To record intended locomotion in immobilized animals, larval zebrafish were paralyzed by immersion in a solution containing bungarotoxin, and either head-embedded in agarose, or suspended in mid-water from three suction pipettes (Methods). Next, two or four suction electrodes were attached to the tail at intersegmental boundaries to record multi-unit activity from motor neuron axons (Figure 1A) (Masino and Fetcho, 2005; Ahrens et al., 2012). (For four electrodes, on each side of the tail, the electrode with the higher signal to noise ratio was selected and the other was ignored). Visual stimulation was delivered using a display underneath the fish. To examine the temporal patterns of motor nerve activity during intended locomotion, we made use of a robust reflexive behavior known as the OMR (Orger et al., 2008). In the OMR, animals respond to whole-field visual motion by turning and swimming along with the direction of the moving stimulus (Figure 1B), which is thought to represent the fish's attempt to maintain a constant position relative to the ground in the presence of a water current. By presenting forward moving gratings, which normally evoke forward swimming, it became apparent that the electrical signature of forward swim bouts consists of symmetrically alternating activity on the left and right sides of the tail at a similar period as that of tail oscillations in freely swimming fish (Figure 1D, left; Figure 1E; pixelized fish indicates fictively behaving animal). Conversely, by presenting left-moving gratings, we observed high-amplitude, long-duration bursts of activity on the left electrode at the start of most swim bouts (Figure 1D, right; Figure 1E), which we interpret as representing the relatively stronger tail bend that initiates a turning behavior (Budick and O'Malley, 2000). These patterns were consistently present during presentation of whole-field motion in the forward, left, and right directions (Figure 1E). Further, after examining eight angles of whole-field motion, we found that statistics of these electrical signals (Figures 1F–H) qualitatively matched behavioral features of freely swimming animals responding to whole-field motion (Orger et al., 2008). This qualitative correspondence encouraged us to attempt to translate the electrical signals into a measure of motor intent (direction of swim), and to use this inferred intent to control its movement in a virtual environment.

To construct a measure of intended swimming and turning, we hypothesized that the difference in power between the left and the right recording channels would provide information about the direction of intended locomotion. Figure 1I shows the difference in power among swims recorded during presentation of whole-field motion in eight different directions. A systematic shift in the power difference is observed when the angle of stimulation changes, peaking during backward-left and backward-right motion, consistent with freely swimming behavior (Orger et al., 2008). During forward motion, the left and right power equilibrates, as is expected from the symmetric undulations of the tail during forward swimming. We also surveyed eight other decoding strategies based on various features of fictive swim bouts (Figure 2). We found that the decoder that was based on the difference in power performed the best. Moreover, this decoder can compute angle change in real time by incrementally increasing or decreasing the heading angle according to the difference in power during each video frame. In contrast, decoders based on parameters such as the duration of the first oscillation of the swim bout only produce an angle change after the first oscillation has terminated. Thus, we constructed a swim decoder where the turning angle was taken to be proportional to a weighted difference in power on the left and right side of the tail and the strength of the forward swim was taken to be proportional to the sum of the power on both sides of the tail (Methods). We next tested whether this decoder is sufficient for driving closed-loop behavior in virtual environments.

The above motor patterns were measured in open-loop, where a fixed stimulus is presented and the fictive behavior monitored. Next we investigated if these results can be used to simulate two-dimensional swimming. To this end, we established a closed-loop swim simulator (Figure 1C), using the above measures as a readout of intended turning and forward swimming. In short, activity on the left side of the tail drove left rotation of the fish in a virtual environment (i.e., right rotation of the visual environment), activity on the right side drove right rotation of the fish (left rotation of the visual environment), and activity on both sides drove virtual forward swimming (backward movement of the visual environment). After tuning the three parameters of the decoder (Methods), i.e., the constants of proportionality for left turns, forward swims, and right turns, the OMR was evoked in a virtual environment and the behavior of the fish compared to the analogous free swimming behavior.

In the free swimming OMR, a whole-field visual stimulus, in this case black and white bars, is moved at a fixed velocity in a fixed direction, which causes the zebrafish to turn and swim along the direction of motion (Video S1). Figure 3A shows the behavior of three fish performing the OMR. The fish swam around in a 10 cm diameter petri dish that was positioned above a display screen (Methods). At a time point when the fish was close to the edge of the dish (but not touching), whole-field motion was initiated in the direction perpendicular to the wall at the location of the fish. In Figure 3A, the paths traced out by the fish are rotated so that the starting position is always at the bottom of the petri dish and the direction of the motion stimulus is upward. Consistent with previous reports of the OMR (Orger et al., 2008), the fish turned and swam in the direction of motion, with the angle between the heading direction and the direction of motion over all 15 s trials centered around 0° (Figures 3B,C). At initiation of stimulus movement, the fish faced a random direction (Figure 3C, dark blue). They then turned to adjust their heading direction toward that of the motion (Figure 3C, blue-red gradient), and this process equilibrated after about 5–10 s (Figure 3C, yellow-red). With this characterization of the free swimming OMR, we now turned to the analysis of the fictive OMR in the virtual environment.

In the virtual reality OMR paradigm (see Video S2 for a more intuitive demonstration of the closed-loop dynamics), the fish was similarly faced with whole-field visual motion, and the virtual fish was periodically “spun round” to randomize the heading direction. The direction of visual motion is then controlled in closed loop by the fish's fictive swims (Figure 1C). For example, if the direction of optic flow is to the left and the fish makes a fictive turn to the left, we rotate the visual environment to the right, resulting in the new direction of optic flow being forward (tail to head). This new optic flow can be slowed down or even accelerated backward by a fictive forward swim, simulating forward swimming. Tracing the virtual swim paths in the simulated environment, a similar pattern appears as in freely swimming fish: paralyzed fish efficiently turn and swim along with the direction of optic flow (compare Figure 3D to the free swimming case in Figure 3A).

We measured several parameters of the OMR in the virtual environment and found that the fictive swimming behavior of paralyzed zebrafish is strikingly similar to that of freely swimming fish. The heading angle is clustered around the angle of visual motion, as in the freely swimming case, and to a similar extent (compare the fictive case in Figures 3E,F to the free case in Figures 3B,C). When the direction of motion in the virtual environment is changed to a random angle, the fictively swimming fish corrects for this over the course of 5–10 s, as is also observed in the freely swimming fish (compare Figure 3F to Figure 3C).

Given the open-loop results (Figure 1), it is not surprising that the paralyzed fish intend to turn and swim in the expected direction. It is, however, remarkable that these neural measures of motor output can be used in a straightforward manner to simulate a two-dimensional assay, such that the fish displays naturalistic behavior, with multiple measures closely matched between free and paralyzed animals.

We imaged the entire hindbrain while zebrafish performed the two-dimensional OMR, in search of neural correlates of left and right optomotor behavior, in a transgenic fish, Tg(elavl3:GCaMP5) (Ahrens et al., 2013), expressing the genetically encoded calcium indicator GCaMP5 (Akerboom et al., 2012) under the elavl3 promoter (Park et al., 2000) in almost all neurons. To this end, we initialized the direction of whole-field motion to 90° (left) and −90° (right) with respect to the fish's body (Figure 4A: left initializations at 0 s and 10 s, right at 20 s and 30 s, with two 40 s repetitions per imaged plane). Subsequently, the fish was able to control its movement in the virtual environment and swim along with the direction of motion. We were able to detect differences in neural activity during the left and right phases by reconstructing the entire hindbrain volumes from the individually imaged planes and subtracting the volumes corresponding to these two periods from one another. Figure 4B shows the hindbrain volume (each panel is the average of three consecutive 5 μm planes; see also Video S3). A left-right segregation of activity during left and right optomotor behavior is observed, including in the inferior olive (white arrows). In the cerebellar cortex, this lateralization partially reverses (blue arrows), probably because climbing fibers from the inferior olive cross to the contralateral side before projecting to the cerebellar cortex. Figure 4C shows activity of the two neural populations, averaged over the 40 s periods. Thus, activity in entire brain regions, and by extension in the entire brain, can be mapped with a two-photon microscope during fictive navigation.

To illustrate the importance of environmental feedback, we compared behavior in open-loop and closed-loop conditions. A closed-loop period was followed by two repetitions of the stimulus in open loop, so that the visual patterns in both open and closed loop were identical, but in closed loop were self-generated. Figure 5 shows two neurons in the hindbrain, each exhibiting different temporal patterns of activity during closed-loop and open-loop replay. Activity of neuron b is determined by motor output; this neuron may drive behavior or receive feedback from motor centers. Importantly, behavior during closed- and open-loop is different (double arrows vs. single arrows; behavior during open loop tends to be more vigorous), so that the function of this neuron cannot be investigated solely by presenting visual stimuli in open-loop. On the other hand, neuron c appears to be driven by visual stimulation, with similar activity patterns during open-loop replay; however, differences still exist, suggesting that this neuron may be partially driven by behavior or other internal processes.

Next, we used phototaxis as a second behavior to test the capabilities of the fictive-virtual paradigm for mimicking natural behavior. Larval zebrafish are attracted to light and avoid darkness (Burgess and Granato, 2007), a behavior that may allow them to avoid deep water, the shadows of predators, and dimly lit areas where it may be hard to detect both predators and food. To quantify phototaxis in freely swimming fish we observed their swimming behavior in an environment consisting of bright and dark areas (Figure 6A), (Methods). We found that animals spent significantly more time in the bright areas than in the dark areas (Figures 6A–D; p < 0.001, t-test, N = 10 fish), although variations existed across the 10 fish tested (Figure 6C).

Transitioning to the fictive virtual-reality paradigm, we implemented a similar environment, consisting of infinitely repeating black tiles on a white background, through which the fish could fictively swim. Unlike in the fictive OMR paradigm, there was no constant flow present, so that the visual environment remained stationary in the absence of fictive locomotion. In this virtual environment, most fish avoided the dark areas (Figures 6E–I, Video S4), preferring to stay in the bright regions (p < 0.001, t-test, N = 12 fish).

Some freely swimming fish avoided the bright inner region of the arena in addition to the dark region (e.g., Figures 6A and C, fish 1,2). Similar behavior was present in some fictively navigating fish (e.g., Figures 6F and H, fish 4). Overall, the degree of darkness avoidance was stronger in freely swimming fish than in fictively behaving fish, as can be seen from the comparison of the normalized occupancy of the dark and bright regions show in Figures 6D,I. This demonstrates that not all quantifiable elements of a behavior are guaranteed to equate between the freely swimming to the fictive case, even if the core features of the behavior are qualitatively preserved.

In summary, we have shown that paralyzed larval zebrafish can traverse a light-modulated environment and show similar attraction and repulsion behavior as freely swimming fish. Combined with the results from the optomotor behavior, this indicates that fictive navigation robustly recapitulates natural behavior.

We have shown that paralyzed zebrafish are able to meaningfully traverse two dimensional environments. We asked whether the activity of neurons in the forebrain carry representations of features of the environment, and whether these can be detected by two-photon calcium imaging while the animals are navigating. In the light-modulated virtual environment, we sought neurons that responded to bright and dark features, with a particular focus on the habenula and pallium (Northcutt et al., 2004; Zhao and Rusak, 2005), and found activity that was both positively and negatively modulated by light level in both regions.

Figure 7A shows the location of a neuron in the habenula of a Tg(elavl3:GCaMP5) fish traversing the light-modulated environment. The activity of this neuron fluctuates over time and is partially determined by whether the fish is on top of a dark patch or bright region (Figure 7B). Representing neuronal activity as a function of location in Figure 7C, elevations in the calcium readout can be observed as the fish enters a dark patch (7 minutes of data shown). Since the environment consists of repeating tiles, ΔF/F can be represented as an average over a single tile (Figure 7D).

Subpopulations of neurons in the pallium also had excitatory responses to darkness, which could be detected during virtual navigation (Figures 7E,F, green). Another population (Figure 7F, red; Figure 7G) responded to light and was suppressed by darkness (5,6 in pallium; 1–4 in anterior midbrain, 1 and 4 are neuropil, the rest are neurons). The fact that activity in these neurons—likely modulated by light level around the fish, but possibly also by other features (e.g., panel E, compare neurons 1,2,4 to neurons 3,5,6)—can be mapped to spatial location as the fish fictively navigates opens up the possibility to study more general representations of features of the environment across the entire brain.

To compare closed- to open-loop conditions, we show in Figures 8A–C the activity of three neurons in the pallium during navigation through the light-modulated environment, and the subsequent stimulus-replay period. Neuron a is faithfully driven by the stimulus—its activity pattern during stimulus replay is identical to the activity during navigation. Neurons b and c are also sensory neurons, with similar activity during navigation and stimulus replay; however, during replay, activity is modulated differently, which may be due to a change in behavioral state, possibly induced by the open-loop condition. For comparison, neuron d, recorded in a different fish during an optomotor paradigm (Ahrens et al., 2012), exhibits very different activity during open-loop stimulus replay, evidenced by the change in behavioral patterns during closed- and open-loop (Figure 8D, bottom). Together with Figure 5, these results emphasize the fact that stimulus delivery alone is not sufficient for investigating neural activity during sensorimotor transformations.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.