The cerebellum implements structured representation of valence to support adaptive behavior control.

The cerebellum implements structured representation of valence to support adaptive behavior control - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Adv . 2026 May 29;12(22):eaeb5860. doi: 10.1126/sciadv.aeb5860 The cerebellum implements structured representation of valence to support adaptive behavior control Wataru Koyama 1, † , Yuki Tanimoto 2, 3 , Tanvir Islam 2 , Makio Torigoe 2, ‡ , Takashi Shimizu 1 , Takashi Kawashima 4 , Hitoshi Okamoto 2, 3, 5 , Masahiko Hibi 1, * Author information Article notes Copyright and License information 1 Department of Biological Science, Graduate School of Science, Nagoya University, Nagoya 464-8602, Japan. 2 RIKEN Center for Brain Science, Wako 351-0198, Japan. 3 Faculty of Science and Engineering, Waseda University, 2-2 Wakamatsu-cho, Shinjuku-ku, Tokyo 162-8489, Japan. 4 Weizmann Institute of Science, Rehovot, Israel. 5 Institute of Neuropsychiatry, 91 Benten-cho, Shinjuku-ku, Tokyo 162-0851, Japan. * Corresponding author. Email: hibi.masahiko.s7@f.mail.nagoya-u.ac.jp † Present address: Neurobiology Department, School of Biological Sciences, University of California, San Diego, La Jolla, CA 92093, USA. ‡ Present address: Bristol Myer Squibb K.K., 1-2-1 Ohtemachi, Chiyoda-ku, Tokyo 100-0004, Japan. Roles Wataru Koyama : Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Software, Validation, Visualization, Writing - original draft Yuki Tanimoto : Methodology Tanvir Islam : Software Makio Torigoe : Methodology, Resources, Software, Validation, Writing - review & editing Takashi Shimizu : Funding acquisition, Writing - review & editing Takashi Kawashima : Formal analysis, Methodology, Software, Supervision, Visualization, Writing - review & editing Hitoshi Okamoto : Conceptualization, Methodology, Resources, Software, Supervision, Validation, Writing - original draft, Writing - review & editing Masahiko Hibi : Conceptualization, Funding acquisition, Project administration, Resources, Supervision, Writing - original draft, Writing - review & editing Received 2025 Aug 18; Accepted 2026 Apr 16; Collection date 2026 May 29. Copyright © 2026 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original U.S. Government Works. Distributed under a Creative Commons Attribution NonCommercial License 4.0 (CC BY-NC). This is an open-access article distributed under the terms of the Creative Commons Attribution-NonCommercial license , which permits use, distribution, and reproduction in any medium, so long as the resultant use is not for commercial advantage and provided the original work is properly cited. PMC Copyright notice PMCID: PMC13220880  PMID: 42213837 Abstract Animals are thought to form internal models in the brain representing safety and danger, enabling avoidance through learned associations. Although the cerebellum has been implicated in active avoidance, its role in constructing internal models remains unclear. Here, we combined calcium imaging and closed-loop virtual reality in adult zebrafish to examine cerebellar circuits underlying avoidance learning. Fish learned to associate visual cues with safety or danger in a cerebellum-dependent manner using electric shocks as instructive signals. Cerebellar output neurons and Purkinje cells in the dorsal cerebellum developed cue-selective responses. Output neurons encoded both valence states, with danger-preferring cells prevalent, whereas Purkinje cells showed stronger danger bias. After cue reversal, neural selectivity adapted to the new rule. Safety- and danger-selective output neurons were spatially segregated, indicating anatomically dedicated units for valence processing. Optogenetic inhibition of shock-related inputs to Purkinje cells impaired rule updating. Together, our results indicate that the cerebellum forms an internal model of safety and danger, enabling adaptive avoidance. Cerebellar circuits identify danger and safety to guide flexible, adaptive behavior. INTRODUCTION Animals associate aversive stimuli with sensory cues, enabling them to avoid danger before it occurs—a process known as active avoidance conditioning. In mammals, this learning involves various brain regions, including the amygdala, medial prefrontal cortex (mPFC), and nucleus accumbens (NAc) ( 1 – 8 ). In zebrafish, telencephalic regions analogous to the mammalian amygdala and isocortex, along with the habenula-raphe circuitry, have also been implicated in active avoidance ( 9 – 12 ). Recent studies suggest that the cerebellum also contributes to this process ( 13 – 15 ). The cerebellum performs diverse functions through conserved neural circuits across vertebrates ( Fig. 1A ) ( 16 ). Purkinje cells (PCs) and granule cells (GCs) are the major GABAergic and glutamatergic neurons in the cerebellum, respectively. GCs receive inputs from mossy fibers (MFs) that originate in various precerebellar nuclei and convey diverse sensory and motor-related information. These signals are transmitted to PCs via parallel fibers (PFs), either directly or via interneurons, providing anticipatory contextual signals. PCs also receive excitatory climbing fiber (CF) inputs from the inferior olive, which encode instructive error signals; CF activity can also reflect positive reinforcement signals ( 17 , 18 ). By integrating PF and CF inputs, PCs regulate the activity of glutamatergic output neurons—deep cerebellar nuclei (DCNs) in mammals or eurydendroid cells (ECs) in zebrafish—which transmit information to other brain regions. While DCNs receive direct MF and CF inputs, ECs receive inputs from PFs and possibly CFs. Despite anatomical differences between species, the integration of MF- and CF-derived signals by PCs—and the relay of this information through output neurons—is a conserved feature across vertebrates. Fig. 1. Cerebellum-dependent control of active avoidance in a closed-loop VR task in adult zebrafish. Open in a new tab ( A ) Schematic of cerebellar circuitry. ( B ) Schematic of the VR setup for head-fixed zebrafish. ( C ) GO/NOGO task in the VR environment. In the NOGO task, fish received a 1-s electric shock if they exited the safety cue (red) area within 10 s. In the GO task, fish received a 1-s shock if they failed to enter the safety cue (red) area from the danger cue (blue) area within 10 s. Trials with shock were classified as failures; those without, as successes. ( D ) Experimental paradigm. The adaptation session included no shocks, followed by three training sessions with shocks delivered at the end of failed trials. ( E ) GO/NOGO task success rates in control WT ( n = 17) and PC-ablated (aldoca:NTR, n = 12) fish. Rates were calculated using a ±3 trial window per task type. The absence of preceding trials at the start of each session results in a brief gap in the moving-average window, which can appear as a step shift in the plotted values. Data include all analyzed fish (learners and nonlearners). Some learner fish exhibited transient decreases in performance after reaching the learning criterion, but overall success rates increased across sessions. Lines and shaded areas: means ± SE. ( F ) Comparison of success rates across training sessions (GO task: P = 0.00353 for training 1; P = 0.00581 for training 2; P = 0.00220 for training 3; Mann-Whitney U test). ** P < 0.01; ns, not significant. Small dots, individual values; large dots and lines, overall means ± SE. The cerebellum participates in a variety of learning paradigms by constructing internal models that simulate the expected consequences of actions before they are executed, enabling anticipatory adjustments and efficient control. In motor learning, such internal models are thought to predict and refine motor output by minimizing sensory prediction errors ( 19 – 22 ). Ito ( 23 ) proposed that similar internal model mechanisms may underlie mental functions to guide behavior. Cerebellar circuits have been implicated in forms of learning beyond motor adaptation. For example, Pavlovian eyeblink conditioning is a well-known example of cerebellum-dependent associative learning ( 17 , 24 – 27 ). In this paradigm, a conditioned stimulus (CS), such as an auditory or visual cue, is paired with an unconditioned stimulus (US), such as a corneal shock, leading the animal to respond to the CS alone. CS information is mainly conveyed to PCs via MFs (and to some extent via CFs), whereas US information is primarily conveyed through CFs. Similarly, in operant conditioning—where rewards guide the association of sensory cues with motor behavior—CFs convey instructive error signals that predict reward or reward omission ( 28 – 31 ), and GCs are shown to encode the expectation of reward ( 32 ). These findings support the view that cerebellar circuits integrate convergent inputs to form internal models that guide learned behavior. However, it remains unclear how the cerebellum integrates inputs to construct internal models representing motivational or emotional valence—such as safety, danger, or reward—and uses them to guide adaptive behavior. Zebrafish have been widely used to investigate cerebellar function. Studies using larvae have shown that PCs are functionally heterogeneous ( 33 , 34 ), that distinct cerebellar regions contribute to different motor control aspects ( 35 ), and that cerebellar circuits are involved in learning-dependent motor planning and adaptation ( 36 , 37 ). These findings support the view that the zebrafish cerebellum implements internal models using sensory feedback to regulate motor behavior. Zebrafish can acquire active avoidance conditioning from larval to adult stages, although learning efficiency varies depending on type of aversive US and experimental configuration ( 9 , 12 , 14 , 15 , 38 ). In these paradigms, fish learn to perform an action to avoid an imminent threat, typically associating a sensory cue with electric shock, aversive heat stimulation, or tactile stimuli. We previously demonstrated that, in adult zebrafish, inhibiting GCs or PCs by expressing botulinum toxin to block neurotransmitter release, or ablating PCs, suppresses active avoidance conditioning ( 13 ). These findings suggest that cerebellar circuits form an internal model representing danger and safety to guide escape responses. Recently, a closed-loop virtual reality (VR) paradigm was developed to enable behavioral experiments in freely swimming fish, where environmental color cues serve as CSs and electric shocks as USs ( 11 ). In the present study, we combined closed-loop VR–based learning, functional Ca 2+ imaging, and optogenetics to examine how cerebellar circuits contribute to constructing internal models that drive adaptive avoidance behavior. RESULTS Active avoidance is dependent on cerebellar neural circuitry We used a slightly modified version of a previously reported closed-loop VR setup for head-fixed adult zebrafish ( Fig. 1B ) ( 11 ). Fish were immobilized using a custom-made harness and placed in a small water tank surrounded by liquid crystal display (LCD) screens on the front, left, and right sides, forming an immersive VR environment. Tail movements were recorded by an overhead infrared camera and translated into forward virtual motion, causing the background images on the screens to shift backward in real time, thereby simulating active swimming in a closed-loop system. In this environment, zebrafish were trained in a GO/NOGO active avoidance task, in which background colors served as visual cues: Blue signaled danger, and red indicated safety ( Fig. 1C ). In GO tasks, fish had to swim from the danger cue (blue) area to the safety cue (red) area within 10 s to avoid a 1-s electric shock. In NOGO tasks, fish were required to remain in the safety cue (red) area for 10 s and received a 1-s shock if they entered the danger cue (blue) area. An initial adaptation session without shock was followed by three training sessions, each consisting of 30 randomized GO and NOGO trials, with shocks administered upon failure ( Fig. 1D ). Consistent with the previous report ( 11 ), the success rate of NOGO trials was high from the first training session. In contrast, the success rate of GO trials was low in the first training session but progressively increased across training sessions. Fish were classified as learners if they successfully achieved at least 8 of 10 consecutive trials in both GO and NOGO tasks at any point during training sessions (fig. S1A). Across all tested fish ( n = 181), 46.4% met this criterion under closed-loop VR conditions, consistent with previous findings ( 11 ). No significant differences in learning performance were observed between nacre (melanophore-deficient, mitfa − / − ) and casper (melanophore- and iridophore-deficient, mitfa − / − ; roy − / − ) strains or between male and female fish (fig. S1, B and C). On the basis of these results, all subsequent analyses were conducted without separating fish by strain or sex. We compared the travel distances of learners and nonlearners during the adaptation session, in which GO and NOGO color cues were presented without shock. Although no statistically significant difference was detected between the two groups (fig. S2), nonlearners tended to exhibit higher travel distances than learners. These results suggest that baseline behavioral differences were not robust at the statistical level; however, subtle differences in baseline activity or spontaneous behavior may exist and could potentially influence subsequent learning performance. Under the head-fixed condition, fish often ceased tail movements, making it difficult to conclude from NOGO trial performance alone that fish actively learned to remain in the safety cue (red) area. However, we detected learning-dependent changes in cerebellar neural activity following repeated NOGO trials and also observed an increased success rate in NOGO trials after reversal of the danger/safety color cues, as described below. To examine cerebellar involvement in active avoidance learning, we tested GO/NOGO task performance in fish with ablation of PCs. We ablated PCs using the transgenic line Tg(aldoca:epNTR-TagRFP) ( 13 ), which expresses nitroreductase (NTR) specifically in PCs under the control of the enhancer and promoter of the PC-specific gene aldolase Ca ( aldoca ), a PC-specific gene. Aldoca is one of the two zebrafish orthologs of aldolase C and corresponds to zebrin II in mammals ( 39 ). In mammals, zebrin II is expressed in a longitudinally striped pattern in PCs ( 40 ), whereas in zebrafish, aldoca is uniformly expressed in all PCs ( 41 , 42 ). PC ablation was induced by metronidazole (MTZ) treatment. PC-ablated fish generated using the same aldoca:NTR transgenic strategy were previously shown to exhibit slower swimming speeds compared to wild-type (WT) controls during free swimming while showing no difference in turning frequency. Moreover, these fish displayed no significant differences in responsiveness to electric shock or in swimming behavior following shock delivery ( 13 ). Compared to WT controls ( n = 17), PC-ablated fish ( n = 12) exhibited significantly impaired GO task learning, with consistently lower success rates across all three training sessions ( P = 0.00353 for training 1; P = 0.00581 for training 2; P = 0.00220 for training 3; Fig. 1, E and F ). To confirm that the observed difference in GO task learning was not attributable to random trial fluctuations, we performed a permutation analysis (1000 shuffles). For each shuffled dataset, fish identities were randomly reassigned, and we assessed whether a statistically significant difference between WT and PC-ablated fish ( P < 0.05) was detected. This procedure was repeated three times independently. Across these analyses, only 22 to 33 of 1000 shuffled datasets showed a significant WT > aldoca:NTR difference (2.2 to 3.3%), indicating that the impairment in GO task learning in PC-ablated fish is unlikely to result from random trial-to-group associations. These results indicate that cerebellar circuits involving PCs are essential for active avoidance learning in adult zebrafish using the VR setup. Distinct responses of cerebellar output neurons to safety and danger cues during the GO/NOGO task We next performed Ca 2+ imaging of ECs, the cerebellar output neurons, using learner fish ( n = 10; Fig. 2A and movie S1). To express the Ca 2+ indicator GCaMP6s, we used the transgenic line hspzGFFgDMC156A ( 43 ), which expresses a modified version of GAL4 in ECs (referred to as EC-GAL4FF ), together with Tg(5xUAS-hsp70l:GCaMP6s) ( 44 ). We imaged ~150-μm deep in the dorsal cerebellum during the GO/NOGO active avoidance task ( Fig. 2A ). Imaging was performed on all fish without preselecting learners, and learning status was determined retrospectively from behavioral performance during the three training sessions. Only fish that met the learner criterion by the end of training session 3 were included in the EC imaging analyses. During adaptation session, ECs showed no clear responses to either safety (red) or danger (blue) cue. However, by the end of training session 3, distinct EC populations had developed selective responses to these cues ( Fig. 2B ). Fig. 2. Distinct responses of cerebellar output neurons (ECs) to safety and danger cues during the GO/NOGO task. Open in a new tab ( A ) Representative Ca 2+ imaging from the dorsal cerebellum of a transgenic zebrafish expressing GCaMP6s in ECs [ EC-GAL4FF; Tg(5xUAS-hsp70l:GCaMP6s) ]. Colored areas indicate detected somatic ROIs. Scale bar, 100 μm. ( B ) Example behavioral and neural data from adaptation and training session 3. Top: wall color in the VR environment where the fish was located. Middle: normalized Δ F / F ( z -score) of ECs sorted by principal components analysis (PCA). Bottom: tail movement intensity. ( C ) Red and blue weights for each EC, computed via Poisson GLM with regressors for time spent in red or blue areas ( n = 804 cells, 10 fish). Dots represent individual cells, colored by red-blue weight balance at training 3. ECs with high red and strong negative blue weights were classified as safety ECs (green); those with high blue and strong negative red weights as danger ECs (purple). ( D ) Mean Δ F / F of safety ECs (253 cells) and danger ECs (551 cells) during successful NOGO, successful GO, and failure GO trials in the adaptation or training session 3 (10 fish). The bottom shows the cumulative average travel distance from the trial onset in the VR environment. Red/blue bar, area colors; yellow bar, shock period. a.u., arbitrary units. To further characterize this activity, we applied a Poisson generalized linear model (GLM) to the neural responses of individual ECs (fig. S3, A to C). Each behavioral variable—such as exposure to safety (red) or danger (blue) cue—was treated as a regressor, representing a specific aspect of the animal’s experience. The model was fitted to the observed neural activity by adjusting the weight of each regressor, which quantifies how strongly each cue influences neural responses. A positive weight indicates increased activity in response to the cue, whereas a negative weight indicates suppression. Accordingly, a high red weight reflects strong activation by the safety cue, while a strongly negative blue weight indicates suppression by the danger cue. Although fluorescence signals can appear sustained across cue transitions due to the slow decay of GCaMP6s, we determined cue selectivity based on the timing of deconvolved neural activity. This approach allowed identifying which cue period most strongly accounted for each cell’s response. By tracking the red and blue weights of all ECs from learner fish across sessions, we found that the initially nonselective population observed during the adaptation session gradually segregated into two distinct subgroups during training sessions: one with increasing red weights and more negative blue weights and the other with increasing blue weights and more negative red weights, reflecting the emergence of color-selective responses through learning ( n = 804 cells, 10 fish) ( Fig. 2C ). On the basis of their weight profiles at training session 3, ECs with relatively higher red weights and more negative blue weights were defined as “safety ECs” (253 cells), whereas those with relatively higher blue weights and more negative red weights were defined as “danger ECs” (551 cells) ( Fig. 2C ), indicating the emergence of both subtypes, with danger ECs being more numerous. To evaluate the robustness of the GLM analysis with respect to trial history, we examined whether baseline fluorescence elevations following electric shocks influenced weight estimation. We therefore excluded trials immediately following shock delivery or trial outcomes. As shown in fig. S4, these exclusions had no substantial effect on the distribution of danger and safety ECs, indicating that trial history did not affect the GLM results. To examine whether these valence-selective EC populations emerged consistently across individuals, we analyzed GLM red and blue weights for each fish separately. As shown in fig. S5, all learner fish ( n = 10) exhibited the emergence of both danger and safety EC populations, with similar relative proportions and temporal emergence patterns across training sessions, indicating that the observed segregation of valence-selective ECs is a robust and reproducible feature and not driven by a subset of individuals. Analysis of the temporal dynamics of EC activity (Δ F / F ) revealed that, as training progressed, danger ECs became active immediately upon entering the danger cue (blue) area during both failed and successful GO trials and were rapidly suppressed upon entering the safety cue (red) area during successful NOGO and GO trials. In contrast, safety ECs were activated upon entering the safety cue area during successful NOGO and GO trials and showed reduced activity in the danger cue area across both GO and NOGO trials ( Fig. 2D and fig. S3, D and E). Notably, danger ECs remained active even during failed GO trials, in which fish stayed in the danger cue area without moving ( Fig. 2D , right). Likewise, safety ECs maintained sustained activity during successful GO trials, even when fish continued swimming after entering the safety cue area ( Fig. 2D , middle). These findings suggest that ECs do not directly control locomotor output but instead provide instructive signals that guide behavior. Even among nonlearner fish ( n = 735 cells, 11 fish), distinct populations of danger and safety ECs were observed by training session 3, with one exception: a single fish that failed to succeed in any GO trials throughout training session 3 (figs. S6 and S7). Analysis of danger and safety cue weight dynamics across all nonlearner fish revealed that, with the exception of this single individual, nonlearners generally developed separable danger and safety EC populations over training (fig. S6). Figure S7 highlights two extreme nonlearner fish that showed no successful GO trials in training session 3 and illustrates their atypical weight trajectories and neural activity patterns. We next performed a more detailed comparison between learners and nonlearners (fig. S8). Although the overall proportions of danger and safety ECs did not differ significantly between groups (fig. S8A), subpopulation analysis revealed significant differences. In nonlearners, there was a reduction in danger ECs that were strongly suppressed by safety cue and in safety ECs that were strongly activated by safety cue (fig. S8, B and C). Consistent with this, safety ECs exhibited reduced neural activity in response to safety cue during training session 3 (fig. S8E). These findings indicate that, although nonlearners are generally capable of forming danger and safety EC populations, their responses to safety cue are weaker. Such reduced responses to safety cue may underlie the behavioral differences observed between learners and nonlearners. Cerebellar output neurons adapt to a new rule and are spatially segregated in the cerebellum To assess neural responses to a rule change, we introduced two reversal sessions following the training sessions, in which the association between color cues and electric shocks was reversed, with red now signaling danger and blue signaling safety ( Fig. 3A ). Using the GLM, we tracked how ECs, previously classified as danger or safety ECs based on their red and blue weights in training session 3, changed their responses in reversal sessions 1 and 2. We found that 89.7% of danger ECs became activated when fish were in the danger cue (red) area and suppressed in the safety cue (blue) area, while 81.0% of safety ECs were activated in the safety cue (blue) area and suppressed in the danger cue (red) area ( Fig. 3, B and C ). These shifts in color selectivity became more pronounced in reversal session 2 compared to reversal session 1 ( Fig. 3, D and E ), indicating that ECs gradually adapt to the new task rule through continued learning after rule reversal. Example trajectories from individual danger and safety ECs across adaptation, training sessions 1 to 3, and reversal sessions 1 and 2 (fig. S9, A and B) illustrate that cue-related activity patterns evolve progressively as animals experience the changing task contingencies. These single-cell trajectories highlight dynamic modulation of EC activity across both learning and reversal phases. Fig. 3. Danger and safety ECs adapt to a reversal rule and are spatially segregated in the cerebellum. Open in a new tab ( A ) GO/NOGO task design with rule reversal. After adaptation and training sessions, color-valence association was reversed: Red (safety) and blue (danger) became red (danger) and blue (safety). In each of reversal sessions 1 and 2, 30 randomized GO and NOGO tasks were conducted with the reversal rule. ( B ) Red and blue weights of individual ECs ( n = 804 cells, 10 fish) across training session 3 and reversal sessions 1 and 2. Dots represent individual cells, colored by red-blue weight balance at training 3. ( C ) Proportions of danger ECs maintaining high weight for the danger color and strong negative weight for the safe color [red triangles in (B)] and safety ECs maintaining high weight for the safe color and strong negative weight for the danger color in reversal session 2 [blue triangles in (B)]. ( D and E ) Changes in red and blue weights across sessions for danger ECs (D) and safety ECs (E). For each fish, mean red and blue weights were calculated and then averaged across all fish. Dots and lines show overall means ± SE. ( F ) Example spatial distribution of danger and safety ECs in the dorsal corpus cerebelli. Dot colors follow classification in Fig. 2 . Dashed line indicates the midline. ( G ) Comparison of the average lateral-axis kernel density of safety EC and danger EC neurons (10 fish). Thin lines represent the distributions for individual fish. ( H ) Distance from each cell’s center to the midline ( P = 0.00117; Welch’s t test). Small dots, per-fish means; large dots and lines, overall means ± SE. To better characterize how neural activity evolves within a session on a finer timescale, we next analyzed danger and safety EC responses across trial blocks (20 trials each) during reversal session 1 (fig. S9, C and D). This block-wise analysis revealed progressive changes in cue-related activity across trials. Notably, safety ECs exhibited a particularly rapid inversion of their responses to the newly assigned danger cue, with clear reversal observed already within the first 20 trials—indicating a faster updating of cue-response mapping compared with danger ECs. Consistent with these neural activity-based findings, GLM analysis across the same trial blocks (fig. S10, A and B) demonstrated that safety ECs showed earlier and more pronounced changes in cue selectivity than danger ECs, further highlighting subtype-specific differences in adaptation speed during reversal learning. A small fraction of ECs (~20%) did not switch their cue preference during reversal. Most of these neurons showed red and blue weights of similar magnitude, and their overall color cue-related weight magnitudes were lower than those of adaptive ECs (fig. S11), suggesting weak cue differentiation. These properties raise the possibility that this subset of ECs may play a limited role in rule updating during learning. We examined the spatial distribution of EC subtypes in the dorsal cerebellum. Although the two populations showed broadly overlapping distributions, safety ECs tended to be located more laterally, whereas danger ECs were slightly more concentrated toward the cerebellar midline ( Fig. 3, F to H , and figs. S6, B and C, and S12). Distinct populations of PCs respond to safety and danger environment We performed Ca 2+ imaging of PCs using a transgenic zebrafish line Tg(aldoca:GAL4FF, myl7:mCherry) , which expresses GAL4FF specifically in PCs ( 43 ) (referred to as PC-GAL4FF ), in combination with Tg(5xUAS-hsp70l:GCaMP6s) ( 44 ). Since PC dendrites extend dorsally and exhibit much stronger GCaMP6s signals than somata, fluorescence from these dendrites obscures somatic signals under spinning disk confocal imaging. Accordingly, we recorded Ca 2+ signals from dendritic regions, which are more accessible for stable imaging. Imaging was conducted during the GO/NOGO active avoidance task, capturing dendritic regions located 50 to 150 μm beneath the dorsal cerebellar surface ( Fig. 4A and movie S2). As with ECs, we found that by the end of training session 3, distinct PC dendrites exhibited selective responses to visual cues associated with danger (blue) or safety (red) ( Fig. 4B ). We defined “danger PC” dendrites as those activated in the danger cue (blue) area and “safety PC” dendrites as those activated in the safety cue (red) area, based on GLM analysis (safety PC dendrites, n = 1014; danger PC dendrites, n = 6404, 6 fish; Fig. 4C ). Although some PC dendrites showed activity during both danger and safety cue periods, we classified each dendrite based on the relative magnitude and sign of its cue-related GLM weights, such that the cue period producing the stronger modeled response determined whether it was assigned as a danger or safety PC. Unlike ECs, however, danger PCs were not strongly suppressed by the safety cue, and safety PCs were not markedly suppressed by the danger cue ( Fig. 4C and fig. S13A). Notably, while ECs exhibited a moderate bias toward danger representation, PC dendrites showed a stronger bias, with a greater proportion selectively responding to danger cues. Tracking these responses across sessions showed that PC dendrites, like ECs, progressively differentiated into these two subtypes during learning ( Fig. 4, D and E ). In the reversal sessions, PC dendrites also showed flexible adaptation to the new rule, with danger and safety PC dendrites gradually shifting their responses in accordance with the reversed color cue–shock association, similar to ECs ( Fig. 4, C to E , and figs. S9, E and F, and S13B). These findings indicate that the functional segregation observed in ECs is established at the level of the PCs that provide input to ECs. However, the difference in the relative proportions of danger and safety representation between PCs and ECs suggests the presence of additional regulatory mechanisms downstream of PCs, which may act to balance or reshape valence information at the EC level ( Figs. 2C and 4C ). Fig. 4. PC dendrites respond differentially to safety and danger cues during the GO/NOGO task. Open in a new tab ( A ) Ca 2+ imaging from the dorsal cerebellum of a transgenic line expressing GCaMP6s in PCs [ PC-GAL4FF; Tg(5xUAS-hsp70l:GCaMP6s) ]. Images were acquired at 50- to 100-μm depth, where PC dendrites are located. Colored areas indicate dendritic segments [region of interest (ROIs)]. Scale bar, 100 μm. ( B ) Example behavioral and neural data from adaptation session and training session 3. Top: wall color in the VR environment. Middle: normalized Δ F / F ( z -score). Bottom: tail movement intensity. ( C ) Red and blue weights of PC dendritic ROIs across sessions ( n = 7418 ROIs, six fish). Each dot represents one ROI; dot color reflects the red-blue weight difference in training 3. ROIs with higher red and strongly negative blue weights were classified as safety PCs (green); those with higher blue and strongly negative red weights as danger PCs (purple). ( D and E ) Changes in red and blue weights for danger (D) and safety (E) PCs across sessions. Dots and lines represent overall means ± SE. PC inputs, likely mediated by CFs from the inferior olive, are thought to act as instructive teaching signals that induce synaptic plasticity and modulate cerebellar circuit activity ( 17 , 26 – 30 , 45 ). To assess whether shock-related input signals function as teaching signals, we examined PC responses using Ca 2+ imaging. During failed GO trials, both safety and danger PC dendrites showed robust activation in response to electric shocks ( Fig. 5, A and B ). Danger PC dendrites exhibited stronger responses with a shorter latency to the electric shock (1.15 ± 0.249 s, mean ± SE), whereas safety PC dendrites exhibited a slightly longer latency (2.12 ± 0.301 s, mean ± SE) ( Fig. 5, A to D ). In contrast, ECs showed similar latencies between subtypes but stronger activation in safety ECs compared to danger ECs ( Fig. 5E ). Consistent with this, GLM analyses also revealed higher shock-related weights in safety ECs (fig. S14). Fig. 5. PCs and ECs respond to instructive teaching signals. Open in a new tab ( A and B ) Mean Δ F / F of danger (A) and safety (B) PCs during failed GO trials, aligned to trial onset. Blue bar, area color; yellow bar and dot window, shock period. ( C ) Definitions of the shock window and shock omission window. ( D and E ) Mean neural activity of PCs (D) and ECs (E) during the shock window across sessions, aligned to the baseline defined as the 1-s period immediately before the shock onset. ( F and G ) Mean neural activity of PCs (F) and ECs (G) during the shock omission window across sessions, aligned to the baseline at trial end. We further analyzed PC dendritic activity in reversal sessions. In successful NOGO trials, fish were presented with visual cues that had previously predicted electric shocks under the original rule, but no shocks were delivered. This omission of the expected shock likely generated a prediction error signal. Safety PC dendrites showed stronger activation than danger PC dendrites ( Fig. 5, C and F , and fig. S15, C and D), suggesting that safety PCs are particularly responsive to prediction errors. These findings indicate that danger and safety PCs may receive distinct types of teaching signals depending on context. A similar pattern was observed in ECs: Safety ECs were strongly activated by shock omission, whereas danger ECs were suppressed ( Fig. 5, C and G , and fig. S15, A and B). Immediately after the rule reversal, both ECs and PCs exhibited strong responses to the omission of the expected shock, with safety-selective neurons showing robust activation compared with danger-selective ones (fig. S15). In safety ECs, these omission-evoked responses declined markedly over subsequent trials as prediction errors diminished (fig. S15B), paralleling their early changes in cue selectivity seen in figs. S9D and S10 (A and B). This rapid attenuation suggests that the faster adaptation of safety ECs to the new rule may be driven by their stronger engagement with early prediction-error signals. Safety PCs showed a similar pattern, exhibiting larger omission responses that decreased as learning progressed (figs. S10, C and D, and S15D), further supporting the idea that prediction-error processing contributes to subtype-specific differences in reversal dynamics across both ECs and PCs. These findings highlight distinct responses to teaching signals between PCs and ECs during both electric shock and shock omission, suggesting that different mechanisms operate at each level of the cerebellar circuit. PC inputs are required to update the rule To examine the role of electric shock–related input to PCs, we optogenetically inhibited this input during reversal learning using Tg(aldoca:GtACR1-EYFP) fish, which express the anion channelrhodopsin Gt ACR1 ( Fig. 6A ) ( 46 ). Fish first underwent training sessions without optogenetic interference. During reversal sessions, 520-nm light was delivered to the cerebellum at the time of electric shock during failure trials in both GO and NOGO tasks to inhibit shock-related inputs to PCs ( Fig. 6B ). Control fish ( n = 14) successfully adapted to the reversal rule despite light illumination. In contrast, Tg(aldoca:GtACR1-EYFP) fish ( n = 13) showed significantly impaired adaptation in both GO and NOGO tasks during reversal sessions ( P = 5.17 × 10 −4 for the GO task in reversal session 1; P = 4.29 × 10 −5 for the GO task in reversal session 2; P = 0.00905 for the NOGO task in reversal session 1; P = 0.0341 for the NOGO task in reversal session 2; Fig. 6, C and D , and fig. S16, A and B). This impairment was most evident in the NOGO task during reversal session 1, particularly in fish that had performed well in the GO task during training session 3. In control fish, those showing high GO success rates in training session 3 also exhibited high NOGO success rates in reversal session 1, indicating effective adaptation to the rule change. However, this relationship was markedly disrupted in Tg(aldoca:GtACR1-EYFP) fish, in which optogenetic inhibition of PC input led to reduced NOGO success despite strong GO performance during training session 3 (fig. S16A). These results support the idea that electric shock–related input to PCs is essential for updating learned associations during reversal learning in the GO/NOGO task. Fig. 6. Optogenetic inhibition of error-related inputs to PCs impairs reversal learning. Open in a new tab ( A ) Expression of Gt ACR1-EYFP in the cerebellum of Tg(aldoca:GtACR1-EYFP) fish. A brain was stained with anti–green fluorescent protein (green) and anti-parvalbumin 7 (Pvalb7, magenta) antibodies. Inset: higher magnification views of the white boxed areas. Scale bars, 200 and 20 μm in the inset. ( B ) Experimental design for the GO/NOGO task with optogenetic interference. After an adaptation session, fish underwent three training sessions without optogenetic stimulation. During reversal sessions, 520-nm light (9 mW/mm 2 ) was applied to the dorsal cerebellum for 1 s at the end of each failed GO or NOGO trial. ( C ) Success rate progression in control fish ( n = 14) and Tg(aldoca:GtACR1-EYFP) fish expressing Gt ACR1 in PCs ( n = 13). Lines show means ± SE. ( D ) Comparison of success rates between groups across sessions ( P = 5.17 × 10 −4 for GO task in reversal 1; P = 4.29 × 10 −5 in reversal 2; P = 0.00905 for NOGO task in reversal 1; P = 0.0341 in reversal 2; Mann-Whitney U test). ( E ) Schematic model illustrating how safety and danger valence information is encoded in cerebellar circuits. During training, PCs receive visual inputs corresponding to danger (blue) and safety (red) cues and become functionally specialized through electric shock–evoked teaching signals. Distinct danger and safety PCs inhibit medial and lateral EC populations, respectively, generating danger and safety ECs that transmit valence information downstream. During reversal, cue-valence associations are switched, and shock-evoked and omission-related prediction error signals update PC selectivity. These updated PC signals reshape EC activity, enabling transmission of reversed valence information. Teaching and prediction error signals may also act directly on ECs, contributing to EC subtype updating. *** P < 0.001, ** P < 0.01, and * P < 0.05. DISCUSSION Cerebellar contributions to active avoidance learning In the present study, a closed-loop VR setup was used to examine active avoidance learning with GO and NOGO tasks. For the GO task, we found that PC-ablated fish showed reduced success rates, indicating that GO task learning depends on PCs. In contrast, the NOGO task exhibited a high success rate from the first training session, and PC-ablated fish did not show significant impairment ( Fig. 1 ). Because remaining in place is counted as success in the NOGO task, zebrafish that did not learn or those with mild motor impairment could still be classified as successful. Although previous work has shown that PC-ablated fish do not display abnormal responses to danger cues, their reduced swimming speed during free swimming ( 13 ) suggests that a potential influence of motor impairment cannot be completely excluded. Under these conditions, early-session performance alone provides limited leverage to rigorously assess learning-dependent cue selectivity in the NOGO task. However, we detected experience-dependent changes in cerebellar neural activity following repeated NOGO trials. During rule-reversal learning of the NOGO task, optogenetic inhibition of PCs reduced learning performance ( Fig. 6 ), supporting the idea that NOGO task learning also depends on cerebellar function. Using Ca 2+ imaging, we demonstrated that during active avoidance learning, distinct subpopulations of ECs and PCs in the dorsal cerebellum emerge that selectively respond to visual cues associated with danger and safety. We show that cerebellar circuits encode an internal model that integrates visual cues and electric shock–based teaching signals to represent these valence states, thereby guiding behavior during active avoidance. Although neurons representing either danger or safety have been identified in various brain regions, such as mPFC, NAc, and amygdala ( 47 – 50 ), our findings reveal a previously unrecognized feature of cerebellar organization: Opposing valence-specific subpopulations coexist within the same type of neurons yet are spatially segregated within the dorsal cerebellum. This suggests that the cerebellum implements a structured representation of valence in mental states to support adaptive behavioral control. In the present study, we focused our analysis on the dorsal cerebellum. Within this area, most of the ECs and PC dendrites exhibited selective responses to either danger or safety cues, albeit with varying degrees ( Figs. 2 and 4 ). It remains to be determined whether this widespread representation reflects a general involvement of cerebellar neurons in processing aversive stimuli such as electric shock, or whether neurons supporting active avoidance conditioning are preferentially localized within the dorsal cerebellum. Further investigation of ECs and PCs in the ventral and caudal regions may help clarify whether the dorsal cerebellum plays a specialized role in fear conditioning. Consistent with this possibility, the vermis at the dorsal midline of the mammalian cerebellum has been implicated in Pavlovian fear conditioning ( 51 , 52 ). The dorsal cerebellum of zebrafish may similarly serve as a key region for controlling fear-related responses. Spatial organization of valence-encoding circuits The somata of danger and safety ECs were spatially segregated, with danger ECs located medially and safety ECs located laterally in the dorsal cerebellum ( Fig. 3 ). In contrast, no clear spatial bias was observed for PC dendrites (fig. S13). Since PCs project to nearby ECs, one might expect that the somata of safety and danger PCs would also show distinct spatial localization. In mammals, PC dendrites are organized in planar structures along the anterior-posterior axis ( 53 , 54 ), whereas in zebrafish, at least during the larval stage, PC dendrites extend more broadly, including along the vertical axis ( 39 ). This anatomical difference may explain the lack of clear spatial segregation in our recordings. It is important to note that in this study, we measured PC activity only in dendrites, which primarily reflect CF and PF inputs, rather than intrinsic PC encoding. This limitation must be considered when interpreting the role of PCs in encoding valence, as dendritic Ca 2+ signals provide insights into sensory and teaching signal processing but do not directly reflect PC outputs. Future Ca 2+ imaging at the level of PC somata may provide further insight. Microzonal architecture and opponent valence representation The distinct spatial segregation of safety and danger ECs suggests the presence of physically separate neural circuits. In mammals and birds, both output neurons (in the DCN) and PCs exhibit region-specific gene expression patterns ( 54 – 59 ), and PCs receive topographically organized CF inputs from the inferior olive ( 60 , 61 ). A similar topographic projection pattern has been observed in larval zebrafish ( 62 ). These findings support the idea that the cerebellum contains functionally organized microcomplexes—sometimes referred to as “microzones”—characterized by spatially organized CF inputs, region-specific MF innervation, and longitudinally striped PC gene expression ( 53 , 56 ). Recent studies in mice have further proposed that such microzones can be functionally grouped into “upbound” and “downbound” microzones, which show opposite changes in PC simple-spike activity and plasticity during learning and exert bidirectional control over downstream nuclei and behavior ( 63 ). Our identification of spatially segregated danger and safety ECs, together with their upstream PC counterparts, suggests that a similar form of opponent, bidirectionally acting microzonal organization may operate in the zebrafish cerebellum to promote or suppress escape behavior depending on the predicted valence of the environment. Although longitudinal molecular markers such as zebrin II ( aldoca ) define PC stripes in mammals, such molecular patterns have not yet been identified in zebrafish PCs. Nonetheless, the medial-to-lateral localization of danger and safety ECs observed in this study, as well as the functional segregation of PC responses, may reflect the presence of latent functional microzones in the zebrafish cerebellum. This medial-lateral separation may reflect the mediolateral positioning of distinct microzones, which receive different CF and MF inputs and project to segregated output channels. In this framework, danger and safety ECs may simply inherit the spatial domains of their respective microzones. Under this model, danger and safety PC-EC pathways could represent zebrafish counterparts of upbound and downbound microzones described in mice, providing bidirectional control over downstream forebrain circuits involved in fear-related decision making. These microzones could be shaped by spatially organized inputs such as CFs and MFs, despite the apparent lack of molecular stripes. Further analysis of the afferent inputs that convey safety and danger cue information, including CFs that carry teaching signals, may help uncover the mechanisms underlying the formation and organization of these valence-specific circuits in the cerebellum. Teaching signals, prediction errors, and internal model formation Learning in the cerebellar circuitry is thought to depend on instructive teaching signals from CFs ( 17 , 26 – 30 , 45 ). In this study, we showed that inhibition of electric shock–related inputs to PCs disrupted adaptation to the rule change, and fish continued to behave according to the previously learned rule ( Fig. 6 ), suggesting that CF inputs associated with electric shocks contribute to learning during training. Although Ca 2+ imaging does not distinguish between inputs from GCs and CFs, prevailing models of cerebellar circuit function ( 21 , 64 ) propose that visual cues eliciting behavioral responses (e.g., escape or stay) are primarily conveyed via GCs but can also be transmitted through CFs, whereas feedback related to behavioral outcomes, such as electric shocks, is mainly conveyed through CFs. Danger PCs responded more strongly and with shorter latency to electric shocks than safety PCs ( Fig. 5 ), suggesting that stronger shock-related teaching signals may drive differentiation into danger PCs, while weaker signals may promote safety PC identity. In reversal sessions, safety PCs exhibited stronger activation than danger PCs when expected shocks were omitted ( Fig. 5 and fig. S15), implying that prediction errors resulting from shock omission may serve as teaching signals that shape safety PC identity. Together, these findings suggest that differences in the nature or strength of teaching signals—and/or the sensitivity of PCs to these signals—underlie their functional differentiation into danger- and safety-responsive subtypes. Here, safety ECs and safety PCs are defined as neurons whose activity corresponds to the safety cue and becomes elevated when the fish enters the safe zone, whereas danger ECs and danger PCs are activated in the danger zone. Given the GABAergic output of PCs, these observations suggest a circuit in which danger PCs suppress the activity of safety ECs, and safety PCs suppress danger ECs ( Fig. 6E ). Such reciprocal inhibition between valence-selective PC and EC subpopulations may contribute to the balanced and adaptive representation of safety and danger states during active avoidance behavior. These opponent interactions resemble the bidirectional control by upbound and downbound microzones described in mammals ( 63 ), in which distinct PC modules drive opposing changes in downstream output pathways during learning. We note that this representation is simplified, and the actual PC-EC interactions are likely to be more complex and influenced by additional inputs. Future studies that map these circuits with greater precision will be essential for elucidating the detailed computational mechanisms underlying cerebellar valence processing. In mammals, CFs have been shown to encode temporal-difference–like prediction error signals, reporting both unexpected delivery and unexpected omission of aversive stimuli during eyeblink conditioning ( 17 ), and coordinated CF activity can signal learned sensorimotor predictions rather than simple movement errors ( 30 ). Other studies have demonstrated that cerebellar GCs and PCs represent expected reward or subjective salience and translate these signals into readiness to act ( 32 , 65 ). In zebrafish, cerebellar circuits have been shown to implement internal models that calibrate feedback controllers in sensorimotor tasks and to support predictive computations for motor planning and shortened reaction times ( 36 , 37 ). Together, these studies establish that cerebellar circuits in vertebrates encode internal models and use error-related teaching signals to adapt behavior. In our task, we interpret the cue-related activity of danger and safety PCs and ECs during training sessions as valence-related expectation signals, reflecting the learned association between sensory cues and future outcomes. During the reversal sessions, however, the rule change introduces a clear mismatch between the previously learned cue–shock relationship and the new contingencies. Under these conditions, the shock-evoked and omission-related responses in PCs and ECs are consistent with prediction error signals that update the internal representation of safety and danger ( Fig. 5 and figs. S9 and S15). These updated valence representations then allow fish to adjust their behavior to the new rule. Thus, our findings demonstrate that cerebellar circuits in adult zebrafish use expectation signals during learning and prediction error signals during rule reversal to flexibly modify valence representations that support adaptive behavior. Integration of sensory and teaching signals in cerebellar output neurons ECs convey signals encoded by PCs to downstream targets and also respond strongly to electric shock–related teaching signals ( Figs. 2 and 5 and fig. S3). As ECs receive inputs from PCs, GCs, and possibly CFs ( 41 , 43 ), their activity likely reflects integrated information from these sources. Although direct CF inputs to ECs have not been demonstrated in zebrafish, anatomical studies in other teleosts and physiological work in mammals indicate that CF-derived signals can influence cerebellar output neurons either indirectly via PCs or, in some cases, through direct projections. In mice, CF activity has been shown to encode the salience and behavioral relevance of sensory cues and to instruct PC simple-spike output during GO/NOGO learning ( 65 ). These findings raise the possibility that CF-derived signals—whether transmitted through PCs or via direct CF projections—may also contribute to shaping valence representations in ECs. Electrophysiological studies support this, showing that EC responses are shaped by both inhibitory and excitatory inputs ( 66 ). Although both ECs and PCs show a greater number of danger-encoding neurons, the bias toward danger representation is more pronounced in PCs ( Figs. 2 and 4 ). This difference suggests that ECs integrate additional inputs beyond those from PCs, contributing to a relatively more balanced representation. These findings indicate that information integration occurs at multiple levels within both PCs and ECs, thereby contributing to the formation of internal models that guide behavior. Although ECs provide information that guides subsequent behavioral decisions, their activity does not consistently align with motor output. As shown in Fig. 2D , danger and safety ECs remain active even when locomotion is absent or continues unchanged, supporting the view that their activity reflects contextual valence rather than movement commands. Interaction between cerebellar and telencephalic valence systems A region within the dorsal telencephalon, considered homologous to the mammalian isocortex, is involved in active avoidance learning ( 11 ). In the VR active avoidance task, neurons responding to danger or safety color cues appeared in the dorsal telencephalon after training, similar to ECs and PCs. However, in reversal sessions, their responses disappeared, and new neuronal populations emerged that responded to the reversed cue-valence associations in fish that successfully adapted to the new rule. These findings suggest that the encoding of danger and safety valence differs between the cerebellum and the dorsal telencephalon. It is therefore intriguing to investigate how cerebellar information contributes to the generation of valence-coding neurons in the dorsal telencephalon. A previous report indicates that the dorsal telencephalon forms circuits with the ventral telencephalon (analogous to the mammalian basal ganglia) and the thalamus and that these telencephalic circuits also receive input from the cerebellum via the thalamus ( 67 ). In mammals, reciprocal networks between the isocortex and cerebellum have been implicated in motor preparation dependent on motor learning ( 68 , 69 ). Furthermore, the mPFC, a part of the dorsal telencephalon, forms distinct neural circuits with the amygdala and basal ganglia (e.g., NAc) that are differentially engaged in the representation of fear, fear extinction, and safety during fear conditioning ( 1 – 8 , 48 , 49 ). These findings raise the possibility that the internal model established in cerebellar circuits provides signals that inform dorsal telencephalic networks involved in fear-related decision-making. Our results highlight a previously unrecognized and essential role for cerebellar circuits in representing mental states and forming internal models that guide adaptive behavior, revealing a key contribution of the cerebellum to associative learning. These findings advance our understanding of how internal models are constructed and flexibly updated across interconnected brain regions during complex learning processes. MATERIALS AND METHODS Animal care and transgenic lines The animal work in this study was approved by the Nagoya University Animal Experiment Committee (approval number: S250015-003) and was conducted following the Regulations on Animal Experiments at Nagoya University. All transgenic lines were generated from zebrafish with nacre ( mitfa w2/w2 ) ( 70 ) or casper ( mitfa w2/w2 ; roy a9/a9 ) ( 71 ) background. Tg(aldoca:NTR-TagRFPT)nub125Tg strain, which expresses zebrafish codon-optimized NTR in PCs was used to induce cell death in cerebellar PCs specifically ( 13 , 72 ). hspzGFFgDMC156A ( 43 ) strain, which expresses a modified Gal4-VP16 (GAL4FF, also referred to as GFF) in ECs within the cerebellum, and Tg(5xUAS-hsp70l:GCaMP6s)nkUAShspzGCMP6s13aTg ( 44 ) strain are crossed to generate hspzGFFgDMC156A;Tg(5xUAS-hsp70l:GCaMP6s) strain, which expresses the calcium indicator GCaMP6s in ECs. To generate the Tg(aldoca:GAL4FF, myl7:mCherry)nub127Tg strain via Tol2 transgenesis, which expresses GAL4FF specifically in PCs, the entry vectors pENTR-L1-aldoca-R5 ( 43 ) and pENTR-L5-GAL4FF-SV40pAS-L2 ( 13 ) were used for Gateway-mediated subcloning into the Tol2 destination vector pBH-R1-R2 ( 73 ). Tg(aldoca:GAL4FF, myl7:mCherry)nub127Tg strain was crossed with the Tg(5xUAS-hsp70l:GCaMP6s)nkUAShspzGCMP6s13aTg strain ( 44 ) to obtain Tg(aldoca:GAL4FF, myl7:mCherry);Tg(5xUAS-hsp70l:GCaMP6s) strain, which expresses GCaMP6s in PCs. To generate the Tg(aldoca:GtACR1-EYFP)nub86Tg strain, the entry vectors pENTR-L1-aldoca-R5 ( 43 ) and pENTR-L5-GtACR1-EYFP-SV40pAS-L2 ( 74 ) were used for Gateway-mediated subcloning into the Tol2 vector pT2ALR-Dest ( 73 ). A total of 25 pg of the Tol2 plasmid and 25 pg of transposase-capped RNA were injected into one-cell stage embryos. Transgene expression was confirmed by the presence of tagged fluorescent proteins. Fish were maintained in a 14/10-hour light/dark cycle 28.5°C. Three- to 12-month-old adult fish were used in this study. Head fixation surgery for adult zebrafish The basic idea and equipment for head fixation surgery were used on the basis of the reported method ( 11 ). Adult zebrafish were quickly anesthetized with an anesthetic mixture [0.04% tricaine methanesulfonate (Nacalai Tesque Inc., catalog no. 01916-32)] and Ringer’s solution (100 mM NaCl, 2.55 mM KCl, 1.53 mM CaCl 2 , 5.0 mM NaHCO 3 , 1.0 mM MgCl 2 ·6H 2 O, and 0.5 mM NaH 2 PO 4 ·2H 2 O) and secured in a custom-made surgical dish. To maintain under anesthesia and keep alive during the surgery, perfusing anesthetizing mixture was performed using a peristaltic pump (ATTO, SJ-1211II-H). The skin tissue above the skull was removed with a microknife. The exposed skull was air dried and treated with a dental surface conditioner (Sun Medical, catalog no. 204610567). The dental adhesive compound mixed with catalyst V (Sun Medical, catalog no. 204610580), quick monomer (Sun Medical, catalog no. 204610399), and polymer powder (Sun Medical, catalog no. 204610563) was applied to the skull above the optic tectum and telencephalon. A steel bar was quickly pressed onto the bonding compound on the skull to adhere the bar to the skull. The fish was subsequently secured to the joint device that connects the attached steel bar to the base of the experimental tank ( 11 ). The tank was filled with Ringer’s solution. The Ringer’s solution in the tank was continuously delivered into the fish’s mouth with a peristaltic pump to recover from anesthesia and assist breathing during the experiment. VR setup for adult zebrafish The VR system for the adult zebrafish was constructed using reported hardware and software ( 11 ). Three LCD displays (Good display, GD70MLXD) were placed at the front, left, and right of the tank where head-fixed fish was placed ( Fig. 1A ). The view from the head-fixed fish in the VR environment was shown in these LCD displays. The VR environment consisted of straight corridor with black stripes at equal interval and perpendicular to the direction of the corridor. The walls of the corridor were displayed in white, red, or blue color. The VR space was constructed using Unity (Unity Technologies). The tail part of the head-fixed fish was monitored by the universal serial bus (USB) camera (BSW20KM11, BUFFALO) placed on top of the tank at 10 Hz. To avoid interference from the LCD display light, an infrared filter (IR 76 FUJI FILTER, FUJIFILM) was attached to the USB camera, and infrared light was placed behind the fish. The propulsion in the VR was calculated using custom software developed with LabVIEW (National Instruments, USA) and MATLAB (MathWorks, USA). The calculation method was previously reported ( 11 ). Briefly, the curvature of the tail was estimated on the basis of the difference in angle between the base and the tip of the tail, and the forward movement was derived from both the frequency and magnitude of the tail bending. At the same time, we also computed the tail movement intensity, which was used as tail kinematics for further analysis. Specifically, the intensity value was calculated as the horizontal displacement in pixels between the upper and lower portions of the extracted tail centerline within a predefined region of interest (ROI). GO/NOGO task The fish were maintained in individual tanks filled with 100 to 200 ml of Ringer’s solution overnight before the experiment. The head-fixed fish underwent GO and NOGO tasks in the virtual corridor ( Fig. 1B ). In the GO task, the wall color of the area where the fish was initially located was set to blue, and the front wall was red. If the fish moved forward through the corridor and entered the red area within 10 s of task initiation, then the trial was considered successful. In the NOGO task, the wall color of the area where the fish was initially located also turned blue, and the front wall turned red. If the fish remained in the red area for the full 10 s without entering the blue area, the trial was considered successful. For both the GO and NOGO tasks, if the fish succeeded, then the entire wall of the corridor turned white for 15-s intertrial interval (ITI). If the fish failed, then a 1-s electric shock (10 V), generated by the NI-6001 (National Instruments, USA), was delivered through electrodes placed on both sides of the fish before the ITI began. By repeating these GO and NOGO tasks, the fish learned the rule that “the blue area signals danger, while the red area signals safety.” After the head-fixed fish fully recovered from the anesthesia, the fish underwent an adaptation session. During the adaptation session, 10 GO tasks and 10 NOGO tasks were performed in a random order, and no electric shock was delivered in the failure trials. After the adaptation session, training sessions were conducted where an electric shock was delivered upon failure. Each training session consisted of 30 GO tasks and 30 NOGO tasks presented in a random order. Three training sessions were performed, with 5- to 10-min intervals between sessions. Fish that successfully completed at least 8 of 10 consecutive trials in both GO and NOGO tasks during the training session were defined as learner fish. After the third training session, two reversal sessions were conducted. In the reversal session, the corridor color associated with the electric shock was switched from blue to red. In the reversal GO task, the fish was initially located in the red area and required to enter the blue area in front of the fish within 10 s to avoid the shock. In the reversal NOGO task, the fish was initially located in the blue area and required to stay in the blue area without entering the red area in front of the fish to avoid the shock. Therefore, the fish learned the opposite rule from the training session, that “the red area signals danger, while the blue area signals safety.” During the tasks, data including “tail movement intensity” and fish position in the VR space were recorded for further analysis. To visualize the progression of success rates across trials in Figs. 1E and 6C , trials were first grouped by task type (GO or NOGO) while retaining their original trial indices. The success rate for each task was calculated on the basis of a sliding window of seven trials, including the target trial and the three preceding and three succeeding trials of the same task type. The resulting success rates were then averaged across all fish to generate the overall time course of task performance. Trials positioned at the beginning or end of each session, which lacked enough neighboring trials to complete the seven-trial window (i.e., the first and last three trials), were excluded from the analysis. Ablation of PCs Adult Tg(aldoca:NTR-TagRFPT) fish and their control WT fish (3 to 12 months old) were treated with 10 mM MTZ (M3761, Sigma-Aldrich) solution for 18 hours and returned to their tank. Ten to 12 days after treatment, the fish were subjected to behavior tests ( 13 ). Ca 2+ imaging for ECs and PCs Ca 2+ imaging was performed using a spinning disk confocal scanner unit (CSU-W1, YOKOGAWA) and an electron multiplying charge-coupled device (EMCCD) camera (Andor iXon Ultra, OXFORD INSTRUMENTS) mounted on an upright microscope (BX61WI, Evident Scientific Inc., Japan). The imaging software Metamorph ver. 7.7 (Molecular Devices) was used to control the number of acquired images, exposure time, gain, and binning. The exposure time and binning were configured to acquire images at 4.5 frames per second. The EMCCD camera generated a pulse signal each time an image was acquired, and this pulse signal was recorded by the LabVIEW software used for the GO/NOGO experiments, allowing synchronization of the Ca 2+ imaging frame numbers with the behavioral data from the GO/NOGO trials. Imaging targeted the soma of ECs located at a depth of ~100 to 150 μm from the surface of the cerebellum, and single-plane images were acquired from a 670 μm–by–670 μm area. The imaging area was centered on the basis of the cranial junction along the midline. Ca 2+ imaging for dendrites of PCs was performed using Tg(aldoca:GAL4FF);Tg(5xUAS-hsp70l:GCaMP6s) fish. Imaging targeted the dendrites of PCs located at a depth of ~50 to 100 μm from the surface of the cerebellum, and single-plane images were acquired from a 340 μm–by–340 μm area. The images were taken from either the left or right hemisphere of the cerebellum based on the cranial junction along the midline. Extraction of neural activity data from Ca 2+ imaging Neural activity data from Ca 2+ imaging were extracted using suite2p v0.14.0 ( https://github.com/MouseLand/suite2p ) ( 75 ), an open-source Python-based software that mainly performs removing motion artifacts (registration), ROI detection, and the calculation of neural activity data (Δ F / F , z -score, and spike estimates). To accurately detect individual PC dendritic segments or EC somata and to perform motion correction with suite2p, the following preprocessing settings and analysis conditions were applied. A reference image used for registration was generated by averaging 2000 to 4000 randomly selected frames. Each image was divided into 64 × 64 pixel blocks, and the displacement within each block was corrected for each frame. For ROI detection, the minimum radius of each ROI was set to 1.5 to 16 μm (2 to 12 pixels), and the signal-to-noise threshold was set to 1.5. When detected ROIs were in close proximity, a merged threshold for brightness fluctuation similarity was set between 1.0 and 1.4, depending on the fish. Raw fluorescence signals extracted from identified ROIs (F_raw) were corrected by subtracting neuropil contamination. Specifically, neuropil signals (F_neu) measured from the immediate surroundings of each ROI were scaled by a factor of 0.8 and subtracted from F_raw to obtain neuropil-corrected fluorescence (Δ F ). Subsequently, the Δ F / F was calculated by normalizing the corrected Δ F relative to this baseline with “preprocess” function in suite2p. In the preprocess function, baseline fluorescence was determined using the “maximin” baseline estimation method with parameters: baseline window = 60.0 s, sigma baseline = 10.0, and percentile baseline = 8.0. The z -score for each ROI’s neural activity was computed by first calculating the mean and SD of the Δ F / F values over the entire recording period for each ROI. Each data point was standardized by subtracting the mean and dividing by the SD. Estimated spike rates were inferred from changes in fluorescence intensity using the OASIS (Online Active Set method to Infer Spikes) algorithm implemented within suite2p. The fluorescence decay constant τ was set to 1.25, considering the properties of the Ca 2+ indicator GCaMP6s. The obtained neural activity data were visualized by suite2p, and the accuracy of motion correction and ROI detection was verified by the “View registered binary” function. When the detected ROIs failed to track the specific area through the experiments due to insufficient registration or significant z -drift, the data were excluded from further analysis. Generalized linear model To analyze the relationship between neural activity and environmental factors or behavioral indicators during GO/NOGO tasks, a Poisson GLM was applied, with estimated spike rates as the response variable and environmental factors and behavioral indicators as explanatory variables (regressors) (fig. S3A). Note that spike rate inference was used exclusively for the GLM analysis; Δ F / F signals in all other analyses were computed directly from fluorescence traces without spike inference. The GLM analysis was performed using the Python-based open-source software GLM_Tensorflow_2 ( https://github.com/sytseng/GLM_Tensorflow_2 ) ( 76 ). Eight environmental factors and behavioral indicators obtained during the GO/NOGO tasks were used as explanatory variables: tail movement intensity over time, the timing of red background, the timing of blue background, the timing of ITI after a successful trial, timing of ITI after a failed trial, the timing of electric shock, the 1 s immediately after a successful trial, and position of the fish within the VR space. The explanatory variables related to tail movement intensity and the position of the fish in the VR space were normalized to a range of 0 to 1. For the timing of the red background, the timing of blue background, the timing of ITIs, the timing of electric shock, and the 1 s after a successful trial, binary functions were used, assigning 1 during the corresponding period and 0 otherwise. Modeling was performed separately for each session, and the coefficients (weights) of each explanatory variable for each session were calculated. The weights for the timing of red background were labeled as red weight, and the weights for the timing of blue background were labeled as blue weight. To achieve sufficient modeling accuracy while avoiding overfitting, L1 regularization (Lasso regression) and L2 regularization (Ridge regression) were applied. The regularization parameters (λ) for both L1 and L2 were selected through cross-validation, following a logarithmic search across a wide dynamic range (from 10 3 to 10 −6 ). To determine the optimal λ, we used the select_model function with se_fraction = 0, which selects the model that minimizes the cross-validated deviance. The Adaptive Moment Estimation (Adam) optimizer was used for model training, with a learning rate of 0.01. For training, 90% of the trials were used as training data, while the remaining 10% were used as test data to evaluate the model’s validity. To evaluate model performance on the test data, we calculated the fraction of deviance explained (FDE), defined as FDE = 1 − (Deviance_model/Deviance_null), which quantifies how much better the fitted model performs compared to a null model that only includes a constant mean term. Using this metric, we confirmed that the average FDE across all ECs and sessions from learner fish ( n = 10) was greater than 0.05 under the above model settings. To ensure that GLM weights reflected cue-selective responses rather than prolonged calcium transients, we fitted the model to OASIS-derived inferred spike rates rather than raw fluorescence. This approach reduces cross-contamination between consecutive cue periods and enables reliable estimation of red and blue weights even when calcium signals appear to extend across color boundaries in individual trials. Optogenetic manipulation Optogenetic manipulation experiments during the GO/NOGO tasks were performed using Tg(aldoca:GtACR1-EYFP) strain, which specifically express Gt ACR1-EYFP in PCs, or control nacre fish. Before the experiment, the experimental tank was positioned under the objective lens so that the cerebellar region was within the field of view (290 μm by 420 μm) monitored by CCD camera (ORCA-R2, Hamamatsu Photonics, Japan) mounted on an upright microscope (BX51WI; Evident Scientific Inc., Japan) and the imaging software StreamPix7 (NorPix Inc., Canada). The GO/NOGO tasks began with one adaptation session without electric shock, followed by three training sessions where a 1-s electric shock was delivered upon task failure. Subsequently, two reversal sessions were conducted, in which the wall color associated with the electric shock was reversed from training session ( Fig. 6B ). In the reversal sessions, during failed trials, 1-s illumination of the cerebellar region with 520-nm light (9 mW/mm 2 ) was delivered simultaneously with a 1-s electric shock. The timing of the illumination was controlled by generating synchronized pulse signals via LabVIEW software. These signals were transmitted to a light-emitting diode (LED) light source (Prizmatix Ltd., Israel) through a pulse generator (Master-9, AMPI, Israel). The stimulus light from the LED was projected onto the cerebellum using a patterned illumination system (LEOPARD, OPTO-LINE Inc., Japan) and controlled by software polyscan 2 (Mightex, Canada) ( 77 ). Immunostaining Cryosections (14-μm thick) were prepared according to the previous publication ( 41 ). The sections were immunostained as previously described ( 41 , 78 ). Anti–green fluorescent protein antibody (1:1000, rat, Nacalai, Japan, catalog no. 04404-84, RRID: AB_10013361) and anti-parvalbumin 7 antibody (1:1000, mouse monoclonal, ascites) ( 41 ) were used as primary antibodies. As secondary antibodies, Alexa Fluor 488 goat anti-rat (1:500, H + L, Invitrogen, Thermo Fisher Scientific, catalog no. A11006, RRID: AB_2534074) and Alexa Fluor 568 goat anti-mouse immunoglobulin G (1:500, Invitrogen, catalog no. A11031, RRID: AB_144696) were used. Imaging was performed using a confocal laser scanning inverted microscope (LSM700, Zeiss, Germany). Neural and behavioral data analysis and statistics Neural activity and behavioral data analyses, as well as statistical tests, were performed using Python in a Jupyter Notebook environment. Data processing and analysis were primarily conducted using the Python libraries NumPy and pandas. Figures were generated using the Python-based plotting library Matplotlib. Statistical significance tests, including Welch’s t tests, Fisher’s exact test, and the Mann-Whitney U test, were performed using the SciPy statistics module. Permutation analysis was implemented in Python to assess whether a statistically significant group difference in GO task learning could arise by chance. Control WT fish ( n = 17) and aldoca:NTR fish ( n = 12) were randomly reassigned into two surrogate groups (group A, n = 17; group B, n = 12), matching the original sample sizes. For each permutation, the mean GO task success rate in training session 3 was calculated for each group. When the mean success rate was higher in group A than in group B, we tested whether this group difference reached statistical significance using the Mann-Whitney U test ( P < 0.05). This shuffling procedure was repeated 1000 times, and the frequency with which a statistically significant WT > aldoca:NTR–like difference was observed was quantified. The entire permutation analysis was repeated three times independently, and the resulting range of frequencies was reported. Artificial intelligence–assisted technologies ChatGPT (GPT-4 and GPT-5, OpenAI) was used solely to assist with language refinement of the manuscript and limited editing of Python code. It was not used for data generation, analysis, interpretation, or figure preparation. All artificial intelligence–assisted content was reviewed and validated by the authors, who take full responsibility for the integrity and accuracy of the work. Acknowledgments We thank K. Kawakami (National Institute of Genetics, Japan) and the National BioResource Project for providing the transgenic zebrafish lines hspzGFFgDMC156A and Tg(UAS:GCaMP6s) , as well as Tol2-related plasmids. We are also grateful to J. Spudich (The University of Texas Health Science Center at Houston) for providing the Gt ACR1-EYFP plasmid. We thank Y. Takayanagi and M. Satoh for support in fish breeding and care and the members of the Hibi Laboratory for helpful discussions. We also thank M. Lovett-Barron, K. A. Martin, and G. Meyerhof for valuable comments on the manuscript. We further acknowledge the Research Equipment Development Group, Technical Center of Nagoya University, for support in manufacturing research instruments. Funding: This work was supported by the Japan Society for the Promotion of Science KAKENHI (JP23K23894 and JP25K02283 to M.H., JP23K05845 to T.S., JP22KJ1601 to W.K., and JP23H04976 to H.O.) and Japan Science and Technology Agency (JPMJCR1753 to M.H.). Author contributions: Conceptualization: W.K., H.O., and M.H. Methodology: W.K., Y.T., M.T., T.K., and H.O. Software: W.K., T.I., M.T., T.K., and H.O. Validation: W.K., M.T., and H.O. Formal analysis: W.K. and T.K. Investigation: W.K. Resources: M.T., H.O., and M.H. Data curation: W.K. Writing—original draft: W.K., H.O., and M.H. Writing—review and editing: M.T., T.S., T.K., H.O., and M.H. Visualization: W.K. and T.K. Supervision: T.K., H.O., and M.H. Project administration: M.H. Funding acquisition: W.K., T.S., H.O., and M.H. Competing interests: H.O. has a research advisory contract with Kao Corporation. All other authors declare that they have no competing interests. Data, code, and materials availability: All data and code needed to evaluate and reproduce the results in the paper are present in the paper, the Supplementary Materials, or in the following Zenodo repository ( https://doi.org/10.5281/zenodo.18317482 ), including the complete datasets underlying the analyses and results. The source data underlying all figure panels are provided as data S1. All custom analysis code used in this study is also available in the same Zenodo repository. Zebrafish lines used in this study are being deposited in the National BioResource Project Zebrafish (NBRP Zebrafish, Japan; https://shigen.nig.ac.jp/zebra/ ) and can be searched by strain name. The following strains are available or will be available: Tg(aldoca:NTR-TagRFPT)nub125Tg [registered as Tg(aldoca:NTR-TagRFPT) ], hspzGFFgDMC156A , Tg(5xUAS-hsp70l:GCaMP6s)nkUAShspzGCMP6s13aTg [registered as UAShspGCaMP6s13A ], and Tg(aldoca:GtACR1-EYFP)nub86Tg [registered as Tg(aldoca:GtACR1-EYFP) ]. Tg(aldoca:GAL4FF, myl7:mCherry)nub127Tg is in the process of deposition and will be available as Tg(aldoca:GAL4FF, myl7:mCherry) . All lines can be requested through NBRP Zebrafish. Supplementary Materials The PDF file includes: Figs. S1 to S16 Legends for movies S1 and S2 Legend for data S1 sciadv.aeb5860_sm.pdf (12.6MB, pdf) Other Supplementary Material for this manuscript includes the following: Movies S1 and S2 sciadv.aeb5860_movies_s1_and_s2.zip (39.1MB, zip) Data S1 sciadv.aeb5860_data_s1.zip (10.4MB, zip) REFERENCES 1. Bravo-Rivera C., Roman-Ortiz C., Brignoni-Perez E., Sotres-Bayon F., Quirk G. J., Neural structures mediating expression and extinction of platform-mediated avoidance. J. Neurosci. 34, 9736–9742 (2014). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Ramirez F., Moscarello J. M., LeDoux J. E., Sears R. M., Active avoidance requires a serial basal amygdala to nucleus accumbens shell circuit. J. Neurosci. 35, 3470–3477 (2015). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Moscarello J. M., LeDoux J. E., Active avoidance learning requires prefrontal suppression of amygdala-mediated defensive reactions. J. Neurosci. 33, 3815–3823 (2013). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Diehl M. M., Bravo-Rivera C., Rodriguez-Romaguera J., Pagan-Rivera P. A., Burgos-Robles A., Roman-Ortiz C., Quirk G. J., Active avoidance requires inhibitory signaling in the rodent prelimbic prefrontal cortex. eLife 7, e34657 (2018). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Gongwer M. W., Klune C. B., Couto J., Jin B., Enos A. S., Chen R., Friedmann D., DeNardo L. A., Brain-wide projections and differential encoding of prefrontal neuronal classes underlying learned and innate threat avoidance. J. Neurosci. 43, 5810–5830 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Jercog D., Winke N., Sung K., Fernandez M. M., Francioni C., Rajot D., Courtin J., Chaudun F., Jercog P. E., Valerio S., Herry C., Dynamical prefrontal population coding during defensive behaviours. Nature 595, 690–694 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 7. Capuzzo G., Floresco S. B., Prelimbic and infralimbic prefrontal regulation of active and inhibitory avoidance and reward-seeking. J. Neurosci. 40, 4773–4787 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Klune C. B., Goodpaster C. M., Gongwer M. W., Gabriel C. J., An J., Chen R., Jones N. S., Williams O. H., Shari M., Ramirez M., Hacking A., Andrade T., Schwarz L. A., DeNardo L. A., Developmentally distinct architectures in top-down pathways controlling threat avoidance. Nat. Neurosci. 28, 823–835 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Amo R., Fredes F., Kinoshita M., Aoki R., Aizawa H., Agetsuma M., Aoki T., Shiraki T., Kakinuma H., Matsuda M., Yamazaki M., Takahoko M., Tsuboi T., Higashijima S., Miyasaka N., Koide T., Yabuki Y., Yoshihara Y., Fukai T., Okamoto H., The habenulo-raphe serotonergic circuit encodes an aversive expectation value essential for adaptive active avoidance of danger. Neuron 84, 1034–1048 (2014). [ DOI ] [ PubMed ] [ Google Scholar ] 10. Aoki T., Kinoshita M., Aoki R., Agetsuma M., Aizawa H., Yamazaki M., Takahoko M., Amo R., Arata A., Higashijima S., Tsuboi T., Okamoto H., Imaging of neural ensemble for the retrieval of a learned behavioral program. Neuron 78, 881–894 (2013). [ DOI ] [ PubMed ] [ Google Scholar ] 11. Torigoe M., Islam T., Kakinuma H., Fung C. C. A., Isomura T., Shimazaki H., Aoki T., Fukai T., Okamoto H., Zebrafish capable of generating future state prediction error show improved active avoidance behavior in virtual reality. Nat. Commun. 12, 5712 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Lal P., Tanabe H., Suster M. L., Ailani D., Kotani Y., Muto A., Itoh M., Iwasaki M., Wada H., Yaksi E., Kawakami K., Identification of a neuronal population in the telencephalon essential for fear conditioning in zebrafish. BMC Biol. 16, 45 (2018). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Koyama W., Hosomi R., Matsuda K., Kawakami K., Hibi M., Shimizu T., Involvement of cerebellar neural circuits in active avoidance conditioning in zebrafish. eNeuro 8, ENEURO.0507-20.2021 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Lin Q., Manley J., Helmreich M., Schlumm F., Li J. M., Robson D. N., Engert F., Schier A., Nöbauer T., Vaziri A., Cerebellar neurodynamics predict decision timing and outcome on the single-trial level. Cell 180, 536–551.e17 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Aizenberg M., Schuman E. M., Cerebellar-dependent learning in larval zebrafish. J. Neurosci. 31, 8708–8712 (2011). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Hibi M., Matsuda K., Takeuchi M., Shimizu T., Murakami Y., Evolutionary mechanisms that generate morphology and neural-circuit diversity of the cerebellum. Dev. Growth Differ. 59, 228–243 (2017). [ DOI ] [ PubMed ] [ Google Scholar ] 17. Ohmae S., Medina J. F., Climbing fibers encode a temporal-difference prediction error during cerebellar learning in mice. Nat. Neurosci. 18, 1798–1803 (2015). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. ten Brinke M. M., Boele H. J., Spanke J. K., Potters J. W., Kornysheva K., Wulff P., AC I. J., Koekkoek S. K., De Zeeuw C. I., Evolving models of Pavlovian conditioning: Cerebellar cortical dynamics in awake behaving mice. Cell Rep. 13, 1977–1988 (2015). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Ito M., Mechanisms of motor learning in the cerebellum. Brain Res. 886, 237–245 (2000). [ DOI ] [ PubMed ] [ Google Scholar ] 20. Ito M., Bases and implications of learning in the cerebellum—Adaptive control and internal model mechanism. Prog. Brain Res. 148, 95–109 (2005). [ DOI ] [ PubMed ] [ Google Scholar ] 21. Ito M., Error detection and representation in the olivo-cerebellar system. Front Neural Circuits 7, 1 (2013). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Wolpert D. M., Kawato M., Multiple paired forward and inverse models for motor control. Neural Netw. 11, 1317–1329 (1998). [ DOI ] [ PubMed ] [ Google Scholar ] 23. Ito M., Control of mental activities by internal models in the cerebellum. Nat. Rev. Neurosci. 9, 304–313 (2008). [ DOI ] [ PubMed ] [ Google Scholar ] 24. McCormick D. A., Thompson R. F., Cerebellum: Essential involvement in the classically conditioned eyelid response. Science 223, 296–299 (1984). [ DOI ] [ PubMed ] [ Google Scholar ] 25. Ten Brinke M. M., Heiney S. A., Wang X., Proietti-Onori M., Boele H. J., Bakermans J., Medina J. F., Gao Z., De Zeeuw C. I., Dynamic modulation of activity in cerebellar nuclei neurons during pavlovian eyeblink conditioning in mice. eLife 6, e28132 (2017). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Kim J. J., Krupa D. J., Thompson R. F., Inhibitory cerebello-olivary projections and blocking effect in classical conditioning. Science 279, 570–573 (1998). [ DOI ] [ PubMed ] [ Google Scholar ] 27. Silva N. T., Ramírez-Buriticá J., Pritchett D. L., Carey M. R., Climbing fibers provide essential instructive signals for associative learning. Nat. Neurosci. 27, 940–951 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Kostadinov D., Beau M., Blanco-Pozo M., Hausser M., Predictive and reactive reward signals conveyed by climbing fiber inputs to cerebellar Purkinje cells. Nat. Neurosci. 22, 950–962 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Tsutsumi S., Hidaka N., Isomura Y., Matsuzaki M., Sakimura K., Kano M., Kitamura K., Modular organization of cerebellar climbing fiber inputs during goal-directed behavior. eLife 8, e47021 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Heffley W., Song E. Y., Xu Z., Taylor B. N., Hughes M. A., McKinney A., Joshua M., Hull C., Coordinated cerebellar climbing fiber activity signals learned sensorimotor predictions. Nat. Neurosci. 21, 1431–1441 (2018). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Ikezoe K., Hidaka N., Manita S., Murakami M., Tsutsumi S., Isomura Y., Kano M., Kitamura K., Cerebellar climbing fibers multiplex movement and reward signals during a voluntary movement task in mice. Commun. Biol. 6, 924 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Wagner M. J., Kim T. H., Savall J., Schnitzer M. J., Luo L., Cerebellar granule cells encode the expectation of reward. Nature 544, 96–100 (2017). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Sengupta M., Thirumalai V., AMPA receptor mediated synaptic excitation drives state-dependent bursting in Purkinje neurons of zebrafish larvae. eLife 4, e09158 (2015). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Chang W., Pedroni A., Hohendorf V., Giacomello S., Hibi M., Köster R. W., Ampatzis K., Functionally distinct Purkinje cell types show temporal precision in encoding locomotion. Proc. Natl. Acad. Sci. U.S.A. 117, 17330–17337 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Matsui H., Namikawa K., Babaryka A., Köster R. W., Functional regionalization of the teleost cerebellum analyzed in vivo. Proc. Natl. Acad. Sci. U.S.A. 111, 11846–11851 (2014). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Narayanan S., Varma A., Thirumalai V., Predictive neural computations in the cerebellum contribute to motor planning and faster behavioral responses in larval zebrafish. Sci. Adv. 10, eadi6470 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Markov D. A., Petrucco L., Kist A. M., Portugues R., A cerebellar internal model calibrates a feedback controller involved in sensorimotor control. Nat. Commun. 12, 6694 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Valente A., Huang K. H., Portugues R., Engert F., Ontogeny of classical and operant learning behaviors in zebrafish. Learn. Mem. 19, 170–177 (2012). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Tanabe K., Kani S., Shimizu T., Bae Y. K., Abe T., Hibi M., Atypical protein kinase C regulates primary dendrite specification of cerebellar Purkinje cells by localizing Golgi apparatus. J. Neurosci. 30, 16983–16992 (2010). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Ahn A. H., Dziennis S., Hawkes R., Herrup K., The cloning of zebrin II reveals its identity with aldolase C. Development 120, 2081–2090 (1994). [ DOI ] [ PubMed ] [ Google Scholar ] 41. Bae Y. K., Kani S., Shimizu T., Tanabe K., Nojima H., Kimura Y., Higashijima S., Hibi M., Anatomy of zebrafish cerebellum and screen for mutations affecting its development. Dev. Biol. 330, 406–426 (2009). [ DOI ] [ PubMed ] [ Google Scholar ] 42. Brochu G., Maler L., Hawkes R., Zebrin II: A polypeptide antigen expressed selectively by Purkinje cells reveals compartments in rat and fish cerebellum. J. Comp. Neurol. 291, 538–552 (1990). [ DOI ] [ PubMed ] [ Google Scholar ] 43. Takeuchi M., Matsuda K., Yamaguchi S., Asakawa K., Miyasaka N., Lal P., Yoshihara Y., Koga A., Kawakami K., Shimizu T., Hibi M., Establishment of Gal4 transgenic zebrafish lines for analysis of development of cerebellar neural circuitry. Dev. Biol. 397, 1–17 (2015). [ DOI ] [ PubMed ] [ Google Scholar ] 44. Muto A., Lal P., Ailani D., Abe G., Itoh M., Kawakami K., Activation of the hypothalamic feeding centre upon visual prey detection. Nat. Commun. 8, 15029 (2017). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. Wagner M. J., Savall J., Hernandez O., Mel G., Inan H., Rumyantsev O., Lecoq J., Kim T. H., Li J. Z., Ramakrishnan C., Deisseroth K., Luo L., Ganguli S., Schnitzer M. J., A neural circuit state change underlying skilled movements. Cell 184, 3731–3747.e21 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Govorunova E. G., Sineshchekov O. A., Janz R., Liu X., Spudich J. L., Natural light-gated anion channels: A family of microbial rhodopsins for advanced optogenetics. Science 349, 647–650 (2015). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 47. Kong E., Monje F. J., Hirsch J., Pollak D. D., Learning not to fear: Neural correlates of learned safety. Neuropsychopharmacology 39, 515–527 (2014). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 48. Duvarci S., Pare D., Amygdala microcircuits controlling learned fear. Neuron 82, 966–980 (2014). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Herry C., Johansen J. P., Encoding of fear learning and memory in distributed neuronal circuits. Nat. Neurosci. 17, 1644–1654 (2014). [ DOI ] [ PubMed ] [ Google Scholar ] 50. Ray M. H., Moaddab M., McDannald M. A., Threat and bidirectional valence signaling in the nucleus accumbens core. J. Neurosci. 42, 817–833 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 51. Sacchetti B., Baldi E., Lorenzini C. A., Bucherelli C., Cerebellar role in fear-conditioning consolidation. Proc. Natl. Acad. Sci. U.S.A. 99, 8406–8411 (2002). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 52. Supple W. F. Jr., Sebastiani L., Kapp B. S., Purkinje cell responses in the anterior cerebellar vermis during Pavlovian fear conditioning in the rabbit. Neuroreport 4, 975–978 (1993). [ DOI ] [ PubMed ] [ Google Scholar ] 53. Apps R., Garwicz M., Anatomical and physiological foundations of cerebellar information processing. Nat. Rev. Neurosci. 6, 297–311 (2005). [ DOI ] [ PubMed ] [ Google Scholar ] 54. Sillitoe R. V., Joyner A. L., Morphology, molecular codes, and circuitry produce the three-dimensional complexity of the cerebellum. Annu. Rev. Cell Dev. Biol. 23, 549–577 (2007). [ DOI ] [ PubMed ] [ Google Scholar ] 55. Kebschull J. M., Richman E. B., Ringach N., Friedmann D., Albarran E., Kolluru S. S., Jones R. C., Allen W. E., Wang Y., Cho S. W., Zhou H., Ding J. B., Chang H. Y., Deisseroth K., Quake S. R., Luo L., Cerebellar nuclei evolved by repeatedly duplicating a conserved cell-type set. Science 370, eabd5059 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 56. Apps R., Hawkes R., Cerebellar cortical organization: A one-map hypothesis. Nat. Rev. Neurosci. 10, 670–681 (2009). [ DOI ] [ PubMed ] [ Google Scholar ] 57. Larouche M., Hawkes R., From clusters to stripes: The developmental origins of adult cerebellar compartmentation. Cerebellum 5, 77–88 (2006). [ DOI ] [ PubMed ] [ Google Scholar ] 58. Sugihara I., Fujita H., Peri- and postnatal development of cerebellar compartments in the mouse. Cerebellum 12, 325–327 (2013). [ DOI ] [ PubMed ] [ Google Scholar ] 59. Karam S. D., Burrows R. C., Logan C., Koblar S., Pasquale E. B., Bothwell M., Eph receptors and ephrins in the developing chick cerebellum: Relationship to sagittal patterning and granule cell migration. J. Neurosci. 20, 6488–6500 (2000). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 60. Sugihara I., Shinoda Y., Molecular, topographic, and functional organization of the cerebellar cortex: A study with combined aldolase C and olivocerebellar labeling. J. Neurosci. 24, 8771–8785 (2004). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 61. Sugihara I., Shinoda Y., Molecular, topographic, and functional organization of the cerebellar nuclei: Analysis by three-dimensional mapping of the olivonuclear projection and aldolase C labeling. J. Neurosci. 27, 9696–9710 (2007). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 62. Félix R., Markov D. A., Renninger S. L., Tomás A. R., Laborde A., Carey M. R., Orger M. B., Portugues R., Structural and functional organization of visual responses in the inferior olive of larval zebrafish. J. Neurosci. 44, e2352212023 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 63. De Zeeuw C. I., Bidirectional learning in upbound and downbound microzones of the cerebellum. Nat. Rev. Neurosci. 22, 92–110 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 64. Ten Brinke M. M., Boele H. J., De Zeeuw C. I., Conditioned climbing fiber responses in cerebellar cortex and nuclei. Neurosci. Lett. 688, 26–36 (2019). [ DOI ] [ PubMed ] [ Google Scholar ] 65. Bina L., Romano V., Hoogland T. M., Bosman L. W. J., De Zeeuw C. I., Purkinje cells translate subjective salience into readiness to act and choice performance. Cell Rep. 37, 110116 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 66. Najac M., McLean D. L., Raman I. M., Synaptic variance and action potential firing of cerebellar output neurons during motor learning in larval zebrafish. Curr. Biol. 33, 3299–3311.e3 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 67. Tanimoto Y., Kakinuma H., Aoki R., Shiraki T., Higashijima S. I., Okamoto H., Transgenic tools targeting the basal ganglia reveal both evolutionary conservation and specialization of neural circuits in zebrafish. Cell Rep. 43, 113916 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 68. Gao Z., Davis C., Thomas A. M., Economo M. N., Abrego A. M., Svoboda K., De Zeeuw C. I., Li N., A cortico-cerebellar loop for motor planning. Nature 563, 113–116 (2018). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 69. Zhu J., Hasanbegovic H., Liu L. D., Gao Z., Li N., Activity map of a cortico-cerebellar loop underlying motor planning. Nat. Neurosci. 26, 1916–1928 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 70. Lister J. A., Robertson C. P., Lepage T., Johnson S. L., Raible D. W., Nacre encodes a zebrafish microphthalmia-related protein that regulates neural-crest-derived pigment cell fate. Development 126, 3757–3767 (1999). [ DOI ] [ PubMed ] [ Google Scholar ] 71. White R. M., Sessa A., Burke C., Bowman T., LeBlanc J., Ceol C., Bourque C., Dovey M., Goessling W., Burns C. E., Zon L. I., Transparent adult zebrafish as a tool for in vivo transplantation analysis. Cell Stem Cell 2, 183–189 (2008). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 72. Tabor K. M., Bergeron S. A., Horstick E. J., Jordan D. C., Aho V., Porkka-Heiskanen T., Haspel G., Burgess H. A., Direct activation of the Mauthner cell by electric field pulses drives ultrarapid escape responses. J. Neurophysiol. 112, 834–844 (2014). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 73. Dohaku R., Yamaguchi M., Yamamoto N., Shimizu T., Osakada F., Hibi M., Tracing of afferent connections in the zebrafish cerebellum using recombinant rabies virus. Front Neural Circuits 13, 30 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 74. Hagio H., Koyama W., Hosaka S., Song A. D., Narantsatsral J., Matsuda K., Shimizu T., Hososhima S., Tsunoda S. P., Kandori H., Hibi M., Optogenetic manipulation of neuronal and cardiomyocyte functions in zebrafish using microbial rhodopsins and adenylyl cyclases. eLife 12, e83975 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 75. M. Pachitariu, C. Stringer, M. Dipoppa, S. Schröder, L. F. Rossi, H. Dalgleish, M. Carandini, K. D. Harris, Suite2p: Beyond 10,000 neurons with standard two-photon microscopy. bioRxiv 061507 [Preprint] (2017). 10.1101/061507. [ DOI ] 76. Tseng S. Y., Chettih S. N., Arlt C., Barroso-Luque R., Harvey C. D., Shared and specialized coding across posterior cortical areas for dynamic navigation decisions. Neuron 110, 2484–2502.e16 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 77. Hagio H., Koyama W., Hosaka S., Song A. D., Narantsatsral J., Matsuda K., Sugihara T., Shimizu T., Koyanagi M., Terakita A., Hibi M., Optogenetic manipulation of Gq- and Gi/o-coupled receptor signaling in neurons and heart muscle cells. eLife 12, e83974 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 78. Kani S., Bae Y. K., Shimizu T., Tanabe K., Satou C., Parsons M. J., Scott E., Higashijima S., Hibi M., Proneural gene-linked neurogenesis in zebrafish cerebellum. Dev. Biol. 343, 1–17 (2010). [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Figs. S1 to S16 Legends for movies S1 and S2 Legend for data S1 sciadv.aeb5860_sm.pdf (12.6MB, pdf) Movies S1 and S2 sciadv.aeb5860_movies_s1_and_s2.zip (39.1MB, zip) Data S1 sciadv.aeb5860_data_s1.zip (10.4MB, zip) Data Availability Statement All data and code needed to evaluate and reproduce the results in the paper are present in the paper, the Supplementary Materials, or in the following Zenodo repository ( https://doi.org/10.5281/zenodo.18317482 ), including the complete datasets underlying the analyses and results. The source data underlying all figure panels are provided as data S1. All custom analysis code used in this study is also available in the same Zenodo repository. Zebrafish lines used in this study are being deposited in the National BioResource Project Zebrafish (NBRP Zebrafish, Japan; https://shigen.nig.ac.jp/zebra/ ) and can be searched by strain name. The following strains are available or will be available: Tg(aldoca:NTR-TagRFPT)nub125Tg [registered as Tg(aldoca:NTR-TagRFPT) ], hspzGFFgDMC156A , Tg(5xUAS-hsp70l:GCaMP6s)nkUAShspzGCMP6s13aTg [registered as UAShspGCaMP6s13A ], and Tg(aldoca:GtACR1-EYFP)nub86Tg [registered as Tg(aldoca:GtACR1-EYFP) ]. Tg(aldoca:GAL4FF, myl7:mCherry)nub127Tg is in the process of deposition and will be available as Tg(aldoca:GAL4FF, myl7:mCherry) . All lines can be requested through NBRP Zebrafish. Articles from Science Advances are provided here courtesy of American Association for the Advancement of Science ACTIONS View on publisher site PDF (3.3 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Probing the linguistic cerebellum: a qualitative review of the effects of cerebellar neurostimulation on language processing.#412111
• Functional neural oscillatory activities reveal the impact of attentional instructions on speech auditory feedback control.#222859
• Reconsidering music in stroke rehabilitation: a scoping review from auditory stimulus to relational process.#301695
• Cerebellum violates Marr-Albus predictions to train synapses on long-term anticipatory goals#214787
• A basolateral amygdala to dorsolateral striatum projection modulates stimulus-evoked motor behavior.#1147183