinkley, 1997) as (c + 1)/(n + 1), where c is the number of bootstrap differences in the relevant tail and n is the number of valid (i.e., non-NaN) bootstrap differences. One-tailed tests evaluated the a priori hypothesis that group i exceeded group j by calculating the proportion of bootstrap differences less than or equal to zero, p(d ≤ 0). Two-tailed tests were computed as $2\times min(p(d\le 0),p(d\ge 0))$, with values capped at 1.0. When multiple pairwise comparisons were performed, raw bootstrap p-values were corrected for multiple testing using the Holm procedure, and corrected p-values below α = 0.05 were considered statistically significant. For relevant figures, asterisks indicate statistical significance based on the Holm-corrected p-value (*p < 0.05, **p < 0.01, and ***p < 0.001).

Code and data availability

Code and data used in the generation of this manuscript are available at https://github.com/kerlin-lab/Gable_2026.

Results

Tuft spikes are accompanied by dendrite-specific, temporally extended input episodes

To measure relationships between excitatory input, tuft spikes, and cortical dynamics, we imaged presynaptic and postsynaptic activity in the ALM cortex of mice performing a delayed directional licking task (Fig. 1A, Guo et al., 2014). During this behavior, task-driven activity patterns in ALM exhibit well-described periods of both extended stability and rapid transition (Li et al., 2016; Inagaki et al., 2022a), making it an especially suitable region to evaluate the relationship between tuft spikes and input dynamics. To precisely measure the timing of presynaptic glutamatergic input, we took advantage of a recently developed genetically encoded glutamate sensor with an affinity well-suited to report fast synaptic glutamate transients (iGluSnFR3, Aggarwal et al., 2023, 2026). To selectively label L5B ET neurons, we injected a retrograde AAV-Cre virus into motor thalamus, as well as Cre-dependent AAVs encoding iGluSnFR3 (a green indicator) and jRGECO1a (a red calcium indicator) into ALM (Fig. 1B). We focused on L5B ET neurons because they have prominent apical tuft dendrites (Landry et al., 1984; Oswald et al., 2013) and because restricting labeling to these neurons ensured that any dendrite-specific synaptic activity we observed did not simply reflect activity from the tuft dendrites of neurons belonging to one sublamina over another.

In mice already expert at the task, we conducted daily high-speed (175 Hz frame rate) imaging of the dendrites in L1 (Fig. 1B,C) during the performance of hundreds of behavioral trials (median = 235; range = 162–302), selecting a new FOV for imaging each session (N = 19 mice, 64 FOV, 611 dendrite segments, 2,591 synapses). Two mice (n = 12 FOVs) only expressed iGluSnFR3 and were included in some analyses. Because iGluSnFR3 undergoes rapid photobleaching (Aggarwal et al., 2023), we restricted imaging to a 2.5 s window spanning the beginning of the delay epoch to 1.25 s after the GO cue (Fig. 1A). Image series were corrected for brain motion, denoised, and manually segmented (Fig. 1C; see Methods). Extracted fluorescence timecourses were then corrected for the effects of bleaching (Fig. S1A,B; see Methods). We analyzed synaptic signals using two different binarized (noise-shifted Heaviside function Θ, see Methods) transformations of the iGluSnFR3 fluorescence. The first measure was derived from the conventional measurement of change from baseline [Θ (ΔF/F₀)] that presumably reflected cumulative glutamate release as convolved by iGluSnFR3 dynamics. The second measure was derived from the positive derivative of fluorescence [Θ (dF/dt)] that presumably reflected rapid increases in glutamate release after periods of relative quiescence (Fig. 1D). The trial-averaged activity of individual synapses (Fig. 1E–G) exhibited selectivity for different task-variables similar to those previously reported for neurons in ALM cortex (Inagaki et al., 2022a; Yang et al., 2022), including synapses consistently driven by the GO cue (Fig. 1E), selective for lick direction (Fig. 1F), and selective for task-outcome (Fig. 1G).

To align synaptic activity with the occurrence of tuft spikes, we inferred the timing of tuft spikes from the onset of transient increases in the jRGECO1a fluorescence within the dendritic shaft (Figs. 1D, 2A,B; see Methods). For the purposes of this study, the term “tuft spike” refers to any regenerative activity within the tuft dendrites, and is therefore agnostic to the spatial extent, engaged conductances, or genesis of the activity. Thus, the tuft spikes we measure may include selectively amplified bAPs, as recently characterized in the distal dendrites of hippocampal neurons through in vivo voltage imaging (Lee et al., 2026; Wu et al., 2026). Dendrite segments were much larger than individual synapses (median = 18 μm; range = 4–105 μm) and our inference algorithm only considered large jRGECO1a transients (>30% ΔF/F₀) as potential spikes, so it is unlikely that the activity of individual synapses or subthreshold cooperativity of a few synapses were mistakenly identified as regenerative events. Consistent with this, tuft spikes were detected at rates ≈100 × lower than synaptic activity (Fig. S1C) with a unimodal amplitude distribution (Fig. S1D), also suggesting that most bAPs were not detected as tuft spikes. Dendrite segments putatively originating from two different neurons were identified by a pairwise calcium session correlation of <0.4. This threshold was the approximate antimode within the bimodal distribution of correlations between segments (Fig. 2C), and previous work has found that even the most independently active compartments of the tuft dendrite exhibit average calcium signal correlations greater than 0.4 (Otor et al., 2022).

The timecourse of mean synaptic activity [Θ (ΔF/F₀)] aligned to tuft spikes was highly variable across individual synapses (Fig. 2D). However, averaging across all synapses revealed a clear peri-spike elevation in synaptic activity (Fig. 2E). Compared to simultaneously recorded dendrites of putative other neurons—henceforth simply referred to as “other” dendrites (OD)—mean synaptic activity on the spiking dendrite (SD) was significantly elevated both before (MD = 2.5 Hz; p = 0.002) and after (MD = 5.2 Hz; p = 6.0 × 10⁻⁴) the spike (Fig. 2F; Table 1). These dynamics could simply reflect the relative timing of trial-aligned activity in dendrites and their respective synaptic inputs. To determine if trial-to-trial fluctuations in synaptic activity were associated with tuft spikes, the mean trial-aligned activity of each synapse was subtracted from its timecourses before alignment with tuft spikes (Fig. 2G). Spike-associated increases in this TSA, Θ (ΔF/F₀)_TSA, were at least 5 × larger on the spiking dendrite than others, both before (MD = 1.9 Hz; p = 6.0 × 10⁻⁴) and after (MD = 3.8 Hz; p = 6.0 × 10⁻⁴) the spike (Fig. 2H–J; Table 1). Restricting analysis to spikes with larger peak ΔF/F₀ (Fig. S1E–G; Table 2) or excluding frames with large lateral brain motion (Fig. S1H–J; Table 2) did not significantly change the spike-associated synaptic activity timecourses. Mean spike-associated Θ (ΔF/F₀)_TSA dynamics were well-fitted by an exponential rise (176 ± 79 ms, half-width at half-max; HWHM) and decay (278 ± 79 ms, HWHM) with a peak that lagged (>101 ± 38 ms) the tuft spike (Fig. 2I). These dynamics were substantially longer than the reported iGluSnFR3 kinetics (rise constant <2 ms; decay constant = 29 ms; Aggarwal et al., 2026) and the mean autocorrelation time constant of the Θ (ΔF/F₀)_TSA signal (Fig. 2K), indicating that tuft spikes were indeed associated with a dendrite-specific elevation in glutamatergic input activity that extended hundreds of milliseconds before and after the spike.

Table 1.

Two-tailed hierarchical bootstrap mean difference tests for population synaptic activity aligned to tuft spikes (Fig. 2)

Comparison	Mean difference (Hz)	p Value
Θ (ΔF/F0) (Fig. 2F)
SDpre versus ODpre	2.5	0.002
SDpost versus ODpost	5.2	6.0 × 10−4
SDpost versus SDpre	2.2	0.009
Θ (ΔF/F0)TSA (Fig. 2J)
SDpre versus ODpre	1.87	6.0 × 10−4
SDpost versus ODpost	3.75	6.0 × 10−4
SDpost versus SDpre	1.72	0.003
Θ (dF/dt)TSA (Fig. 2N)
SDpre versus ODpre	1.15	6.0 × 10−4
SDpost versus ODpost	1.33	6.0 × 10−4
SDpre versus SDpost	0.08	0.75
Male Θ (dF/dt)TSA
SDpre versus ODpre	1.4	6.0 × 10−4
SDpost versus ODpost	1.5	8.0 × 10−4
SDpre versus SDpost	0.31	0.32
Female Θ (dF/dt)TSA
SDpre versus ODpre	0.753	0.005
SDpost versus ODpost	1.11	0.01
SDpre versus SDpost	−0.25	0.6

p Values for each of the three measures separately corrected using Holm Method. SD n = 826 synapses. OD n = 1,922 synapses. SD_male n = 482. SD_female n = 344. OD_male n = 1,139. OD_female n = 783.

Table 2.

Two-tailed hierarchical bootstrap mean difference tests for tuft spike-associated synaptic activity control analyses (Fig. S1)

Comparison	Mean difference (Hz)	p value
Event-threshold comparison (Fig. S1G)
Low thresh.pre vs high thresh.pre	1.0	0.3
Low thresh.post vs high thresh.post	0.01	1
Motion-removal control (Fig. S1J)
Controlpre vs motion removedpre	0.2	1
Controlpost vs motion removedpost	0.08	1
Multi-branch tuft spikes (Fig. S1L)
SDpre vs ODpre	0.9	0.16
SDpost vs ODpost	2.0	0.004

p values separately corrected using Holm method for each group of hypotheses (tuft spike event threshold, motion control, multi-branch spikes). Low threshold synapse n = 826. High threshold synapse n = 333. Motion (both groups) synapse n = 826. Multi-branch: SD synapse n = 485; OD synapse n = 1,332.

Abrupt increases in excitatory input are generally more effective at triggering regenerative activity than gradual increases in input (Mainen and Sejnowski, 1995; Gasparini and Magee, 2006; Losonczy and Magee, 2006). To determine whether abrupt increases in synaptic activity from trial-to-trial were associated with tuft spikes, we calculated TSA of the binarized positive derivative, Θ (dF/dt)_TSA. The spike-associated increase in Θ (dF/dt)_TSA exhibited a more rapid rise (68 ± 49 ms, HWHM) and decay (71 ± 26 ms, HWHM) than the spike-associated increase in Θ (ΔF/F₀)_TSA, with a peak that was closer (32 ± 13 ms) to spike onset (Fig. 2L–O). This increase was specific to the spiking dendrite (Fig. 2N; Table 1), both before (MD = 1.1 Hz; p = 6.0 × 10⁻⁴) and after (MD = 1.3 Hz; p = 6.0 × 10⁻⁴) spike onset. Furthermore, we confirmed that these results generalized to both male and female mice (Table 1). Thus, in addition to being associated with an extended elevation of dendrite-specific input activity, tuft spikes were tightly associated (within ≈ 140 ms) with trial-to-trial abrupt increases in input activity that were highly dendrite-specific.

The high dendrite-specificity of the spike-associated synaptic activity could reflect synchronous activity among adjacent inputs directly driving local dendritic spikes. Indeed, pairwise trial-to-trial and task-driven correlations between synapses did decline as a function of distance (Fig. S2A,B; Wilson et al., 2016; Kerlin et al., 2019). However, most calcium transients detected in the dendritic shaft of the tuft dendrites extend across large regions of the tuft (Francioni et al., 2019; Kerlin et al., 2019; Otor et al., 2022). When we restricted our analysis to dendrites with multiple branches in the FOV and multi-branch calcium transients (see Methods), the synaptic activity associated with these confirmed multi-branch events was still dendrite-specific (Fig. S1K,L; Table 2). This suggests that a significant portion of the activity associated with tuft spikes may be selective for the broader tuft dendrites of particular postsynaptic neurons, rather than solely selective at the level of small dendritic segments.

In summary, tuft spikes were associated with both short-timescale and long-timescale increases in glutamatergic synaptic activity on the spiking tuft, while synaptic activity changed minimally on the tuft dendrites of other neurons belonging to the same L5B subclass.

Inputs SC to the population are the most active at tuft spike onset

After establishing that an abrupt increase in input was associated with the precise timing of large-scale tuft spikes, we sought to identify functionally-defined networks that might contribute to synchronizing this input. The activity of some cortical neurons is SC to the activity of the population (Okun et al., 2015) and these neurons are densely connected to each other, forming a “rich-club” topology (Finkelstein et al., 2026). We hypothesized that the activity of highly coupled synapses—synchronized by their dense connectivity—might be particularly associated with the generation of tuft spikes.

We calculated the population coupling of each synapse as the correlation between its Θ (dF/dt)_TSA and the net Θ (dF/dt)_TSA of all other synapses within the FOV, excluding synapses on the same dendrite within 10 μm (Fig. 3A,B) to minimize any sampling bias introduced by distance-dependent pairwise correlations (Fig. S2A,B). Similar to the distribution of coupling among local cortical neurons (Okun et al., 2015), the coupling of synapses to the L5 ET tuft input population was centered near zero, with a long tail of highly coupled synapses (Fig. 3C). Synapses from all sessions were then pooled and assigned percentile ranks based on coupling strength. This coupling rank estimate was robust to various potential technical artifacts, including synapses per FOV, signal amplitude, and brain motion (Fig. S3A–D). We next asked whether synapses with similar coupling strengths were preferentially coactive. As found in local cortical populations (Okun et al., 2015), highly coupled synapses were more correlated with each other than with other synapses (Fig. 3D; Table 3). Surprisingly, synapses with the lowest coupling also exhibited a higher correlation among themselves (Fig. 3D; Table 3). These enhanced correlations among the lowest couplers were also found in continuous imaging of spontaneous synaptic activity (Fig. S3E; Table 4) and publicly available silicon probe recordings of ALM activity (Fig. S4; Table 5; Li et al., 2015a), indicating that this may be a general feature of cortical activity rather than specific to our task, labeled synaptic population, or recording method. Thus, our data suggest the existence of two large-scale functional networks composed of synapses at opposite extremes of the distribution of population coupling. To pool synapses into groups with respect to this organization, we categorized synapses based on their coupling rank as SC (top 20%), weakly coupled (WC; middle 60%), or nonpositive coupled (NC; bottom 20%). The shorthand identifier “nonpositive” was adopted since this group contained synapses with a population coupling that was either negative or statistically indistinguishable from zero.

Table 3.

One-tailed hierarchical bootstrap mean difference tests for same and different coupling range pairwise noise correlations (Fig. 3D)

Bin	Mean difference (ρ)	p Value
0.00–14.29%	0.02	7.0 × 10−3
14.29–28.57%	0.002	0.11
28.57–42.86%	0.0007	0.24
42.86–57.14%	0.002	0.11
57.14–71.43%	0.01	7.0 × 10−3
71.43–85.71%	0.02	7.0 × 10−3
85.71–100.00%	0.06	7.0 × 10−3

p Values corrected using Holm method. Same coupling range n = 9,561 pairs. Different coupling range n = 71,212 pairs.

Table 4.

One-tailed hierarchical bootstrap mean difference tests for same and different coupling range pairwise correlations during spontaneous behavior (Fig. S3E)

Bin	Mean difference (ρ)	p value
0.00%–14.29%	0.1	7.0 × 10−4
14.29%–28.57%	−0.005	1
28.57%–42.86%	−0.006	1
42.86%–57.14%	0.0007	1
57.14%–71.43%	0.05	7.0 × 10−4
71.43%–85.71%	0.05	7.0 × 10−4
85.71%–100.00%	0.1	7.0 × 10−4

p values corrected using Holm method. Mice n = 1. Session n = 3. Same coupling range n = 1,965 pairs. Different coupling range n = 22,424 pairs.

Table 5.

One-tailed hierarchical bootstrap mean difference tests for same and different coupling range pairwise somatic spiking correlations in ALM (Fig. S4B)

Bin	Mean difference (ρ)	p value
0.00%–14.29%	0.06	7.0 × 10−4
14.29%–28.57%	0.004	0.12
28.57%–42.86%	0.001	0.61
42.86%–57.14%	−0.002	0.76
57.14%–71.43%	0.02	7.0 × 10−4
71.43%–85.71%	0.06	7.0 × 10−4
85.71%–100.00%	0.1	7.0 × 10−4

p values corrected using Holm method. Mice n = 11. Session n = 20. Same coupling range n = 753 neuron pairs. Different coupling range n = 6,586 neuron pairs.

Next, we compared mean Θ (dF/dt)_TSA aligned to tuft spike onset across coupling groups (Fig. 3E,F). SC synapses exhibited a higher rate of abrupt increases in activity at tuft spike onset (window: −50 to +150 ms) than the other coupling groups (SC_SD vs WC_SD: MD = 1.1 Hz, p = 0.002; SC_SD vs NC_SD: MD = 1.0 Hz, p = 0.02). Furthermore, we observed no indication that this observation was sex specific (Table 6). The differences between coupler groups were far less pronounced using Θ (ΔF/F₀) activity, suggesting highly coupled synapses have more trial-to-trial precise input timing with respect to tuft spike onsets (Fig. S3G). Interestingly, this spike-associated increase in SC activity was still highly dendrite-specific (SC_SD vs SC_OD: MD = 1.5 Hz, p = 6 × 10⁻⁴). This suggests that although SC synapses are relatively strongly correlated with each other (same vs different coupling range MD = 0.06 ρ, p = 7.0 × 10⁻³; Fig. 3D), at the onset of the tuft spikes the activity of SC synapses is driven along a dendrite-specific dimension in the activity space, and this dimension is distinct from the dimension that best captures their population-wide coactivity. Although there was a weak tendency for synapses of the same coupling group to cluster on the same dendrite segment, it was not statistically significant (same vs across dendrite MD = −0.053 a.u., p = 0.19; Fig. S2C), suggesting that SC synapses were not strongly targeted to particular dendrite segments.

Table 6.

Two-tailed hierarchical bootstrap mean difference tests for synaptic activity aligned to tuft spikes by coupling group (Fig. 3F)

Comparison	Mean difference (Hz)	p Value
SCSD versus WCSD	1.1	0.002
Male SCSD versus WCSD	0.93	0.05
Female SCSD versus WCSD	1.5	0.06
SCSD versus NCSD	1.0	0.02
NCSD versus WCSD	0.08	0.7
SCSD versus SCOD	1.5	6.0 × 10−4
WCSD versus WCOD	0.59	6.0 × 10−4
NCSD versus NCOD	0.77	6.0 × 10−4

Between coupler-group SD and SD versus OD p values separately corrected using Holm method. NC_SD n = 160 synapses, NC_OD n = 384, WC_SD n = 522, WC_OD n = 1,156, SC_SD n = 144, SC_OD n = 382. Male NC_SD n = 83 synapses, WC_SD n = 314, SC_SD n = 85, Female NC_SD n = 77 synapses, WC_SD n = 208, SC_SD n = 59.

Spike-associated activity occurs along distinct dimensions partitioned by task-selectivity and population coupling

After discovering a relationship between the activity of synapses at the onset of tuft spikes and the population coupling of synapses, we sought to determine whether different coupling groups encoded distinct task-variables. Trial-averaged Θ (ΔF/F₀) timecourses revealed clear qualitative differences in the timing of activity between coupler groups (Figs. 4A–F, S5A). SC synapses were predominantly active immediately following the GO cue (Fig. 4D), whereas NC synapses were predominantly active during a later epoch when mice typically collected reward (Fig. 4F). The timing of WC synapse activity was intermediate between the two other groups (Fig. 4E). Individual synapses of the different coupling groups also had qualitatively distinct selectivity for trial-type (Fig. 4A–C). Importantly, these differences in the average magnitude of early-response activity did not trivially define synaptic coupling, as coupling ranks remained stable after excluding this epoch (Fig. S3F).

Figure 4.

Synapses with different population coupling have distinct task-selectivity. A, Example SC synapse showing condition-invariant GO activity. Top: Θ (ΔF/F₀) activity for individual trials for each trial condition. Bottom: trial-averaged time courses for each trial condition. Error shading: bootstrap trial SEM. B, Same as (A) except a WC synapse with lick-direction selective activity. C, Same as (A) except an NC synapse showing positive reward-outcome activity. D–F, Peak-time-sorted heatmaps of trial-averaged time courses [Θ (ΔF/F₀)] of individual synapses for correct trials (n: NC = 518, WC = 1,555, SC = 518 synapses). G, Mean GO Mode, Response, and Outcome Direction activity for synapses grouped by coupling. Gray shading indicates the epoch used to calculate activity-direction weights. GO mode (−0.25 to 0.25 s) projected onto all trial conditions and averaged. Response direction (0–0.6 s) projected onto CR and CL trials and averaged; CL trials multiplied by −1 to directionally align selectivity. Outcome Direction (0.6–1.2 s) projected onto CR and CL trials and averaged (GO Mode and Outcome Direction n: NC = 434, WC = 1,267, SC = 479 synapses; Response Direction n: NC = 518, WC = 1,555, SC = 518 synapses). H, Mean activity in gray windows shown in (G), except GO mode is 0–0.25 s from GO to isolate post-GO activity. For (G–H), error shading indicates hierarchical bootstrap SEM and statistics were calculated with two-tailed hierarchical difference test (Table 7).

To quantitatively compare the encoding of task-variables across the coupling groups, we calculated dimensions within the activity space of all synapses (Fig. S5B; Methods; Li et al., 2016; Inagaki et al., 2022a) that best distinguished the onset of the GO cue (GO Mode), the lick direction (Response Direction), or the task-outcome (Outcome Direction). Projections of all synaptic activity along these dimensions (Fig. S5C) were qualitatively similar to those of local neural activity in ALM cortex (Inagaki et al., 2022a; Yang et al., 2022). Projecting the activity of different coupling groups along these dimensions revealed divergences in task-coding (Figs. 4G,H, S5D; Table 7). The encoding of the GO Mode was greater for SC synapses than NC synapses (Fig. 4G,H; MD = 0.006 a.u.; p = 0.002), whereas encoding of the Outcome Direction was greater for NC synapses than SC synapses (Fig. 4G,H; MD = 0.002 a.u.; p = 6 × 10⁻⁴). The task-encoding of WC synapses tended to be intermediate between the two other groups, and Response Direction encoding was similar across groups (Fig. 4G,H).Next, we investigated whether synaptic activity along these task-related dimensions was associated with the onset of tuft spikes in a manner dependent on coupling group. Despite biases in task encoding, the activity of every coupling group had substantial projections along multiple dimensions (Figs. 4G, S5D), and thus, a group’s trial-to-trial activity associated with tuft spikes could still be biased to different task-related dimensions than its trial-averaged activity. However, the spike-associated projections of Θ (dF/dt)_TSA along the GO Mode for both SC_SD synapses and WC_SD synapses were greater than for NC_SD synapses (Fig. 5A,B; Table 8; SC_SD vs NC_SD: MD = 0.005 a.u., p = 0.003; WC_SD vs NC_SD: MD = 0.002 a.u., p = 0.007). In contrast, spike-associated projections along the Outcome Direction were greater for NC_SD synapses than the other two groups (NC_SD vs SC_SD: MD = 0.001 a.u., p = 0.02; NC_SD vs WC_SD: MD = 0.001 a.u., p = 0.02). Projections were dendrite-specific for all coupling groups along the GO Mode (Fig. 5B; Table 8; SC_SD vs SC_OD: MD = 0.004 a.u., p = 0.02; WC_SD vs WC_OD: MD = 0.002 a.u., p = 0.01; NC_SD vs NC_OD: MD = 0.0006 a.u., p = 0.02). Projections along the Outcome direction were not significantly dendrite-specific for any coupling group (Fig. 5D; Table 8), likely reflecting greater noise in the estimation of the outcome mode as a result of low numbers of error trials per session.

Table 7.

Two-tailed hierarchical difference tests for synaptic activity modes aligned to GO cue (Fig. 4H)

Comparison	Mean difference (a.u.)	p Value
GO Mode
SC versus WC	0.0035	0.07
SC versus NC	0.0060	0.002
WC versus NC	0.0026	0.07
Response Direction
SC versus WC	0.0000	1.0
SC versus NC	0.0008	0.8
WC versus NC	0.0008	0.6
Outcome Direction
NC versus SC	0.0020	6.0 × 10−4
WC versus SC	0.0010	6.0 × 10−4
NC versus WC	0.0009	0.3

p Values for each activity mode separately corrected using Holm method. GO Mode and Outcome Direction synapse n: NC = 434, WC = 1,267, SC = 479; Response Direction synapse n: NC = 518, WC = 1,555, SC = 518.

Figure 5.

Spike-associated synaptic activity along task-selective dimensions. A, Mean GO Mode projected synaptic activity aligned to tuft spikes by coupling group using Θ (dF/dt)_TSA for SD (solid lines; n: SC_SD = 90, WC_SD = 287, NC_SD = 98 synapses) and OD (dashed lines; n: SC_OD = 355, WC_OD = 1,040, NC_OD = 353 synapses). B, Mean activity in gray windows (−50 to 150 ms from onset) shown in (A). C,D, Same as (A,B), except projected along the Outcome Direction. n: SC_SD = 83, WC_SD = 269, NC_SD = 96, SC_OD = 328, WC_OD = 1,000, NC_OD = 342. Error shading indicates session-level hierarchical bootstrap SEM and statistics used a one-tailed hierarchical difference test (Table 8).

Table 8.

One-tailed hierarchical difference tests for GO Mode and Outcome Direction synaptic activity aligned to tuft spikes (Fig. 5)

Comparison	Mean difference (a.u.)	p Value
GO Mode (Fig. 5B)
SCSD versus WCSD	0.003	0.1
SCSD versus NCSD	0.005	0.003
WCSD versus NCSD	0.002	0.007
SCSD versus SCOD	0.004	0.02
WCSD versus WCOD	0.002	0.01
NCSD versus NCOD	0.0006	0.02
Outcome Direction (Fig. 5D)
NCSD versus SCSD	0.001	0.02
NCSD versus WCSD	0.001	0.02
WCSD versus SCSD	0	0.6
NCSD versus NCOD	0.001	0.05
WCSD versus WCOD	−0.0003	1
SCSD versus SCOD	−0.0002	1

p Values separately corrected using Holm method for each group of hypotheses (GO Mode SD across coupler groups, GO Mode SD vs OD, Outcome Direction SD across coupler groups, Outcome Direction SD vs OD). GO Mode synapse n: SC_SD = 90, WC_SD = 287, NC_SD = 98, SC_OD = 355, WC_OD = 1,040, NC_OD = 353. Outcome Direction synapse n: SC_SD = 83, WC_SD = 269, NC_SD = 96, SC_OD = 328, WC_OD = 1,000, NC_OD = 342.

Taken together, these results suggested that two large-scale networks defined by population coupling encoded different aspects of the task: SC inputs preferentially encoded the transition between the preparation epoch and action epoch of the behavioral task, whereas NC synapses preferentially encoded task-outcome. These functional differences extended to the spike-associated activity of the synapses coupled to these two networks, and this activity remained dendrite-specific even after partitioning by both task-selectivity and population coupling.

Discussion

Our results uncover new relationships between spikes in the tuft dendrites and the activity of networks partitioned by coupling and task-coding, providing critical insights into potential roles for dendritic spikes in network-level operations, such as flexible behavior and learning.

Simultaneous high-speed glutamate and calcium imaging allowed us to map the input dynamics accompanying tuft spikes in L5 neurons. Recent studies of synaptic plasticity in layer 2/3 neurons have characterized input timing with respect to bAPs (Cornejo et al., 2022; Hedrick et al., 2024; Wright et al., 2025), but did not investigate trial-to-trial correlations. Thus, our recordings are among the first to precisely measure neuron-specific, trial-to-trial glutamatergic input statistics associated with the generation of postsynaptic events during behavior. We found that tuft spikes were associated with a dendrite-specific increase in input activity within ≈200 ms of the spike, in addition to the precise onset of additional activity within ≈70 ms of the spike. The specificity of this glutamatergic signal may reflect anatomically selective synaptic contact with nearby axons (Lee et al., 2016; Ding et al., 2025) or selective presynaptic scaling, which is known to positively correlate with postsynaptic scaling (Hardingham et al., 2010; Holler et al., 2021). Analysis of putative multi-branch Ca²⁺ transients (Fig. S1K,L) and the low frequency of tuft spikes (≈0.01 Hz; Fig. S1C) suggest that most tuft spikes we detected were likely multi-branch regenerative events, and not isolated local dendritic spikes or every bAP. These multi-branch events could be triggered by many mechanisms, including multiple concurrent NMDA spikes, Ca²⁺ spikes originating from the apical nexus, amplified bAPs, or combinations thereof (Stuyt et al., 2022).

It is possible that the dendrite-specific increase in input associated with tuft spikes reflects shared selectivity between the inputs and the postsynaptic neuron for unmeasured behaviors that varied from trial-to-trial. However, inputs encoding the unmeasured behavior would have to be precisely targeted to individual dendrites to a degree that is difficult to reconcile with existing evidence. Although there is some clustering of inputs by functional selectivity at scales of ≈10 μm (Fig. S2A,B; Wilson et al., 2016; Kerlin et al., 2019), the tuft dendrites of individual neurons receive inputs with diverse selectivity for behavioral variables across large spatial scales (Kerlin et al., 2019).

Alternatively, the dendrite-specific increase in input that accompanies tuft spikes could reflect trial-to-trial variability in the activation of subnetworks that are not exclusively defined by their behavioral selectivity. Both feedforward (Murphy and Miller, 2009; Hennequin et al., 2014; Daie et al., 2023) and recurrent (Douglas et al., 1995; Chance et al., 1999; Guo et al., 2017; Peron et al., 2020; Oldenburg et al., 2024; Deveau et al., 2026) subnetworks have been proposed as critical circuit mechanisms of signal amplification, persistence, and active filtering. The tuft dendrites of individual L5 ET neurons may receive targeted input from amplifying feedforward subnetworks (Daie et al., 2023) and cortical-thalamic-cortical loops (Guo et al., 2018), possibly receiving the output of a chain of loops that traverses the cortical hierarchy (Shepherd and Yamawaki, 2021; Young et al., 2021). Furthermore, recent work has demonstrated that the network influence of individual cortical neurons, rather than their behavioral selectivity, best predicts their influence on behavior (Daie et al., 2021, 2023; Gauld et al., 2024; Bounds and Adesnik, 2025; Pereira-Obilinovic et al., 2025). Thus, the dendritic spikes and concurrent dendrite-specific increase in input we observe may reflect trial-to-trial variable activation of different subnetworks—perhaps due to metastable recurrent dynamics (Friston, 1997; Mazzucato et al., 2015; Brinkman et al., 2022)—within intermediate dimensions that eventually drive the same action (Pereira-Obilinovic et al., 2025). Our results would also be consistent with models in which nonlinear dendritic integration and recurrent activity continuously interact to bind subnetworks (Antic et al., 2018) or perform learning-related calculations that traverse the cortical hierarchy (Payeur et al., 2021).

Interestingly, we also observed that average tuft spike-associated input peaked ≈100 ms after the onset of the tuft spike and then persisted over 280 ms (Fig. 2), well beyond the reported kinetics of iGluSnFR3 (rise <2 ms; decay = 29 ms; Aggarwal et al., 2026). Feedforward inhibition to the distal dendrites of the tuft (Palmer et al., 2012; Anastasiades et al., 2021; Rindner et al., 2022) could prevent input late within an input episode from triggering tuft spikes—similar to the temporal gating of action potentials by feedforward inhibition (Pouille and Scanziani, 2001; Wehr and Zador, 2003; Gabernet et al., 2005). Alternatively, this timing could reflect tuft spike-triggered bursts (Larkum et al., 1999) directly driving looped input (Guo et al., 2018; Payeur et al., 2021) or a “prepared state” that facilitates further activity in the subnetwork (Antic et al., 2018). Furthermore, recent work has found that learning modulates the excitability of the tuft dendrites of specific “engram” L5 ensembles (Rosier et al., 2025).

We uncovered additional evidence suggesting relationships between tuft spikes and network topology by analyzing the population coupling of inputs to the tuft dendrites of L5B neurons. The input population contained a SC subpopulation, similar to the rich-club topology observed in local cortical activity (Yassin et al., 2010; Gal et al., 2017; Hafizi et al., 2021; Tu et al., 2025; Finkelstein et al., 2026). Tuft spikes were more precisely aligned to the activity of SC inputs than WC inputs, suggesting that the more synchronous activity of SC inputs may be particularly effective at generating tuft spikes. Surprisingly, simultaneously recorded SC inputs that did not synapse on the spiking tuft exhibited minimally increased activity (Figs. 3E,F, 5A,B). SC neurons may encompass multiple rich-clubs that are interconnected (Gu et al., 2019; Tu et al., 2025) and these “hub” neurons may provide synchronized—but subnetwork-specific—input to tuft dendrites that is especially effective at triggering tuft spikes. SC neurons also exhibit greater plasticity during learning (Barnes et al., 2015; Grosmark and Buzsáki, 2016; Sweeney and Clopath, 2020), plastic changes that our data indicate may be efficiently communicated to the tuft dendrites. The labeling and imaging strategies of our study were not designed to evaluate tuft spike associated functional plasticity. However, future studies could utilize sparser labeling approaches and indicators that precisely report both subthreshold and suprathreshold activity (e.g., voltage indicators; Lee et al., 2026; Wu et al., 2026) to measure the impact of tuft spikes on the selectivity of dendrites in premotor cortex.

SC inputs were particularly active along a “GO” mode (Figs. 4G,H, 5A,B, S5D), a dimension that reflects the rapid transition between motor-preparatory and action-ready population activity patterns and that is not selective for trial-type (Inagaki et al., 2022a). Notably, rich-club neurons in layer 2/3 of ALM are also weakly selective for trial-type (Finkelstein et al., 2026). Our results indicate one possible mechanism by which rich-club activity could trigger rapid population transitions: by facilitating the generation of tuft spikes, which then trigger powerful burst firing (Larkum et al., 1999) or “prepared-state” (Antic et al., 2018) dynamics. Interestingly, we recently found that activity in the tuft of L5 ET neurons in ALM is more precisely aligned to the timing of the GO cue than somatic activity (Scheib et al., 2026), consistent with a potential role for tuft activity in facilitating state transitions.

The anatomical identity of the SC inputs to the tuft remains unknown. However, converging evidence suggests that thalamocortical inputs play a key role in coordinating a preparation-to-action state transition within ALM (Takahashi et al., 2021; Guzulaitis et al., 2022; Inagaki et al., 2022a) as well as state transitions in other cortical regions (Tanaka et al., 2018; Logiaco et al., 2021; Neske and Cardin, 2025). Considering that projections from VM thalamus to ALM cortex preferentially target the apical tuft of L5 ET neurons (Guo et al., 2018), it is possible that thalamic inputs are overrepresented within the SC population of inputs to the tuft.

Surprisingly, we also found slightly enhanced pairwise correlation between inputs in the lowest quintile of population coupling (Figs. 3D, S3E)—a topology not unique to synaptic inputs (Fig. S4B)—that preferentially encoded task-outcome (Figs. 4G,H, 5C,D, S5D). It is possible these inputs are coupled with previously identified mesoscale cortical networks that distribute emotion-related information and exhibit similar dynamics (Chen et al., 2024; Kauvar et al., 2025). Nonpositive coupling synapses also exhibited activity temporally aligned with tuft spikes (Fig. 3E,F), but unlike strong couplers, this activity was along the outcome direction rather than the GO mode (Fig. 5A–D). The outcome information conveyed by nonpositive coupling synapses may facilitate learning-related computations in the tuft (Guerguiev et al., 2017) or coordinate post-outcome state transitions, similar to the role we propose for strong couplers.

A previous study also reported high noise correlation among outcome-encoding neurons, but did not assess their population coupling (Ramesh et al., 2018). Another study (Kells et al., 2019) reported that WC neurons were the most selective for gross movement, an inconsistency with our results that may reflect differences between behavioral paradigms and sampling densities used to estimate coupling values.

In summary, we found that tuft spikes are accompanied by temporally extended input episodes that are highly dendrite-specific, suggesting the activation of sparse subnetworks. Yet, the synapses most associated with dendritic spikes belonged to one of two populations: one with large-scale coactivity that encoded action initiation and another that encoded task-outcome. This suggests that tuft spikes may be particularly driven by “hub” neurons that are both embedded in sparse subnetworks and synchronized to large-scale functional networks. Future experiments could employ targeted optogenetics to investigate causal relationships between input coupling and the generation of tuft spikes.