Ruben Portugues

Innate sensorimotor policies enable rapid and reliable behavior but can become ineffective when environments change. How these policies interact with feedback-driven processes to support behavioral flexibility remains un-clear. Using closed-loop virtual reality in larval zebrafish, we reversed the visual consequences of self-generated turns. Larvae initially followed their innate mapping and then rapidly switched to corrective turns in the opposite direction. Purkinje cell ablation abolished this correction while preserving innate optomotor responses. Imaging revealed outcome-invariant sensorimotor representations in granule cells and an outcome-dependent modulation of inferior olive neurons and Purkinje cells. Correction weakened as the optic-flow angle increased, despite persistent error-related representations. The innate behavioral repertoire also shifted as the optic-flow angle increased, making corrective turns less likely. An actor-critic model initialized with the measured behavioral repertoire reproduced the limits of correction. Together, these results show that adaptive behavior emerges from the interaction between an innate feed-forward policy and an olivo-cerebellar instructive signal, revealing innate behavioral structure as a fundamental constraint on cerebellar adaptation.

Improving behavioral performance relies on the ability to associate decisions with their positive and negative outcomes. Although neurons that associate actions with their consequences have been identified across multiple brain regions, the circuit-level mechanisms underlying this integration remain poorly understood. Here, using an operant thermoregulatory assay in larval zebrafish, we show that the dorsal habenula-interpeduncular nucleus (dHb-IPN) pathway is necessary for associating actions with subsequent thermal reward. A population of intermediate IPN (iIPN) neurons encodes actions only when they lead to reward, suggesting a role in establishing action-outcome associations. This integration is temporally precise, requiring actions and their outcomes to occur within approximately one second in both the behavioral assay and iIPN responses. We then combine circuit mapping and axon imaging to show that actions and reward signals are conveyed by GABAergic prepontine and glutamatergic dHb neurons, respectively. Finally, the integration between action and reward signals relies on presynaptic GABA_B receptor-mediated modulation of dHb axon terminals by prepontine neurons. These results link a crucial computation for adaptive behavior to a specific circuit mechanism.

The ultimate goal of sensory coding is to extract and represent the cues required for adaptive motor output. This suggests that sensory codes and behavioral outcomes may align, and a variety of studies have argued that both biological and engineered sensory systems represent stimuli similarly when they play similar roles in behavior. However, the extent to which behavioral demands determine sensory coding throughout the brain is largely unknown. Here we propose that behavioral alignment is a general principle that organizes sensory representations, and we show that carefully measured behavior can predictively account for visual encoding across the entire zebrafish brain. We discover population codes that represent visual motion stimuli according to the optomotor responses elicited by them, indicating that information required for behavioral selection is explicitly encoded in sensory populations. Brain-like neuronal responses result when sensory codes are optimized for efficiency and aligned to behavior. These results provide a paradigm for understanding sensory representations through the behaviors they drive.

Vision can inform animals as they navigate their environment. Landmarks can be used to maintain heading, while optic flow can be integrated to estimate turning. Although it has been shown that head direction (HD) neurons in various species use these visual cues^1,2, the circuit mechanisms underlying this process in vertebrates remain unknown. Here we asked whether and how the recently identified HD cells in the larval zebrafish³, one of the smallest vertebrate models, incorporate visual information. By combining two-photon microscopy with a panoramic virtual reality setup, we demonstrate that the zebrafish HD cells can reliably track the orientation of multiple visual scenes, exploiting both visual landmarks and optic flow cues. The mapping between landmark cues and heading estimates is idiosyncratic across fish and experience dependent. Furthermore, we show that landmark tracking requires the lateralized projection from the habenula to the interpeduncular nucleus⁴, a structure innervated by HD neuron processes³. The physiological and morphological parallels suggest that a Hebbian mechanism similar to the fly ring neurons^5,6 is at work in the habenula axons. Overall, our observation that hindbrain HD cells of larval zebrafish can utilize the visual cues despite the lack of an elaborate visual telencephalon sheds new light on the evolution of navigation circuitry in vertebrates.

From insects to fish to mammals, many species have an internal compass: a set of recurrently connected neurons that combine motor feedback, vestibular signals, and external cues to compute the animal′s heading direction. Whether the underlying mechanism is universal across different species is unresolved. In Drosophila, for instance, the central complex contains three connected, yet anatomically separate, functional neuron rings. Two rings receive countervailing velocity signals that shift neuronal activity bumps around all three rings. An alternative, classical continuous-attractor model invokes only one ring but requires velocity-modulated synaptic connections. A single-ring compass has been discovered in the anterior hindbrain of zebrafish. Using theory and experiments, we find, however, that three rings, including two shifter rings, are intermingled on the same anatomical scaffold. Zebrafish and Drosophila, therefore, use the same basic compass mechanism, suggestive of convergent evolution.

Sensory information is fundamental for navigation. Visual motion is used by animals to estimate their traveling distance and direction, and landmarks allow animals to tether their location and orientation to their environment. How such signals are integrated in the vertebrate brain is poorly understood. Here we investigate the representation of directional whole field visual motion and landmark position in the larval zebrafish head direction circuit. Using calcium imaging we show that these stimuli are represented in the habenula, interpeduncular nucleus and anterior hindbrain. In the dorsal interpeduncular nucleus, both stimuli are topographically arranged and align with the representation of the heading signal. Neuronal ablations show that the landmark responses, but not the whole field motion responses, require intact habenula input. Our findings suggest the interpeduncular nucleus as a site for integration of the heading signal with visual information, shedding light on how navigational signals are processed in the vertebrate brain.

Despite the large evolutionary distance between vertebrates and insects, the visual systems of these two taxa bear remarkable similarities that have been noted repeatedly, including by pioneering neuroanatomists such as Ramón y Cajal. Fuelled by the advent of transgenic approaches in neuroscience, studies of visual system anatomy and function in both vertebrates and insects have made dramatic progress during the past two decades, revealing even deeper analogies between their visual systems than were noted by earlier observers. Such across-taxa comparisons have tended to focus on either elementary motion detection or relatively peripheral layers of the visual systems. By contrast, the aims of this Review are to expand the scope of this comparison to pathways outside visual motion detection, as well as to deeper visual structures. To achieve these aims, we primarily discuss examples from recent work in larval zebrafish (Danio rerio) and the fruitfly (Drosophila melanogaster), a pair of genetically tractable model organisms with comparatively sized, small brains. In particular, we argue that the brains of both vertebrates and insects are equipped with third-order visual structures that specialize in shared behavioural tasks, including postural and course stabilization, approach and avoidance, and some other behaviours. These wider analogies between the two distant taxa highlight shared behavioural goals and associated evolutionary constraints and suggest that studies on vertebrate and insect vision have a lot to inspire each other.

The visual stabilization behavior in the larval zebrafish reflects the history of optic flow experienced in the recent past. This integrative process has gained traction in recent years as a simplified, tractable model of working memory and decision-making. Yet, its algorithmic bases are poorly understood. In the present study, we first demonstrate that fish integrate externally generated optic flow over time, disregarding self-generated optic flow. This observation suggests that the fish integrates optic flow to better estimate the state of the environment, rather than to keep track of their position. Second, through reverse correlation and delay-based paradigms, we reveal multiple timescales involved in the optic flow integration. With the help of quantitative modeling, we show that the fish becomes more forgetful about the past optic flow in a more dynamic visual environment. Next, with whole-brain, light-sheet calcium imaging, we find that optic flow selective neurons throughout the brain exhibit neural signatures of motor efference copies, mirroring the behavioral findings. Lastly, with two photon calcium imaging, we show that inferior olive neurons integrate forward and backward flow separately, giving clues to how the multiple timescales of the optic flow integration is implemented. Overall, the results here refine our algorithmic and functional understanding of the history dependence of the visual stabilization behaviors in the larval zebrafish, paving the way for deciphering its circuit implementations.

Improving behavioral performance relies on the ability to associate decisions with their positive and negative outcomes. Although neurons that associate actions with their consequences have been identified across multiple brain regions, the circuit-level mechanisms underlying this integration remain poorly understood. Here, using an operant thermoregulatory assay, we show that larval zebrafish maintain a short-term memory of action-outcome associations and that the dorsal habenula-interpeduncular nucleus (dHb–IPN) pathway is necessary for this process. Consistently, a population of intermediate IPN neurons encodes actions only when they lead to a thermal reward, suggesting a major role in establishing short-term associations and influencing subsequent decisions. We then combine circuit mapping and axon imaging to show that actions and reward signals are conveyed by GABAergic prepontine and glutamatergic dHb neurons, respectively. Finally, the integration between motor and sensory streams relies on presynaptic GABA _B receptor–mediated modulation of dHb axon terminals by prepontine neurons. These results link a crucial computation for adaptive behavior to a specific circuit mechanism.