Highlights d Zebrafish hunting consists of stereotyped transitions in a behavioral continuum d Chaining successive hunting bouts depends on short-term high-acuity visual cues d Larvae capture prey from below with stereotyped jaw and tail movements d Precise positioning of prey in the strike zone depends on binocular visual cues

Retinal axon projections form a map of the visual environment in the tectum. A zebrafish larva typically detects a prey object in its peripheral visual field. As it turns and swims towards the prey, the stimulus enters the central, binocular area, and seemingly expands in size. By volumetric calcium imaging, we show that posterior tectal neurons, which serve to detect prey at a distance, tend to respond to small objects and intrinsically compute their direction of movement. Neurons in anterior tectum, where the prey image is represented shortly before the capture strike, are tuned to larger object sizes and are frequently not direction-selective, indicating that mainly interocular comparisons serve to compute an object’s movement at close range. The tectal feature map originates from a linear combination of diverse, functionally specialized, lamina-specific, and topographically ordered retinal ganglion cell synaptic inputs. We conclude that local cell-type composition and connectivity across the tectum are adapted to the processing of location-dependent, behaviorally relevant object features.

No abstract

Visual objects naturally compete for the brain's attention, and selecting just one 16 of them for a behavioural response is often crucial for the animal's survival 1 . The 17 neural correlate of such stimulus prioritisation might take the form of a saliency 18 map by which responses to one target are enhanced relative to distractors in 19 other parts of the visual field 2 . Single-cell responses consistent with this type of 20 computation have been observed in the tectum of primates, birds, turtles and 21 lamprey 2-7 . However, the exact circuit implementation has remained unclear. 22Here we investigated the underlying neuronal mechanism presenting larval 23 zebrafish with two simultaneous looming stimuli, each of which was able to 24 trigger directed escapes on their own. Behaviour tracking revealed that the fish 25 respond to these competing stimuli predominantly with a winner-take-all 26 strategy. Using brain-wide functional recordings, we discovered neurons in the 27 tectum whose responses to the target stimulus were non-linearly modulated by 28 the saliency of the distractor. When the two stimuli were presented monocularly 29 in different positions of the visual field, stimulus selection was already apparent 30 in the activity of retinal ganglion cell axons, a likely consequence of antagonistic 31 mechanisms operating outside the classical receptive field 8,9 . When the two 32 stimuli were presented binocularly, i.e., on opposite sides of the fish, our 33 analysis indicates that a loop involving excitatory and inhibitory neurons in the 34 nucleus isthmi (NI) and the tectum weighed stimulus saliencies across 35 hemispheres. Consistent with focal enhancement and global suppression, 36 glutamatergic NI cells branch locally in the tectum, whereas GABAergic NI cells 37 project broadly across both tectal hemispheres. Moreover, holographic 38 optogenetic stimulation confirmed that glutamatergic NI neurons can modulate 39 visual responses in the tectum. Together, our study shows, for the first time, 40 context-dependent contributions of retinotectal and isthmotectal circuits to the 41 computation of the visual saliency map, a prerequisite for stimulus-driven, 42 bottom-up attention. 43Dark, looming stimuli are strongly aversive stimuli for zebrafish larvae 10,11 and other 44 animals 12,13 , probably mimicking an approaching predator or an object on a collision 45 course. In our setup, single looming disks presented from below and on one side of a 46 free-swimming animal were highly effective in driving an escape response to the 47GABAGABA ? Midline Excitatory tectal neurons Inhibitory tectal neurons Eye NI Tectum NI Fee bac ro ections (NI to tectum) Tectum to NI

23Animals build behavioral sequences out of simple stereotyped actions. A comprehensive 24 characterization of these actions and the rules underlying their temporal organization is necessary 25 to understand sensorimotor transformations performed by the brain. Here, we use unsupervised 26 methods to study behavioral sequences in zebrafish larvae. Generating a map of swim bouts, we 27 reveal that fish modulate their tail movements along a continuum. We cluster bouts that share 28 common kinematic features and contribute to similar behavioral sequences into seven modules. 29Behavioral sequences comprising a subset of modules bring prey into the anterior dorsal visual 30 field of the larvae. Fish then release a capture maneuver comprising a stereotyped jaw movement 31 and fine-tuned stereotyped tail movements to capture prey at various distances. We demonstrate 32 that changes to chaining dynamics, but not module production, underlie prey capture deficits in 33 two visually impaired mutants. Our analysis thus reveals the temporal organization of a vertebrate 34 hunting behavior, with the implication that different neural architectures underlie prey pursuit and 35 capture. 36 Patterson et al., 2013;Szigeti et al., 2015). In either case, actions must be chained into sequences 59 that reliably achieve the desired goal of the animal. Stereotyped, reproducible behavioral 60 sequences have been explained with serial models, in which one action triggers the next in the 61 chain via feed-forward neural mechanisms (Long et al., 2010). In contrast, flexible sequences, in 62 which the ordering of modules might be different each time the behavior occurs, have been 63 explained using hierarchical models. In hierarchical models, switching between behavioral 64 modules is stochastic, but may be influenced by longer-term behavioral states or sensory stimuli 65 received by the animal (Berman et al., 2016;Seeds et al., 2014;Tao et al., 2019; Wiltschko et al., 66 2015). 67 68Capturing prey is an essential behavior for the survival of many animals and is innate. The 69 behavior is also complex, requiring the localization, pursuit and capture of a prey object, often 70 encode bouts with similar postural dynamics. In this space, tight clusters would suggest that 131 larvae can only generate a limited number of stereotyped bout types, whereas a diffuse cloud 132 would suggest that larvae can continuously modulate the kinematics of their bouts. To distinguish 133 these possibilities, we developed a pipeline for determining the structure of the behavioral 134 manifold (Figure 2A; see Methods). Our algorithm consists of three steps: alignment, clustering 135 and embedding. In the first step, we calculate the distance between each pair of bouts in the 136 three-dimensional postural space using dynamic time warping (DTW) (Jouary and Sumbre, 2016; 137