Introduction

During the acquisition and refinement of complex behavior, predictions about the state of the external world and the efficacy of actions are tested and revised. Take answering a phone as an example. From sensory input (i.e., the ring tone) a latent aspect of the world (i.e., ‘call being received’) is inferred, as is an optimal course of action (i.e., ‘swipe upward until phone unlocks’). However, when an action unfolds, errors in the inferred state (e.g., it was actually your coworker’s phone) or inferred action (e.g., a larger swipe is required to unlock) are identified, triggering corrective action and storage of information about the experience that will guide future estimates (i.e., learning). The frontal cortex plays a critical role in such flexible motor behaviors (Lopes et al., 2023; Mizes et al., 2024; Dragoi et al., 2026; Dasgupta et al. (2025)) and corrective action (Heindorf et al., 2018; Bollu et al., 2021), especially in the early stages of learning a new motor skill (Kawai et al., 2015; Hwang et al., 2019).

The circuit mechanisms of flexible behavior and rapid learning may depend on the nonlinear integration of signals in the apical tuft dendrites of layer 5 (L5) cortical pyramidal neurons. The tuft dendrites of L5 neurons receive distinct synaptic inputs (Spratling, 2002; Petreanu et al., 2009) and support regenerative events – tuft spikes – that depend on precise patterns of synaptic input and back-propagating action potentials (Larkum et al., 1999a; Larkum et al., 1999b; Stuart and Häusser, 2001; Larkum et al., 2009; Farinella et al., 2014; Stuart and Spruston, 2015). These tuft spikes facilitate burst firing (Larkum et al., 1999b) and generate a large influx of calcium, impacting both the ongoing output and the future function of neurons. In sensory cortices, evidence indicates that tuft spikes condition the sensory responses of L5 somata on top-down contextual signals (Larkum, 2013; Xu et al., 2012; Ranganathan et al., 2018; Fişek et al., 2023; Takahashi et al., 2016; Zhang et al., 2025) and can rapidly induce persistent changes in the selectivity of L5 neurons (Xiao et al., 2025; Yaeger et al., 2025). Furthermore, during learning, activity within the tuft of L5 neurons in primary sensory cortex changes along sensory dimensions relevant to a newly learned task (Benezra et al., 2024).

Tuft activity may serve different functions in the frontal cortex, however. Intracortical and thalamic connectivity motifs involving the tuft are different in the frontal cortex compared to primary sensory cortices (Shepherd and Yamawaki, 2021; Geng et al., 2022). These connectivity motifs may support long-timescale and rotational frontal cortex dynamics (Murray et al., 2014; Shepherd and Yamawaki, 2021) related to action planning and online movement correction (Riehle and Requin, 1989; Churchland et al., 2006; Kaufman et al., 2016; Heindorf et al., 2018; Ames et al., 2019; Sauerbrei et al., 2020). Tuft connectivity also changes during motor skill learning (Xu et al., 2009), concurrent with changes in the activity patterns of L5 neurons (Costa et al., 2004; Peters et al., 2017). Recent models propose that tuft spikes may encode both action initiation information (Takahashi et al., 2021; Guzulaitis and Palmer, 2023) and teaching information (Rao, 2024; Senn et al., 2024) that are distinct from the final output of neurons encoded in somatic spiking. Changes in laminar dynamics throughout learning also differ between the frontal and sensory cortex (Pollak et al., 2025), suggesting that representations within L5 neurons could change in both a compartment-specific and area-specific manner. However, functional encoding in the tuft dendrites and somata has not been directly compared during motor learning. As a result, it remains unclear what distinct information, if any, is conveyed by spikes in the tuft dendrites of L5 neurons within the frontal cortex.

We addressed this knowledge gap by performing longitudinal two-photon (2P) calcium imaging of apical tuft dendrites and somata of L5 extratelencephalic (ET) neurons in anterolateral motor cortex (ALM), a premotor region in the frontal cortex of mice. L5 ET neurons play a central role in encoding the trajectory of upcoming and ongoing movement (Evarts, 1968; Li et al., 2015) and are likely to be involved in the implementation of new dexterous actions (Peters et al., 2017; Li et al., 2017; Serradj et al., 2023). To measure activity associated with the acquisition of new motor trajectories, we trained mice to expert performance on a cued directional licking task and then shifted the lick targets. Relative to somatic activity, tuft activity showed weaker associations with the timing of the initial action but stronger associations with instructional cues and corrective actions, revealing compartmental differences that had not previously been reported. The gain of task-modulated activity within the two compartments also changed in opposite directions during learning. Our delineation of functional differences between tuft activity and somatic output provides critical new evidence to guide circuit models of credit assignment, motor control, and motor skill learning.

Results

Contingency-dependent motor skill learning

To better understand the role of tuft dendrite activity in motor control and learning, we organized our behavioral paradigm to induce a discrete change in movement trajectories under conditions known to require the premotor cortex (Guo et al., 2014b; Li et al., 2015; Makino et al., 2017; Bollu et al., 2021; Dasgupta et al., 2025). We monitored neural activity and tongue trajectories before (i.e., pre-shift) and after (i.e., post-shift) relocation of two lickport targets. Mice were trained to perform an auditory-cued directional licking task (Figure 1A; (Guo et al., 2014b; Inagaki et al., 2018) with two target lickports approximately equidistant from the mouth (Figure 1A,B). Once the animals reached expert performance, we collected neural recordings and high-speed videography of tongue kinematics across multiple pre-shift behavioral sessions (Figure 1B; N = 22 mice, N = 4 median pre-shift sessions per animal, N = 306 median trials per session; see Methods). To extract a low-dimensional summary of each lick trajectory that was intuitive and relevant to the task outcome, we calculated the angle of the tongue tip as it exited the mouth at a threshold distance (Figure 1C). As expected, exit angles during correct trials formed two distinct clusters in the space of exit angle (Figure 1D). For each animal, we used linear discriminant analysis to define a choice boundary that distinguished right and left licks during pre-shift performance (Figure 1D; Figure 1—figure supplement 1A).

Figure 1.

To induce motor skill learning, we shifted the location of both lickports to the right (Figure 1E). 2P imaging and task timing, as well as contingencies between cue, choice, and reward, all remained unchanged after the shift (N = 20 mice, N = 4.5 median post-shift sessions per animal, N = 302 median trials per session). The new left lickport position encroached on space that had been occupied by the tongue during pre-shift licks towards the right port (Figure 1E,F). Thus, on right-cued trials, mice continuing to follow the pre-shift motor plan would contact the left port (Figure 1E,F; Figure 1—figure supplement 1B) and receive punishment (i.e., timeout). We used the exit angle choice boundary (CB) to distinguish this type of motor error – where the animal intended to lick the right port but contacted the left port – from decision errors, where the animal simply chose the left port (Figure 1F). Exit angles after the shift exhibited a bimodal distribution across error trials (Figure 1G,H; Figure 1—figure supplement 1C,D), supporting this distinction in error type. The frequency of motor errors on right-cued trials increased significantly after the shift (median pre-shift 0.06, median post-shift 0.42, p < 0.001; Figure 1I; Figure 1—figure supplement 1C,D), reflecting the new challenge of avoiding the left lickport. Exit angles of licks directed towards the correct port (see Methods) shifted significantly rightward after the shift (late post-shift 4.11 ± 0.99°, p < 0.001; Figure 1J; Figure 1—figure supplement 1), indicating that mice had learned a new contingency-dependent tongue trajectory.

Longitudinal 2P calcium imaging in the apical tuft dendrites and somata of L5 ET neurons in ALM cortex

To monitor activity in L5 ET ALM neurons, we injected retrograde adeno-associated virus (AAV) encoding Cre recombinase into left ventromedial (VM) thalamus and AAV encoding Cre-dependent GCaMP8m (Zhang et al., 2023) into left ALM (Figure 2A). In post hoc histological sections, fluorescent neurons were restricted to left ALM cortex and no axonal labeling was detected in the corpus callosum (Figure 2B), indicating that our injections labeled L5 ET neurons and not intratelencephalic (IT) neurons. Daily imaging sessions alternated between L1 dendritic tuft and L5 somatic imaging (Figure 2D). To mitigate the effects of axial brain motion and capture more tuft dendrites at high imaging speeds (44.6 Hz frame rate), we implemented an excitation pulse splitting strategy that enabled extended depth-of-field (≈ 12 µm) 2P imaging in L1 (Figure 2E,F; Figure 2—figure supplement 1). Imaging of L5 somata was conducted with a standard 2P depth-of-field and interleaved line blanking (see Methods) to achieve high-fidelity imaging deep within the cortex at powers that were safe for longitudinal imaging (Figure 2F).

Figure 2.

To accurately track the activity of densely labeled dendritic arbors across many imaging sessions, we developed a custom image processing and segmentation pipeline that we used for both compartments (Figure 2—figure supplement 2A; see Methods). We used non-negative matrix factorization (NMF) to extract spatial components for each compartment (i.e. regions of interest; ROIs) from which we could isolate temporal components (i.e. traces) that minimized contamination from crossing or neighboring structures (Figure 2G; Figure 2—figure supplement 2A,E-F). To mitigate the impact of compartmental differences in calcium signal kinetics on downstream analyses, we deconvolved (Pnevmatikakis et al., 2016) the temporal components into impulses (Figure 2H-K; Figure 2—figure supplement 3A-C; see Methods). Estimated measurement noise and impulse frequencies were stable across days in both dendrites and somata (Figure 2—figure supplement 3D,E), indicating that our imaging conditions did not cause systematic changes in brightness or overall activity. Impulse rates were significantly higher in somata compared to tuft dendrites (median dendrites: 0.138 Hz N = 402 ROIs, N = 21 animals; median somata: 0.240 Hz N = 528 ROIs, N = 20 animals; p < 0.05 K-S test), consistent with previous measurements from L5 neurons in the visual cortex (Francioni et al., 2019). Together, this longitudinal imaging paradigm provided us with a uniquely comprehensive readout of population activity in both L1 tuft dendrites and L5 ET somata across learning.

Tuft activity tracks sensory cues more than planned action

Functional diversity across tuft dendrites has been observed in small numbers of L5 neurons in frontal cortex (Cichon and Gan, 2015; Kerlin et al., 2019; Otor et al., 2022). However, systematic differences in the sensorimotor selectivity of tuft dendrites and somata of these neurons have not yet been investigated. To address this, we first focused on analysis of data collected during the pre-shift period. Individual tuft dendrite ROIs (Figure 3A-C) and somata (Figure 3D-F) exhibited trial-type-specific activity that was generally consistent across sessions. To visualize task-related activity across the population of both dendrites and somata, we collected correct response trial types, calculated the mean GO-cue-aligned responses and sorted ROIs according to their trial-type preference and peak activity (Figure 3G,H; see Methods). Both tuft and somata populations exhibited diverse selectivity for trial type and a range of activity peaks that spanned the duration of trials. Upon close inspection of trial-to-trial activity, we observed that the onset of tuft activity was consistently time-locked to the GO cue (vertical green line; Figure 3B). This was in contrast to somatic activity, which had more variable timing (Figure 3E). We hypothesized that peri-GO tuft activity was more temporally aligned with the GO cue and that somatic activity was more aligned with the trial-to-trial timing of the initiation of action (Figure 3I,J). To systematically test this, we modeled the activity of each dendrite and soma as the sum of two nonnegative response functions: one response locked to the timing of the GO cue and another locked to the timing of first port contact (Figure 4A; see Methods). GO-associated activity sharply increased immediately after the GO cue in both dendrites and somata (Figure 4B,C; somata activity 2.58 ± 0.36, p < 0.001; dendrites 2.78 ± 0.24, p < 0.001; null hypothesis 1.0; between compartments p = 0.67), whereas contact-associated activity during this period was greater in somata than dendrites (Figure 4B,C; somata activity 2.02 ± 0.18, p < 0.001; dendrites 1.15 ± 0.13, p = 0.25; null hypothesis 1.0; between compartments p < 0.001) and not significantly different than baseline in dendrites. Thus, we found that peri-GO activity in the tufts is predominantly aligned with the GO-cue, whereas somata activity reflected both the timing of the GO-cue and the timing of action. To our knowledge, this is unique evidence of a strong functional distinction between tuft activity and somatic activity during motor behavior.

Figure 3.

Figure 4.

To further understand the behavioral features encoded by activity in tuft dendrites and somata, we next considered activity during both correct trials and error trials (Figure 4D,E). Pre-shift, almost all errors were decision errors (decision error median 81.3% of all errors). Thus, behavioral trials could be categorized into four different trial types (Figure 4D,E) which differed in sensory instruction (i.e. initial cue), animal choice (i.e. port licked), and the task-outcome (i.e. presence or absence of reward). We modeled the mean activity of dendrite and soma ROIs during each trial type as a linear combination of these three task variables (Chen et al., 2017; Yang et al., 2022). Since estimates of the selectivity of individual dendrites and somata tended to be noisy, we focused on population-level analyses of selectivity in an “activity space” where each dimension corresponded to the activity of individual dendrites or somata (Stopfer et al., 2003). From the linear model, we calculated coding directions (CDs) at each timepoint that maximally separated Sensory, Choice, and Outcome activity (Figure 4F; Figure 4—figure supplement 1A) and estimated the selectivity along each dimension across time (Figure 4—figure supplement 1B,C; see Methods). Tuft dendrites exhibited higher sensory selectivity than somata during presentation of the initial auditory tones (Figure 4F,G; dendrites 0.029 ± 0.013; somata 0.008 ± 0.003; inter-compartment p < 0.05), and higher relative sensory selectivity compared to choice and task-outcome selectivity during the response epoch (Figure 4G-I).

In summary, we found systematic differences in the baseline functional representations within the tuft dendrites and somata. Tuft activity was strongly sensory-selective and aligned to the timing of instructional cues. Somatic activity also encoded these features, but in addition contained strong representations of choice, task-outcome, and action timing – representations that were weaker in the dendrites.

Distinct encoding of corrective action in tuft activity

We next investigated whether activity in the tuft dendrites reflected motor corrections made during skill learning (i.e., post-shift). Theories (Körding and König, 2001; Guerguiev et al., 2017; Sacramento et al., 2017) and recent experiments (Francioni et al., 2026) have suggested that apical dendrites are a critical locus of credit assignment calculations in the cortex. However, the nature of these calculations and their relationship with precise behavioral events remain unclear. In our motor learning paradigm, mice not only made behaviorally distinct errors across trials (Figure 1H), but also exhibited distinct behaviors after a motor error (Figure 5A). We focused our analysis on trials in which the exit angle of the first lick indicated intent to target the correct lickport (i.e., based on the choice boundary; Figure 1H). These trials were then divided into three distinct trial types based on the success or failure of the first lick and the nature of any subsequent lick (Figure 5A). On Correct Right (CR) trials, the first lick made contact with the correct port and reward was delivered. On Correction Attempted (CA) trials, the first lick made contact with the incorrect port, and the mouse chose to direct a second lick towards the correct port. We interpreted these licks as a corrective action, because the tongue exit angle shifted further towards the correct target (Figure 5B). Abandoned Port (AP) trials were the same as CA trials, except the mouse either did not make a second attempt or the second lick was directed towards the incorrect port.

Figure 5.

Qualitative inspection of dendritic activity during these trial types revealed a sparse subset of dendrites exhibiting especially strong activation on CA trials after the first lick, but not on other trial types (Figure 5C,D; Figure 5—figure supplement 1). In contrast, the most strongly activated somata were selectively active on CR trials (Figure 5E,F; Figure 5—figure supplement 1). To quantify these differences, we again utilized population-level analyses. We focused on differences in activity between CR and CA trials, since neither compartment was particularly active on AP trials. We calculated a CD for each time bin that best distinguished CR and CA trials and then projected CR and CA trials along the CD (Figure 5G; Figure 5—figure supplement 2A). The projection of CA dendritic activity along this dimension was large after the first lick (Figure 5H,I; see Methods for normalization details). This was similar in magnitude to projection along a CD that distinguished CR and Correct Left (CL) trials (Figure 5—figure supplement 2B,C) – whereas projection of CR activity was comparatively weak. Projections of somatic activity along this dimension exhibited the opposite pattern (Figure 5H,I). The dendritic CA projection was significantly larger in magnitude than the somatic CA projection (Figure 5I; dendrites: −0.759 ± 0.18; somata: −0.074 ± 0.10; p < 0.01), while the somatic CR projection was significantly larger than the dendritic CR projection (dendrites: 0.175 ± 0.14; somata: 0.872 ± 0.08; p < 0.05).

The CD that distinguishes activity on CR trials and CA trials could reflect selectivity for many different behavioral variables that differ between the two trial-types. First, it could reflect simple selectivity for task-outcome (i.e., reward or punishment). However, projections of activity on CL and Decision Error Right (DER) trials onto this CD were near zero for both dendritic and somatic activity (Figure 5—figure supplement 2C), indicating that the CD reflects selectivity for task-outcome that is specific to the contingency (i.e., “lick right on right-cued trial”). To determine if activity was additionally selective for the presence or absence of corrective action, we projected AP activity along the same CD, since initial action and task outcome were the same on AP and CA trials. For somata, the projection of AP activity was similar to the projection of CA activity (Figure 5H,I; AP −0.123 ± 0.10, p = 0.115; CA −0.074 ± 0.10, p = 0.442), suggesting that this CD did not distinguish between follow-up actions in somata. However, for dendrites, the projection of AP activity was lower than the projection of CA activity (Figure 5H,I, AP −0.118 ± 0.10, p = 0.38; CA −0.759 ± 0.18, p < 0.05), indicating that dendritic activation along this dimension was selectively associated with corrective action. In summary, tuft dendrites encode a highly selective representation of corrective action, consistent with theories that posit a central role for tuft activity in credit assignment and learning.

Divergent changes in tuft and somata function across motor skill learning

Learning can trigger transient and sustained reorganizations of cortical representations (Costa et al., 2004), including changes in the activity of the tuft dendrites of L5 neurons (Lacefield et al., 2019; Benezra et al., 2024; Schoenfeld et al., 2021). However, whether the tuft dendrites and somata of L5 neurons undergo similar or divergent functional changes during motor skill learning remains unknown. To address this, we compared the trial-aligned activity of individual tufts (Figure 6A) and somata (Figure 6B) before and after shifting the lickports. We focused our analysis on correct trials because the lick pattern (i.e., one contact lick followed by consummatory licking) on correct trials remained similar across the shift, facilitating clearer interpretations of any change in activity pattern.

Figure 6.

Learning can be accompanied by many different changes in cortical representations, such as changes in the prevalence, selectivity, and responsiveness (i.e., gain) of feature-sensitive neurons. To distinguish between these options, we first identified the ≈0.5 s time window with the largest difference in activity between CR and CL trials for each dendrite or soma (see Methods). In this time window, we then calculated a selectivity index (SI) in each training epoch (i.e., pre- or post-shift) as the mean difference between CR and CL responses divided by their sum. If the SI was reliable (95% CI < 0.66), the dendrite or soma was considered “task-modulated” in that training epoch regardless of the SI value. In both dendrites and somata, the post-shift distribution of SI shifted towards the edges compared to pre-shift (Figure 6C). A very small fraction of dendrites or somata changed preference across the shift (Figure 6C; dendrites N = 5 or 3%; somata N = 7 or 4%), indicating minimal population remapping. The prevalence of task-modulated dendrites and somata also did not significantly change across the shift (Figure 6D; Δ selective dendrites 11.98 ± 8.08%, p = 0.065; somata 6.78 ± 5.75% p = 0.212). However, SI values in both compartments increased (Figure 6E, Δ SI dendrites 5.43±2.15%, p < 0.05; somata 11.16±3.73%, p < 0.01). Surprisingly, the port shift induced divergent changes in the responsiveness of both compartments (Figure 6F; dendrites −17.75 ± 5.01%, p < 0.001; somata 12.06 ± 4.52%, p < 0.01; between compartments p < 0.001). Dendritic responses decreased, whereas somatic responses increased. The overall impulse rates of dendrites and somata during behavior remained steady across the shift (Figure 2—figure supplement 3E), so these changes were presumably specific to trial-type selective activity. We then calculated net selectivity as the difference between CR and CL responses (i.e., without normalization by the sum). The net selectivity of dendrites did not change significantly across the shift (Figure 6G; −8.88 ± 6.04%, p = 0.127), whereas the net selectivity of somata increased (Figure 6G; 27.05 ± 6.04%, p < 0.001; between compartments p < 0.001).

To measure net selectivity across the population without any filtering for task-modulation, we again returned to population-level analyses in activity space. Within each training epoch, we calculated the CD that best distinguished activity between CR and CL trials (Figure 6H), cross-validated projections of CR and CL activity along these CDs (Figure 6—figure supplement 1A,B), and determined the selectivity projection as the difference between these projections (Figure 6I). The difference between the pre- and post-shift selectivity (Figure 6J) confirmed that net selectivity did not change across training epochs in the tuft dendrites, but increased substantially in the somata (Figure 6K; dendrites −3.08 ± 9.89%, p = 0.724; somata 32.18 ± 7.60%, p < 0.001; between compartments p < 0.01). This increase in the magnitude of the net selectivity of somata occurred early in learning (first 2 training sessions after the shift) and then returned to near-baseline in later sessions (Figure 6—figure supplement 2), consistent with previous studies of L5 neuron output in frontal cortex during learning (Costa et al., 2004; Peters et al., 2017; Hwang et al., 2021). Changes across learning in the direction of the CD were not significantly different between dendrites and somata (Figure 6—figure supplement 3), consistent with their similar magnitude of changes in SI across the shift (Figure 6E).

In summary, the tuft dendrites and somata exhibited divergent changes in the gain of task-modulated responses across learning. Increases in dendrite selectivity were tightly balanced by decreases in gain, suggesting that tuft activity is regulated to maintain consistent net selectivity across learning. In contrast, the net selectivity of the somata increased transiently during early learning.

Discussion

Previous work identified diverse behavior-related dendritic activity that was localized to different compartments within individual layer 5 neurons in the frontal cortex (Kerlin et al., 2019; Otor et al., 2022). However, any systematic functional differences between regenerative activity in the tuft dendrites and somatic output remained unclear. In the present study, we uncovered systematic differences between the tuft dendrites and somata in sensorimotor selectivity, dynamics, corrective signaling, and response gain across different phases of motor learning. These differences in encoding and functional plasticity suggest specialized roles for tuft computation in motor control and motor learning, and provide a critical new foundation for further investigation of the contributions of dendritic computation to behavior.

Selectivity and timing of L5 tuft activity during expert behavior

We found that, during the sensory cue period of the licking task, tuft dendrites of L5 neurons in ALM cortex are more sensory-selective than somata. This is consistent with the general role for tuft spikes in conditioning the activity of L5 somata on feedback signals (Larkum, 2013; Xu et al., 2012; Manita et al., 2015; Ranganathan et al., 2018; Fişek et al., 2023; Zhang et al., 2025) and recent work indicating that suppression of L1 activity in ALM cortex prevents relearning of context-dependent rules (Maristany De Las Casas et al., 2025).

More surprisingly, we found that tuft activity in ALM cortex exhibited distinct timing that was associated with the GO cue, but not the trial-to-trial timing of action initiation. In contrast, somatic activity was associated with both GO cue and action initiation. Interestingly, ventromedial (VM) motor thalamus provides stronger input to the apical tuft than to the basal dendrites of Layer 5 ET neurons in ALM (Guo et al., 2018). Cue-triggered VM input rapidly shifts ALM activity along a distinct “GO” mode prior to movement onset (Inagaki et al., 2022b). However, this shift along the “GO” mode is not sufficient to initiate movement. Instructive activity along a different activity mode, dominated by the output activity of L5 ET neurons, prepares and executes directed movement (Li et al., 2015; Inagaki et al., 2022b). Our finding that tuft activity in L5 ET neurons in ALM is predominantly aligned to the GO cue raises the possibility that it could drive or reflect the transition along the “GO” mode. Tuft spikes could drive bAP-activated calcium (BAC) burst firing (Larkum et al., 1999b), signaling the state transition within a movement-null space (Kaufman et al., 2014) to downstream structures, as well as back to L5 IT and ET neurons in ALM via low-latency loops (Guo et al., 2018). Subsequent somatic firing, uncoupled from the tuft, could control action within a movement-potent space. Consistent with these ideas, recent models posit that sparsely active (Logiaco et al., 2021) or weakly selective (Pereira-Obilinovic et al., 2025) thalamic inputs can rapidly transition the cortical network towards particular future actions.

Representation of corrective action in the tuft of L5 ET neurons

Frontal cortex is a critical node in an action control system that selectively minimizes goal-relevant errors (Todorov and Jordan, 2002; Scott, 2012) through both prospective internal (Kao et al., 2021) and ongoing sensory (Sauerbrei et al., 2020; Kalidindi et al., 2021) feedback control. Information regarding prediction errors must be utilized to generate real-time corrections and to remodel the context-specific networks that generated the action in order to improve future movement. How corrective action is represented in the cortex is still debated (Perich et al., 2024; Ames et al., 2019; Kudryashova et al., 2025; Malonis et al., 2021) and circuit mechanisms linking corrective action to the plasticity of individual neurons remain unclear.

We found that strong activation of the tuft population was selectively associated with corrective action, while the somatic population was particularly active during reward and generally active during action (i.e., regardless of whether that action was corrective or not). The deconvolved tuft signal associated with a corrective lick peaked ≈200 ms after the average time of second port contact and remained elevated for another ≈500 ms. This is longer than GCaMP8m kinetics (Figure 2—figure supplement 3A-C; Zhang et al., 2023), suggesting that most of the tuft signal occurred after the initiation of the corrective movement. Thus, our results favor a model in which the tuft spikes associated with corrective action are involved in post-correction computations (e.g., driving plasticity) rather than driving the corrective action itself. Our finding that strong tuft activity selectively occurred during corrective action suggests that nonlinear integration of error (i.e., “target missed”), context (i.e., “right port is the target”), and movement (i.e., “second lick is rightward”) occurs either within the tuft or upstream of it. A recent study in retrosplenial cortex found that dendritic spikes in neurons were modulated by performance signals related to the causal role of each neuron in a neurofeedback-based task (Francioni et al., 2026). Although we do not know whether the strong correction-related activity we observed played a causal role in learning, the specificity of the behavioral circumstances that generated this signal could be consistent with a responsibility-weighted error signal to the subnetwork of L5 ET neurons that controlled the contingency-specific action (i.e., rightward choice). If this correction-associated signal, observed in the tuft dendrites of L5 ET neurons, indeed reflects a credit assignment calculation, many afferent inputs to the tuft could provide the necessary information for such a calculation, including the prediction error (Manita et al., 2015; Ako et al., 2025; Heindorf et al., 2018; Levy et al., 2020), the context-specific goal (Moberg and Takahashi, 2022; Inagaki et al., 2022a), and an efference copy of past output (Guo et al., 2018). Now that this correction-associated tuft signal has been identified, future work can determine the precise circuit mechanisms generating it, as well as determine whether it is indeed a teaching-related signal or reflects some other correction-associated computation.

Motor skill learning triggers compartmentalized changes in representation and gain

Our results demonstrate that across motor skill learning, the selectivity of tuft dendrites of L5 ET neurons in ALM cortex increases. This is similar to the enhanced selectivity for behaviorally relevant task dimensions after learning that has been observed in the tuft dendrites of L5 neurons in the somatosensory cortex (Lacefield et al., 2019; Benezra et al., 2024; Schoenfeld et al., 2021). Our recordings from the somata of L5 ET neurons in premotor cortex revealed a similar increase in selectivity. However, we found that motor skill learning was accompanied by a decrease in the response gain of task-modulated dendrites and an increase in the gain of task-modulated somata. To our knowledge, these divergent changes in the gain and net selectivity of L5 tuft dendrites and somata of frontal cortex have not been previously reported. The increase in the somatic net selectivity declined later in learning, consistent with previous findings in motor cortex (Peters et al., 2017).

Most surprisingly, the increase in selectivity of selective tuft dendrites was accompanied by a matched decrease in overall responsiveness, such that net selectivity remained stable across learning. This suggests that, even within a particular behavioral context, total dendritic activity and the sparseness of that activity may be tightly regulated. Interestingly, a recent study of L5 neurons of the auditory cortex found decreased, but more synchronous, activity across tufts during recall of conditioned stimuli and associated this change with an increase in I_h-current (Rosier et al., 2025). L5 neurons in visual cortex also exhibit divergent compartmental changes in activity following passive exposure to pattern-violating stimuli (Gillon et al., 2024).

We did not investigate causality, so the precise relationship between the neural changes we observe during learning and the behavior of the animal remains unclear. Given the port shift direction relative to the hemisphere we imaged, the increased net selectivity we observe could be directing licking trajectories towards the new port locations. Alternatively, the changes in net selectivity could reflect increased engagement of ALM cortex due to increased uncertainty about upcoming action (Dragoi et al., 2026). Since total activity did not change, the specific increase in the gain of neurons selective for the task could be considered a form of “motor attention” (Dragoi et al., 2026). This could also explain the decrease in the gain of the tuft dendrites, as some pathways that are more active during increased arousal or attention suppress the apical dendrites while simultaneously increasing the gain of the basal compartment (Malina et al., 2021).

Study limitations

In order to detect sparse representations in the tuft, we adopted a dense labeling strategy that precluded unambiguous tracing of dendrites. The dendritic signals we measured could reflect bAPs and regenerative events that generate global, hemi-tree, or local branch calcium influx (Hill et al., 2013; Otor et al., 2022). Based on previous calcium imaging of L5 tufts in ALM cortex of mice engaged in similar tasks (Kerlin et al., 2019; Maristany De Las Casas et al., 2025), we suspect that most of the activity we measured was coincident with global tuft or hemi-tree events, as well as somatic spiking. However, the distinct correction-associated and cue-aligned signals we observed in the tuft represent only a fraction of total tuft activity and could be produced by many different mechanisms. Future studies that directly manipulate dendrite conductances could begin to identify precise mechanisms and investigate causal relationships with behavior.

Methods

All experiments were conducted in accordance with the NIH Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee of the University of Minnesota. Data are from 22 mice (C57BL/6J, Jackson Laboratories, stock #000664) of either sex (82 ± 20 days old at the beginning of experiments). Eleven were male, and eleven female. There were no sex-specific differences in the results and data were pooled across sex.

Surgery and Virus Injection

Cranial window surgeries were conducted as previously described (Daie et al., 2023). We injected viral retrograde Cre recombinase (pENN-AAV-hSyn-Cre-WPRE-hGH, Addgene; 1.9×10¹² vg/mL; 200 nL) through a burr hole in the skull into left thalamus (−1.5 mm posterior, 0.8 mm lateral, 3.8 mm deep relative to bregma). We next performed a 3 mm craniotomy over left ALM (2.5 mm anterior, 1.5 mm lateral relative to bregma) and injected the exposed cortex with FLEX-GCaMP8m (pGP-AAV-syn-FLEX-jGCaMP8m-WPRE, Addgene; 1.7 × 10¹² vg/mL; 75 nL per injection; 9 injections total in a 3×3 grid; depth of 700 µm below pia). A triple coverglass window (#1 coverglass, 2.5 mm / 2.5 mm / 3 mm; Potomac Photonics and Warner Instruments) covering the cortex and a headbar were affixed to the skull with adhesive (C&B Metabond; Parkell) and dental acrylic (Jet; Lang Dental).

Behavioral Paradigm

During training and recording, mice were placed on water restriction (1 mL per day). Initial training and task structure were similar to previous studies (Guo et al., 2014a; Guo et al., 2014b; Inagaki et al., 2018) and shown in Figure 1A. All task epochs and triggers were controlled by a BPOD State Machine (Sanworks). At the beginning of each trial, mice were presented with an auditory sequence (5 tones of 0.15 s with 0.1 s between tones; total duration 1.2 s) of 3 kHz (left trial) or 18 kHz (right trial) pure tones. After presentation of these “sample period” tones, mice had to withhold licking for a 1.25 s delay until the presentation of an auditory GO cue (carrier frequency: 6 kHz, modulating frequency: 360 Hz) lasting 0.1 s. Port contact prior to the GO cue reset the timer for that epoch and these trials with “early” licking were excluded from further analysis. Mice had 2 s to respond by licking one of two lickports with internal edges (the typical locations of the first tongue contact) separated by a distance of 2 mm. Tongue contact with the correct port first (as reported by a custom electrical lick detector) resulted in a water reward (3 µL) at that port. Contact with the incorrect port first resulted in a 15 s timeout. Trials in which the mouse failed to respond after the GO cue were also excluded from further analysis. To ensure the animals learned a stereotyped licking behavior the lickports were fixed in position with respect to their head across pre-shift training sessions. Mice typically learned this task over the course of ≈25 training sessions. Once task performance reached threshold (>75% correct), baseline tongue trajectory and neural data (see Methods) was collected over ≈4 sessions. After recording baseline sessions, both lickports were shifted by 1 mm to the right of the mouse on the hundredth trial of the “shift” session (session 0). Across all sessions, all trials prior to the trial of lickport shift were considered “pre-shift” and all trials subsequent to that trial were considered “post-shift” for subsequent analysis. For more detailed analyses of learning dynamics, we specified 4 phases of learning: two phases before the shift (pre-shift early: >1 session before the session of the shift; pre-shift late: the session immediately prior and session of the shift up to the shift trial) and two phases after the shift (post-shift early: session of the shift after the shift trial and the immediately subsequent session; post-shift late: >1 session and less than 3 sessions after the shift session).

Behavior Analysis

Tongue Tracking

Two CMOS cameras (FLIR Blackfly S, BFS-U3-04S2M) were located to the side and below the mouse (Figure 1A,B). Custom software (Tran, 2026) captured timestamped video at 500 Hz from both cameras for 2 s after every GO cue. The animal was illuminated with infrared LEDs. For analysis of the tongue trajectory, we first collected all recorded images from both the side and bottom cameras across all sessions and all animals. A custom convolutional neural network (Dinh, 2026) determined whether the tongue was visible outside of the mouth in each frame. To extract the tongue position in 3D, we then trained a second neural network using DeepLabCut (Mathis et al., 2018) on a subset of frames in which the tongue was visible. This network was used to estimate the tongue’s 3D coordinates for all frames containing a visible tongue. Using these outputs, we constructed a time-resolved trajectory for each lick synchronized with electrical lickport signals to determine if and when each lick contacted a lickport.

Lick Angle Quantification and Classification

To reduce each 3D tongue trajectory to an intuitive estimate of movement towards the lickports, we calculated the exit angle (i.e., azimuth θ and elevation ϕ) of the tongue as it extended beyond a critical distance from the mouth. This distance was the minimum distance from the mouth to the closest lickport position across both pre-shift and post-shift lickport positions. Linear Discriminant Analysis (LDA; python scikit-learn; Pedregosa et al., 2011) of pre-shift exit angles on correct trials was used to determine a choice boundary. Exit angle with respect to this boundary was used to define the intended target of the lick, regardless of the subsequent port contact.

Exit angle distance from the choice boundary (Figure 1—figure supplement 1) was calculated as follows. First, the lick coordinates (θ, ϕ) were projected onto the LDA decision line in 2D angle space. Then, each lick and its projection were converted into 3D unit vectors (v_lick, v_proj), originating from the animal’s mouth. We defined the lick angle as the angular separation (ω) between these vectors:

where sgn is the sign of the LDA decision function, assigning positive or negative values based on the classified lick target (positive = right, negative = left). For population analyses, we normalized each animal’s lick angle by subtracting the average lick angle calculated from the entire pre-shift period. This relative lick angle was then averaged across all trials within each training epoch, and then averaged across animals. For SEM estimates and hypothesis tests, we bootstrapped this process 10000× by drawing animals with replacement.

Gaussian Mixture Model Analysis

To assess whether the lick angle distributions were unimodal or bimodal, we fit one- and two-component Gaussian mixture models (GMMs) to the lick angle data using maximum likelihood estimation. Models were compared using the Bayesian Information Criterion (BIC). The difference in BIC between the one- and two-component models (ΔBIC = BIC₁ − BIC₂) was used to assess relative model fit, with positive values indicating evidence in favor of the two-component model.

Functional 2P calcium imaging

For high-speed extended depth-of-field imaging of L1 dendritic tufts, we built a custom 2P pulse splitter system for multi-foci simultaneous excitation combined with high-speed scanning. This extended-depth-of-field 2P imaging approach maintained a high excitation efficiency by confining excitation light to four diffraction-limited foci. This allowed us to use average infrared laser powers that were suitable for long-term imaging. In brief, excitation light (960 nm; 100 fs; 80 MHz; InsightX3, Spectra Physics) was temporally multiplexed using a custom pulse splitter consisting of a series of polarization gated delay lines that produced an output pulse train of 320 MHz (see Figure 2E; Figure 2—figure supplement 1). Optical relays within each delay line imparted a variable divergence, resulting in the generation of four foci equally separated in time and space that spanned an approximate depth of 12 µm. For more details, see Figure 2—figure supplement 1. L1 imaging was conducted with a 25× microscope objective (Olympus XLPLN25XWMP2 25× NA 1.05) and average laser power exiting the objective of between 20–50 mW (mean: 34 mW, 0.11 nJ pulse energy per focus). The frame rate was 44.6 Hz and the field of view (FOV) was 210 µm × 210 µm × ≈ 12 µm.

A different configuration was used to achieve high signal-to-noise ratio recordings from somata in L5 at reasonable average laser powers. We blocked 80 MHz excitation light from reaching the sample on 3 out of every 4 imaging lines scanned by our resonant mirror. This was a cost-effective alternative to maintaining a separate, lower repetition rate light source. L5 imaging was conducted with a 16× microscope objective (Nikon CFI75 LWD, NA 0.8) and average laser power exiting the objective (with 75% line blanking) was 9 mW to 38 mW (mean: 22 mW, 1.1 nJ pulse energy). The frame rate was 22.8 Hz and the FOV was 328 µm × 328 µm.

In both imaging configurations, lateral scanning was performed by a resonant mirror (CRS 12 kHz, Novanta Photonics) conjugated to a set of galvanometer mirrors (6215H 5mm, Novanta Photonics). Emission photons were detected using a photomultiplier tube (PMT; PMT2101, Thorlabs), digitized (NI-5771/NI-7975R/PXIe-1092, National Instruments) and recorded with ScanImage (MBF Biosciences).

Anatomical 2P imaging

After completion of functional imaging, a z-stack was collected (328 µm×328 µm FOV; 44.6 Hz frame rate; 100 frames per step; 2 µm steps) from pia to L5 using a 16× microscope objective (Nikon CFI75 LWD, NA 0.8). Frames were registered within each slice, mean-projected within slice, aligned to neighboring slices, and then mean-projected along the stack y-axis.

Histology

After completion of all data collection, animals were deeply anesthetized with ketamine-xylazine and a transcardiac perfusion was performed via the left ventricle with saline followed by 1:10 buffered formalin. The brain was extracted and post-fixed in formalin overnight at 4 °C, then rinsed in phosphate-buffered saline (PBS). Coronal sections of 60–100 µm thickness were collected with a vibratome and mounted in Vectashield (Vector Laboratories). GCaMP8m fluorescence was imaged (excitation wavelength: 488 nm) using an upright Nikon Ti2 confocal microscope with a Plan APO 10× NA 0.45 air objective. Whole coronal sections were imaged using tiles to cover the entire slice, stitched using NIS-Element (Nikon Imaging), and finally aligned to the Allen Brain Atlas (Wang et al., 2020). Histology images were spatially smoothed for display purposes (Gaussian σ = 2 µm; Figure 2B,C)

Image Analysis

In order to extract activity traces for dendritic tufts and somata, image series underwent a combination of rigid registration, denoising, non-rigid warp correction, segmentation and finally deconvolution (Figure 2—figure supplement 2A). Front-end data intake and metadata management was performed using custom Matlab (Mathworks) scripts. All other imaging data were processed using custom Python scripts.

Rigid registration

All image series were subjected to an initial rigid registration (phase cross-correlation; no upsampling; python scikit-image; (Walt et al., 2014)). All imaging from the same animal and same layer was registered to a common reference target. Offsets were estimated from spatially filtered (Gaussian σ ≤ 1 µm) frames, adaptively filtered to suppress registration noise, and then used to transform unfiltered frames.

Denoising

Following rigid registration, image series were denoised using DeepInterpolation (Lecoq et al., 2021), modified to minimize L2 (rather than L1) loss and to receive as its input, 10 mean projections of frames collected from non-overlapping exponential (base 1.75; exponents: 0–9) frame windows both preceding and following the interpolated frame.

Non-rigid registration

Our segmentation methods (see Methods) required that all dendritic imaging from the same animal be very precisely aligned. Small within-session and across-session instabilities (e.g., resonant mirror temperature changes, vessel dilation) resulted in image warp that was significant for longitudinal dendritic imaging. Warp calculations were performed on contrast-enhanced frames using symmetric diffeomorphic registration (python DIPY; Garyfallidis et al., 2014; Avants et al., 2008), 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. Comparisons of denoised and warp corrected images can be seen in Figure 2—figure supplement 2B-D. Due to their relatively larger size and spherical shape, soma imaging did not require warp correction.

Segmentation and demixing

Segmentation and demixing were performed on an image series comprising all imaging sessions from a given animal and given layer. This image series was cropped to contain only pixels with valid data across the entire image series (Figure 2—figure supplement 2B). We used a software package optimized for performing NMF on large datasets within a distributed computing environment (Chennupati et al., 2020), to calculate a single factorization (Figure 2—figure supplement 2A). NMF was performed on a compressed version (N = 3 frame binning), normalized, and baseline-corrected version of the image series (see Figure 2—figure supplement 2A). Unlike piecemeal approaches to NMF-based segmentation of large dendritic imaging datasets (Buchanan et al., 2018), the output of this complete NMF pipeline was not prone to spatial discontinuity artifacts from block stitching. We calculated a single factorization of at least 100 initial components within each FOV to ensure that even weakly active dendrites or somata would be identified. Minimally active components and components with a spatial distribution indicating that they were noise-dominated, were excluded from final components. Initial components with temporal correlation > 0.8, spatial correlation > 0.5, and binarized spatial overlap > 50% were merged to generate a final set of ROIs (Figure 2G). For each imaging session, temporal components were calculated from denoised and warp-corrected image series using Hierarchical Alternating Least Squares (HALS; python scikit-learn; Pedregosa et al., 2011) with the ROIs as fixed spatial components. Finally, a binarized mask derived from the spatial component of each ROI was used to extract a demixed activity trace for each ROI from the matrix product of each ROI’s spatial and temporal components. ROIs were excluded from further analysis if they corresponded to axons in L1, basal dendrites in L5, or were lost intermittently during imaging sessions.

We assume that individual ROIs rarely included segments of dendrite from more than one neuron. Tuft activity is sparse (Francioni et al., 2019) and GCaMP8m kinetics are fast (Zhang et al., 2023), therefore, the activity of the tuft dendrites of two different neurons would have to be highly temporally correlated (> 0.8 Pearson correlation) to merge into a single ROI. Furthermore, components extracted from individual imaging sessions had consistent spatial structure (Figure 2—figure supplement 2F), suggesting that the structure of ROIs was robust to measurement noise and spurious within-session correlations. However, we do not assume that dendritic segments belonging to the same neuron always merged into one ROI. In particular, L5 ET neurons with early bifurcating apical dendrites can exhibit highly independent activity across two hemi-trees (Otor et al., 2022) and may be segmented into two ROIs by our approach. In summary, different ROIs are putatively the dendrites of different neurons or highly independent dendritic compartments of the same neuron.

Deconvolution

An initial ΔF/F₀ was calculated from each ROI’s demixed activity trace, where F₀ was the 20^th percentile of F within a rolling 100 s window. Residual contamination of the estimated F₀ by periods of high activity was corrected by iteratively identifying periods of activity (> 1.0 s.d. of the noise) in a high-pass (1 Hz) filtered version of ΔF/F₀, and then excluding those periods from the next iteration of F₀ estimation. Corrected ΔF/F traces were then deconvolved using constrained FOOPSI (CaImAn; Giovannucci et al., 2019) set to fit an AR(2) model. Deconvolved activity below a noise threshold (defined by bimodal Gaussian fitting of all deconvolved values; python scipy, Virtanen et al., 2020) was set to zero. The resulting graded “impulse” traces were used for all subsequent analysis.

Trial-alignment and ROI grouping

Individual trial traces were aligned to either the GO cue frame (green line/label in all figures) or the first lickport contact frame (FL; dark yellow line/label in all figures). Trial-type averages were calculated using the mean of N = 1000 bootstraps with replacement. For Figure 3G,H, ROIs were excluded if they had N < 10 pre-shift CR trials, N < 10 pre-shift CL trials, or r < 0.5 correlation between trial averages of 50/50 trial count splits (Yang et al., 2022). In Figure 3G,H, ROIs were classified and sorted according to the timing of peak activity and whether this peak activity was in CR or CL GO-aligned averages. For Figure 5D,F, ROIs were excluded if they had N < 10 CA trials and sorted according to mean CR−CA activity between 0 s and 1 s relative to first-lick contact (see Figure 5—figure supplement 1).

Identification of GO- and contact-associated activity

The deconvolved activity of each ROI was modeled as the sum of GO- and contact-aligned impulse response functions (IRFs). In brief, vectors marking GO-cue times and first lick tongue contact times (FL) across all pre-shift CR or pre-shift CL trials were converted into lagged Toeplitz design matrices. These design matrices were then fit to the activity of each ROI using non-negative linear regression and the resulting coefficients were taken as the estimated IRFs. To compare the contribution of GO- and contact-associated activity with respect to a common temporal alignment, all contact-IRFs for a given animal were shifted by the temporal offset between the GO-cue and the mean FL time for that trial-type (CR or CL) and animal. Thus, both GO- and contact-associated trial-aligned activity were realigned to the timing of the GO-cue. Mean CR and mean CL activities were then averaged together, generating a single GO-associated activity trace and a single contact-associated activity trace for each ROI. To compare relative peri-GO activity across dendrites and somata (Figure 4B,C), activity was divided by the mean of both GO-associated and contact-associated activity during the first half (0.625 s) of the delay epoch.

Coding directions and projections of task-variables

The mean activity of ROIs during each trial type was modeled as a linear combination of three task variables: sensory cue, choice, and outcome (Chen et al., 2017; Yang et al., 2022). At each timepoint t in the trial, three task-variable selectivity CDs were defined as:

where x_k(t) denotes the trial-averaged activity for trial type k at time t. For each timepoint, selectivity CDs were then orthogonalized by QR decomposition in the order O(t), c(t), s(t).

We then projected trial-averaged activity at t onto the corresponding s(t), c(t), and O(t) CDs. To cross-validate projections, CDs were calculated from a randomly selected 50% of trials and then the remaining 50% of trials were projected onto the CD. Population selectivity along each CD at t was then calculated by replacing x_k(t) in Equations 2–4 with the corresponding projections of trial-averaged activity along each CD. For estimation of SEM and statistical comparisons, data were hierarchically bootstrapped at the top level (animals) N_boot = 10000 times and 50/50 trial splits were recalculated each bootstrap iteration.

Coding directions and projections that discriminate trial types

To compare population dynamics between two trial types, we calculated the CDs along which activity maximally discriminated between the trial types (Stopfer et al., 2003; Li et al., 2016). CDs were defined as:

where x_A(t) and x_B(t) denote the trial-averaged activity for trial types A and B at time t.

To calculate selectivity projections, x_A(t) and x_B(t) were projected onto CD(t) and then subtracted from each other. All projections were cross-validated by calculating CDs from a randomly selected 50% of trials and then projecting onto the remaining 50% of trials. For estimation of SEM and statistical comparisons, data were hierarchically bootstrapped at the top level (animals) N_boot = 10000 times and 50/50 trial splits were recalculated each bootstrap iteration.

Coding Magnitude and Correlation

To quantify the overall strength of population selectivity, we computed the magnitude of the coding direction at each timepoint t as the L2-norm of the CD vector.

To quantify the correlation between CDs across training epochs, we calculated the cosine similarity between the mean-subtracted CDs, which was equivalent to the Pearson correlation. Empirical measurements of correlation will be systematically lower than true underlying correlations due to noise. To place across-epoch correlations into context, we estimated a maximum observable correlation (i.e., correlation “ceiling”) based on the within-epoch reliability of the CD estimate. Within-epoch reliability was defined as the correlation between two CDs derived from a random 50/50 trial-level split of the epoch. The maximum observable cross-epoch correlation was defined as:

where r¹(t) and r²(t) denote the within-epoch correlations of training epoch 1 and training epoch 2. The cross-epoch correlation (r_cross(t)) was then normalized by this correlation ceiling to obtain a reliability-corrected correlation estimate:

Selectivity Index, Responsiveness, and Net Selectivity

To quantify the selectivity of individual ROIs (Figure 6C-H), we first calculated the mean trial-averaged activity of each ROI during non-overlapping 0.5 s bins spanning the trial duration. At each time bin t we computed a Selectivity Index (SI) for ROI and for each training epoch (pre-shift or post-shift):

where x_CR(t) and x_CL(t) are the trial-averaged activities at time bin t for CR and CL trials, respectively. The 95% confidence intervals (CIs) for each SI(t) were estimated by resampling trials with replacement (N_boot = 10000). When the number of CR and CL trials differed across training epochs, trials were randomly subsampled to ensure matching trial counts and balanced estimates. SI(t) estimates with 95% CI width > 0.66 or minimal total activity (x_CR(t) + x_CL(t) < 0.1 impulses) were considered unreliable and were excluded. Within a training epoch, ROIs with any reliable SI(t) estimates were considered “task-modulated”. For each ROI, the time bin t across both training epochs that had the largest difference in activity between the two trial types (x_CR(t) − x_CL(t)) and a reliable SI(t) was defined as the characteristic response time, t_response, for that ROI. The final pre-shift SI and post-shift SI for each ROI were defined as SI(t_response) drawn from the respective SI(t) estimates of each epoch.

The characteristic response time (t_response) was then used to calculate the Responsiveness and Net Selectivity for each ROI:

and

Sample Sizes

Relevant sample sizes can be found in the appropriate figure legends. Overall sizes for the numbers of animals and ROIs can be found in Table 1 while numbers of sessions and trials can be found in Table 2 and Table 3 respectively.

Table 1.

Table 2.

Table 3.

Binning and Statistics

In all plots, the dendrite and soma data were binned with matching duration time bins. Data were binned at: 44 ms (Figure 3; Figure 4D,E; Figure 5C,F), 90 ms (Figure 4B), 224 ms (Figure 4F; Figure 5D,F,H,I; Figure 5—figure supplement 1; Figure 5—figure supplement 2; Figure 6I,J; Figure 6— figure supplement 1B; Figure 6—figure supplement 2B; Figure 6—figure supplement 3B) or 493 ms (Figure 6C-G). Unless otherwise noted, data are presented as bootstrap mean ± standard error of N_boot = 1000 bootstraps for trace averages and N_boot = 10000 bootstraps for statistical tests. Unless otherwise noted, * indicates p < 0.05 estimated by hierarchical bootstrap at the level of animals. Horizontal gray dashed lines indicate the comparison value for significance tests. Corrections for multiple comparisons were performed using a Holm–Bonferroni method. Comparisons between probability distributions were performed using a Kolmogorov-Smirnov (K-S) test (Figure 2—figure supplement 3B,C).

Figure supplements

Figure 1—figure supplement 1.

Figure 2—figure supplement 1.

Figure 2—figure supplement 2.

Figure 2—figure supplement 3.

Figure 4—figure supplement 1.

Figure 5—figure supplement 1.

Figure 5—figure supplement 2.

Figure 6—figure supplement 1.

Figure 6—figure supplement 2.

Figure 6—figure supplement 3.

Data availability

Code and data intermediates used in the generation of this manuscript are available at https://github.com/kerlin-lab/Scheib_2026. The complete raw image series are approximately 40 TB and will be available on request.

Acknowledgements

We thank Timothy Ebner, Matthew Chafee, and Thomas Naselaris for comments on the manuscript. This work was supported with funding from the Whitehall Foundation (Grant 2020-05-71), NINDS R01NS127902, as well as the resources and staff at the University Imaging Centers (UIC; RRID:SCR_020997) and Minnesota Supercomputing Institute (MSI) at the University of Minnesota.

Note

This reviewed preprint has been updated to correct the GitHub URL in the Data availabilty.

Additional information

Funding

HHS | NIH | National Institute of Neurological Disorders and Stroke (NINDS) (R01NS127902)

• Aaron Kerlin

Whitehall Foundation (The Whitehall Foundation) (2020-05-71)

• Aaron Kerlin