Distributed plasticity drives visual habituation learning in larval zebrafish

Summary

Habituation is a simple form of learning, where animals learn to reduce their responses to repeated innocuous stimuli [1]. Habituation is thought to occur via at least two temporally and molecularly distinct mechanisms, which lead to short-term memories that last for seconds to minutes, and long-term memories that last for hours or longer [1,2]. Here we focus on long-term habituation, which, due to the extended time course, necessitates stable alterations to circuit properties [2–4]. In its simplest form, long-term habituation could result from a plasticity event at a single point in a circuit, and many studies have focused on identifying the site and underlying mechanism of plasticity [5–10]. However, it is possible that these individual sites are only one of many points in the circuit where plasticity is occurring. Indeed, studies of short-term habituation in C. elegans indicate that in this paradigm multiple genetically separable mechanisms operate to adapt specific aspects of behaviour [11–13]. Here, we use a visual assay in which larval zebrafish habituate their response to sudden reductions in illumination (dark flashes) [14,15]. Through behavioural analyses, we find that multiple components of the dark flash response habituate independently of one another using different molecular mechanisms. This is consistent with a modular model, in which habituation originates from multiple independent processes, each adapting specific components of behaviour. This may allow animals to more specifically or flexibly habituate based on stimulus context or internal states.

Graphical Abstract

Animals habituate to repeated stimuli, learning to ignore potential distractions. Randlett et al. show that multiple components of the larval zebrafish dark flash response habituate independently using different molecular mechanisms. This reveals that habituation learning is a modular process that selectively adapts specific components of behavior.

Results

High-throughput quantification of dark flash habituation

When exposed to a sudden transition to whole-field darkness (dark flash), larval zebrafish startle. These startles are characterized by a large angle body bend followed by a swim forward in the new direction (Figure 1A, Video S1)[16]. These movements were originally classified as “O-bend” maneuvers, but a recent behavioural clustering analysis indicates that a second kinematically distinct movement termed a Spot Avoidance Turn (SAT) can also be elicited by dark flashes [17]. Although the neural circuitry underlying dark flash response has not been well described, it is known to be retina-dependent [18]. At the reticulospinal level, the dark flash response does not require the Mauthner neuron that drives the acoustic escape response [16], but does require the smaller and more ventromedially located reticulospinal neurons that also drive spontaneous turning behaviours (RoV3, MiV1, and MiV2) [19,20]. Importantly, larval zebrafish exhibit protein synthesis-dependent long-term habituation to dark flashes, which, similar to memory formation in Drosophila and mice, requires Neurofibromatosis 1 (Nf1) dependent cAMP- and RAS-mediated plasticity [15,21,22].

Figure 1. Habituation of the dark flash response.

A) When stimulated with a dark flash, larval zebrafish execute a high-amplitude turn, which habituates with training[14].

B) Time projection images from a 0.9s recording of the same 300 larvae for the first flash (Naive Response), and the 240^th flash (Habituated Response). Motion is visible as orange streaks in the image tracing the path travelled by the larvae. Larvae that do not move are visible as stationary white larvae (insets). Larvae were recorded in 300-well plates, and images were background subtracted to remove the behaviour plates.

C) The response probability across the population of larvae decreased both within the 60-flash training blocks, and successively across the 4 blocks of training. Each dot represents the proportion of larvae that respond to each stimulus, which are delivered at 1 minute interstimulus intervals (ISI). Memory was evident at the Re-test Block 5-hours after training, where larvae have not recovered to untrained levels (those in Block 1). Arrows = 10^th stimulus in each block. See also Figure S1, Video S1, S2.

We developed a high-throughput behavioural setup that can track 600 larvae in individual wells, deliver visual and acoustic stimuli, and track stimulus responses at 560 Hz (Figure 1B, see Methods). This allows us to maintain individual larval identity throughout the experiment, to monitor behaviour over days, and to unambiguously classify stimulus responses using postural reconstruction of the bending axis of the larvae. Adapting an established dark flash habituation assay [14,15], we developed a paradigm that consists of 4 training blocks of 60 dark flashes at 1 minute interstimulus intervals, with blocks separated by 62 minutes of rest. This spaced training paradigm induced habituation, which we quantified as the progressively decreasing probability of executing a dark flash response (Figure 1B,C, Video S2). Fitting curves to each block with an exponential function (Figure 2A) revealed that after each rest period the response returned to near maximal levels, but then decreased more rapidly, and to lower values, during subsequent training blocks. This is consistent with previous observations of long-term habituation, which often manifests as a faster rate of rehabituation [1]. Memory retention was tested in a “Re-test” block 5 hours after training (Figure 1C), which revealed that dark flash responses did not recover to naive levels (those observed in the first block), but rather exhibited even greater reductions compared to the last block of training. We confirmed that this effect is not due to experimental time by comparing larvae that had undergone the 4-block training protocol with controls that had not been trained, either 5 or 28 hours after the training period (Figure S1A). In both experiments, trained larvae responded less frequently, indicating these reductions are based on training experience, though by 28 hours there was a significant recovery towards untrained levels. Therefore, the habituation paradigm presented here induces memory that lasts robustly for 5 hours, and effects are seen for up to 28 hours.

Figure 2. Habituation kinetics of different dark flash response components.

A) Exponential fits of habituation curves for each of the 4 training blocks and the re-test block plotted overlapping in time, colour coded by training block (thick line), thin line = raw data (mean across 3120 larvae), insets = mean response per block, for the probability of executing a dark flash response,

B) the proportion of larvae executing at least two responses,

C) the latency from the stimulus to the initiation of the response (note that the latency values increase, indicating habituation),

D) the proportion of larvae executing a “Simple Response”,

E) the duration of the movement,

F) the displacement of the larvae,

G) the reorientation angle achieved by the movement, and

H) the maximal bend amplitude.

I) Exponential fit curves for block one habituation performance plotted across components. Data was normalized such that initial response is equal to 1, and the minimal response observed in any flash is 0.

J) Violin plots of the distributions of response recovery during the 5 hr. retention window, computed as the mean values across larvae for the trials in the Re-test block, divided by those in the last training block (block 4). Values greater than 1 reflect a recovery of the response. See also Figure S2, Video S3.

Decreasing responsiveness is characteristic of habituation, but it is important to rule out fatigue as an alternative explanation. To that end, we monitored spontaneous movements (when no stimuli are delivered), the response to acoustic tap stimuli, and the ability of larvae to detect and respond to visual motion stimuli using the optomotor response (OMR) [23]. In each case, we did not detect reduced responses in the trained larvae compared to controls, indicating that fatigue, or a generalization of habituation to other behaviours does not occur (Figure S1B–I). In fact, rather than a fatigue induced reduction in motility, we observed small increases in displacement, turning rate and acoustic tap responses after training, indicating that dark flash habituation training may be slightly arousing to the animals. Importantly, as the OMR is also a retina-dependent behaviour that depends on the detection of light/dark transitions in the retina, and as the OMR is unaffected by our training protocol, we conclude that habituation does not affect vision globally. This furthermore indicates that habituation is unlikely to occur at the general sensory neuron/photoreceptor-level, but rather selectively within the dark flash response circuitry.

Multiple dark flash response components habituate

Habituation can be measured as a binary reduction in response (as above), or alternatively via decreases in the magnitude of the response [1]. To further analyse this process, we asked how many other aspects of the response habituate, and subsequently how these might be related to one another. In total we identified 8 components of the dark flash response that habituate (Figure 2A:H, Figure S2A):

1) Probability of responding to the stimulus, as discussed above. 2) Double Responses: zebrafish larvae move in bouts, separated by periods of inactivity. By tracking behaviour over the full 1 second of the dark flash, we observed that many larvae do not execute a single response, but rather execute multiple large-angle turns separated by at least 50 ms of inactivity (Video S2). The proportion of larvae executing at least two responses habituates (Figure 2B). 3) Latency: consistent with previous results[15], we observed that the latency from the stimulus to the response habituates, increasing by an average of 197 msec (first stimulus vs. last stimulus of block 4). Similar to components 1 and 2, we saw progressive accumulation of habituation across blocks, but retention after 5 hours was not as robust, though it still remained habituated compared to naive levels (Figure 2C). We also note that while acoustic responses can be parsed into kinematically distinct response types based on latency (Short- and Long-Latency C-bends [24]), we do not observe a bimodal distribution in the dark-flash response latencies, indicating such relationship does not exist for this response (Figure S2B). 4) Proportion of Simple Responses: in response to a dark flash, some larvae perform multiple high-amplitude bends in the same direction followed by a swim, as opposed to the classic “O-bend” response that involves only one such bend before the swim. We term these “Compound Responses” and “Simple Responses,” respectively (Figure S2C, Video S2). Before training, 61% of the responses are Simple Responses, which decreases to 41% with training (Figure 2D). We observed across-block habituation during training, but retention is poorer than components 1–3, and fully recovers to untrained levels after 5 hours.

We also observed habituation in dark flash response components related to the kinematic magnitude of the movement, including its: 5) Duration, 6) Displacement, 7) Reorientation, and 8) Bend Amplitude (Figure 2E–H). Unlike components 1–4, habituation learning of these latter kinematic components occurred mostly during the first block, and did not decrease appreciably after block 2. Memory in these components was retained after 5 hours, with varying degrees of recovery. These results demonstrate that the dark flash response, rather than being an “all or nothing” response is actually composed of multiple behaviour components capable of adaptation during habituation. We noticed a significant degree of variability in learning and memory kinetics (Figure 2I, J), consistent with the idea that habituation of individual components, rather than resulting redundantly from a single mechanism, instead results from multiple mechanisms with differential kinetics of adoption and decay.

Habituation occurs at multiple circuit loci

The differences in the kinetics of habituation that we have observed could still be explained by a single-site plasticity model, where plasticity occurs at a single locus that is upstream of multiple independent circuit branches. Differences in the synaptic and molecular makeup of these downstream branches could result in differential rates habituation. To further explore the separability of the components of dark flash habituation, we took advantage of spontaneous variability present in our dataset. Namely, while the majority of individual larvae habituate, there is considerable spread in the learning distributions (Figure 3A). We reasoned that if habituation occurs at a single circuit locus, then the learning performance of the different response components would be correlated across larvae. In such a scenario, larvae would vary in their ability to habituate at this locus, but individual larvae would exhibit a consistent level of habituation across all behaviour components. Alternatively, if habituation of individual behaviour components occurs at distinct loci within the circuit, then learning performance should be independent of one another in any individual larva. Consequently, this should result in a lack of correlation in learning performance for individual components across larvae.

Figure 3. Habituation of different response components occurs independently.

A) Percent habituation histograms of individual larvae across the 8 dark flash response components (n=3120 larvae). Asterisk marks the uptick in “Double Responses”, reflecting the individuals that show 100% habituation.

B) Scatter plots and correlation coefficient comparing percent habituation for the “Displacement” and “Duration” components. Colormap reflects the density of points (R = 0.74, p<1×10⁻¹⁰, Spearman’s rho).

C) Same analysis as B, revealing the “Probability” and “Duration” components are not correlated (R = 0.01, p = 0.48, Spearman’s rho).

D) Hierarchically clustered correlation matrix (Spearman’s rho), comparing percent habituation of the response components.

E) Weak but significant anti-correlation observed between the “Bend Amplitude” and “Probability” components (R = −0.17, p<1×10⁻²⁰).

F) Conceptual model for how anti-correlations of habituation performance could arise. If two plasticity loci exist in series and are habituating, in larvae where habituation at the upstream locus is stronger than average, this would result in less than average habituation training at the downstream node, and vice versa.

G) Simulation results modeling plasticity at two loci that follow the kinetics of habituation for ‘Probability’, plus Gaussian noise. Training at the downstream locus depends on how much habituation occurs at the upstream node, resulting in anti-correlations in learning performance. The magnitude of the anti-correlations increases with learning variability at the upstream locus, and decreasing variability at the downstream locus (n = 10000 runs per comparison).

H) Re-sampling of the original 3200 larvae into random 100-larvae subsets over 2000000 iterations. Variance in percent habituation for “Probability” scales with the magnitude of the anti-correlation between percent habituation for “Probability” and “Bend Amplitude (p < 1×10⁻¹⁰, Spearman’s rho). See also Figure S2D.

Our analysis indicates that both scenarios occur. We observed strong positive correlations between some components, such as “Displacement” and “Movement Duration” (Figure 3B). These two components also show similar learning and retention kinetics (Figure 2I,J), further supporting the idea that they habituate through a single mechanism. However, other components, such as “Probability” and “Movement Duration” showed no correlation (Figure 3C), demonstrating that the capacity of a larva to learn to respond less frequently is uncoupled from its ability to learn to respond with a shorter movement. Analysis of all pairwise comparisons (Figure 3D) revealed that learning was largely correlated across the kinematic components (5–8). “Probability” and “Double Responses” are also correlated, while “Latency” and “The Proportion of Simple Responses” are not strongly correlated to any other group. We also observed similar patterns when analyzing how correlated the different components are across fish when analyzing only the responses to the first flash (Figure S2D). This indicates that these components are separately regulated modules of behaviour, and it is not the process of habituation alone that uncouples them. This leads us to conclude that the dark flash response is composed of multiple separately regulated components, and that plasticity exerted at four or more distinct loci acts to independently modulate these components during habituation.

We also observed weak negative correlations between habituation of some response components, most prominently between “Probability” and “Bend Amplitude” (Figure 3E). One possible explanation for such subtle anti-correlations might be due to the circuit architecture of habituation. For example, if different plasticity loci in the circuit operate in parallel, we would expect to observe no correlation in learning performance for their respective components. Alternatively, if the loci are arranged in series, habituation at the upstream locus will reduce the amount of training signal that reaches the downstream locus. Since habituation results from repeated stimulation, this would result in a negative relationship between upstream plasticity and downstream training (Figure 3F). To demonstrate that this can occur, we modelled two habituating neurons connected in series. Both neurons acted as habituation loci with the same habituation kinetics, with random noise added to simulate learning variability. This simulation is in line with the idea that anti-correlated distributions manifest from such an architecture, and that the magnitude of the anti-correlation increases with variability at the upstream plasticity locus (Figure 3G). If we assume from a sensorimotor perspective that the locus that habituates “Probability” is upstream of “Bend Amplitude”, then this simple model predicts that the magnitude of the anti-correlations in habituation for “Probability” and “Bend Amplitude” would increase with the variability for “Probability”. Using iterative sub-sampling of groups of 100 larvae, we observed that indeed, the magnitude of the anti-correlations increases along with the variance in percent habituation for “Probability” (Figure 3H). Combined, these results support a model by which habituation results from distributed effects spread across multiple circuit loci, some of which operate in parallel in the circuit (no correlation in learning performance across larvae), while others operate in series, resulting in negative correlations.

Habituation of different dark flash response components are molecularly separable

Although our results indicate that multiple sites in the dark flash response circuit exhibit plasticity independently during dark flash habituation, it is unclear if these distinct events use separate molecular pathways. If different molecular pathways operate, then it should be possible to identify manipulations that differentially affect different response components. To test this, we analyzed neurofibromatosis 1 (nf1) mutants, which fail to habituate the latency of their dark flash responses [15]. When we analyzed habituation in nf1a;nf1b double homozygous (nf1) mutants, we found that not all components are equally affected (Figure 4A). In fact, while learning performance is strongly inhibited for “Latency” (Figure 4B), learning performance is indistinguishable from controls for “Displacement” (Figure 4C), strongly suggesting that individual components of dark flash habituation are regulated via distinct molecular mechanisms. We note that nf1 mutants also show alterations in the naive response to the first flash for some components, including a significantly longer latency (Figure S3). However, a lack of habituation for “Latency” is not due to a ceiling effect, as sibling controls surpass mutant values during training (Figure 4B).

Figure 4. Habituation of different response components are separable genetically, pharmacologically, by stimulus strength, and circadian phase.

A) Cumulative plot of the differences in habituation rate for the response components. These plots display the cumulative average differences in the mean response across larvae of nf1a;nf1b double homozygous mutants (n=41 larvae) compared to sibling controls (n = 678 larvae). Difference from 0 reflects divergence in response across the 240 dark flash stimuli in the 4 training blocks, with negative values reflecting a failure to habituate. The width of the line are bootstrapped 99.5% confidence intervals. The grey boxed region reflects the expected non-significant effect size from a negative control experiment (see panel I). Mutants fail to habituate some components, most profoundly for the “Latency” metric.

B) Raw data (dots = mean across larvae for each stimulus) and smoothing spline fits (solid lines), demonstrating that nf1a;nf1b double mutants fail to habituate when measuring “Latency” (note that increasing latencies indicate habituation).

C) nf1a;nf1b double mutants habituate normally when measuring “Displacement”.

D) Cumulative difference plots for treatment with Haloperidol (10 μM, n = 80 larvae) vs. vehicle controls (0.1% DMSO, n = 140 larvae).

E) Treatment with haloperidol increases habituation performance when measuring “Latency”, and

F) decreases habituation when measuring “Bend Amplitude”.

G) Cumulative difference plots comparing 0.1% DMSO vehicle controls and after treatment with Pimozide (1μM, n = 140 treated larvae, n = 160 controls), and

H) Clozapine (10 μM, n = 120 treated larvae, n = 160 controls).

I) Negative control experiment comparing wild type larvae in the same experiment that are given no treatments (n = 150 larvae, both groups). These plots do not diverge consistently from +/− 0.05, which is taken as an empirically derived threshold for a meaningful effect size.

J) Cumulative difference plots comparing larvae given an 80% dark flash (light intensity transitions from 100% to 20%), with larvae given a normal 100% dark flash (n = 300 larvae, both groups).

K) Raw data (dots = mean across larvae for each stimulus) and smoothing spline fits (solid lines), showing how larvae habituate more rapidly to the weaker stimulus for the “Probability” component, while

L) habituating the “Bend Amplitude” similarly to controls

M) Cumulative difference plots comparing larvae in the subjective night phase of the circadian cycle with those in the subjective day (n = 150 larvae, both groups).

N) Larvae in the subjective night phase show increased habituation performance for the “Probability” component, and

O) much weaker effects on “Bend Amplitude”. See also Figure S3

To further generalize these findings, we next performed a set of pharmacological manipulations. Due to their previously identified roles in zebrafish behavioural plasticity and habituation, we tested antipsychotic drugs that act as antagonists of the dopamine and serotonin systems [14,24,25]. Specifically, treatment with Haloperidol, a dopamine D2 receptor antagonist, had a wide range of effects on habituation (Figure 4D). Remarkably, these effects include oppositely signed effects for different response components, such as increased habituation for “Latency” (Figure 4E) and decreased habituation for “Bend Amplitude” (Figure 4F). Similarly, treatment with Pimozide and Clozapine, which also antagonize the dopamine D2 receptor, had separable effects across different behaviour components (Figure 4G,H). These pharmacological experiments confirm that habituation of different response components occurs via different molecular mechanisms.

Habituation of different response components are separably modulated by stimulus strength and the circadian rhythm

While we have shown that habituation of different response components is independent, it was unclear what utility such independence might serve. We reasoned that having a modular system might allow an animal to adapt its behaviour with more flexibility or specificity in different contexts. To test this, we first asked how habituation rates change when the stimulus is weakened. Instead of delivering a full dark flash, we decreased the illumination by only 80%. This weakened stimulus is still strong enough to reliably elicit responses (Figure S3F, [16]), and causes the larvae to habituate more rapidly (Figure 4J–L). However, the effect was selective for the “Probability” component, while other components, including “Bend Amplitude”, showed much less modulation. This indicates that the nature of the stimulus can alter the habituation rate of different behavioural components in different ways.

Finally, to further test our hypothesis that a modular system enables animals to adapt habituation behaviour in a more context-dependent manner, we compared habituation rates during different phases of the circadian rhythm. The circadian phase modulates the endogenous arousal level of zebrafish larvae [26], and may also alter the salience of a dark flash since darkness is an expected condition at night. Specifically, we raised a subset of larvae on a reversed night/day light cycle and subjected them to our habituation assay together with their normally raised siblings. By testing the behaviour of larvae during either their subjective day, or their subjective night, we found that there was a circadian influence on habituation (Figure 4M–O). Similar to the effect of a weakened stimulus, we identified selective effects on the habituation rate of the “Probability” component, which showed significantly stronger habituation during the night phase. This could allow the larvae to more rapidly cease responding during the night, or to weaker dark flashes, while continuing to adapt kinematic-related components at the normal rate. Thus, modularity in habituation can allow for the adaptation of specific behavioural components based on both the context of the animal and the stimulus.

Discussion

Studies of habituation in several species have focused on identifying the site and underlying mechanism of plasticity. These efforts have generally associated habituation with plasticity in upstream sensory-related brain areas, including: depression of the sensory-to-motorneuron synapse in the Aplysia Gill/Siphon withdrawal reflex [5,6], depression of the sensory-to-interneuron synapse in C. elegans [8], and enhanced GABAergic inhibition in olfactory glomeruli in the Drosophila antennal lobe [7]. However, one recent study in mice associated habituation to a visual stimulus with synaptic potentiation in the visual cortex, indicating that habituation does not occur via sensory depression in this system [10]. While there is substantial data confirming the importance of these specific loci, it is possible that these individual sites are only one of many points in the circuit where plasticity is occurring.

The experiments presented here quantify behaviour across thousands of larval zebrafish and demonstrate that dark flash habituation occurs via multiple plasticity events, where each of these events acts to suppress a specific component of behaviour. These different plasticity events manifest in differential kinetics of learning and forgetting and a lack of correlated learning across behavioural components. Particularly surprising was the separability of “Probability” and “Latency”, since in the simplest model, “Latency” is a direct function of “Probability” (as in a Poisson process, similar to [27]). Since the brain can adapt these two aspects of behaviour separately, the decision of whether to respond to a stimulus appears to be uncoupled from the decision of when to respond. While a previous study in bullfrogs found similar rates of habituation for different behavioral components in a population of animals [28], our method examined correlation of response component habituation within individuals, and thus serves as a more direct test for a common underlying site of plasticity.

Our results indicate that the brain not only implements plasticity in multiple circuit loci, but also does this via multiple molecular mechanisms. We found that some components of habituation require Nf1, while others do not. Furthermore, our experiments with antipsychotic drugs indicate that habituation of only a subset of components involves signalling through dopamine and/or serotonin receptors. This opens the path for a whole series of detailed investigations on the precise nature and mechanistic role of these pathways that will be the subject of future studies. However, as these drugs all antagonize the dopamine D2 receptor, and all increase habituation for “Latency”, it is likely that dopamine signalling negatively regulates this aspect of habituation. The effects of opposite sign that a single drug can have across different components suggest that the same molecular pathways are capable of influencing different plasticity events in oppositely signed manners. Alternatively, these effects may relate to the promiscuous nature of these drugs, which can affect many targets [29]. Considering that these drugs are used to treat schizophrenia, and that there are well-established connections between habituation and schizophrenia (as well as other psychiatric disorders including autism [30,31]), our approach that allows to disambiguate specific behavioural components in a high-throughput assay may have important relevance for translational approaches. For example, it might aid efforts aimed at identifying more selective therapeutic compounds that share targets with known beneficial pharmaceuticals, but that act with greater molecular and behaviour-modifying precision.

Habituation of different response components may result from plasticity at different synapses within the same neurons, but a more parsimonious mechanism would involve different neurons that are part of parallel or serial pathways within the circuit. The negative correlations in habituation performance that we observed in some components also support a model where plasticity occurs in different neurons that are arranged in sequence within a sensory-motor path. This can be explained by a simple model, where individual variations in plasticity at upstream neurons result in variable levels of activation and thereby variable opportunity for habituation at downstream neurons. The physical location of plasticity sites remains to be determined. However, it is tempting to speculate that plasticity regulating the release of the dark flash response, such as “Probability” and “Latency”, might exist more towards the earlier sensory-related parts of the circuit, while regulation of kinematic parameters might occur downstream towards the motor circuitry in the hindbrain or spinal cord.

In light of recent work in C. elegans, where multiple genetically separable mechanisms have been shown to underlie short-term habituation [11–13], we propose that such modularity is a conserved feature of habituation. Thus, to accurately identify and characterize the possible neural implementation of habituation, it is important to consider a movement bout not as a single behavioural output, but rather as a combination of multiple independent modules. Additionally, while behavioural classification in larval zebrafish often considers the entire bout as the unit of behaviour [17], these results suggest that sub-components of bouts may represent an important unit of behaviour in this system. Similarly, in our analyses, we have treated all high-amplitude turns exhibited in response to a dark flash as a single “response type” or bout, which has multiple components that are modulated during habituation. Alternatively, it is possible that dark flashes elicit multiple distinct bout types with different neural circuit underpinnings (O-bends vs. SATs [17], the “Simple” and “Complex” responses we observe here), and that habituation acts to shift the proportions of bout types expressed. This question may be resolved when we have a better understanding of the circuit elements underlying response components or bout types.

Why might the zebrafish brain have evolved such a seemingly complex strategy to habituate? Perhaps plasticity to repeated stimulation is simply a pervasive adaptation at many synapses in a circuit, and we can observe these multiple effects when analyzing behaviour in a multi-component manner. Alternatively, approaching habituation in a modular way would facilitate behavioural flexibility. This would allow for specific adaptations rather than a simple global reduction in responses, perhaps tuned based on brain state, stimulus, or environmental context. Indeed, we found that habituation of “Probability” is modulated by the circadian rhythm and dark flash intensity, while other response components are not. This demonstrates that habituation acts in a modular fashion to tune the habituation rate of different components of behaviour based on context. Thus, our results reveal that the strategies taken by even relatively simple larval zebrafish brains to habituate require a surprisingly complex combination of independent plasticity events distributed across the circuit.

STAR Methods

CONTACT FOR REAGENT AND RESOURCE SHARING

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Michael Granato (granatom@pennmedicine.upenn.edu).

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Experiments were conducted on 5–6 dpf larval zebrafish (Danio rerio, TLF strain) raised in E3 medium at 29°C on a 14:10 hr light cycle. For each experiment, larvae from multiple clutches (5–20 mating pairs) were collected, and the clutch populations were mixed evenly between treatment and control groups during experiments. For the circadian experiments (Figure 4 J–L), larvae tested in the subjective day were raised on a 9am-ON, 11pm-OFF light cycle, while larvae tested in the subjective night were raised on an 9pm-ON, 11am-OFF light cycle, and both groups were tested beginning at ~12:30pm. Breeding adult zebrafish were maintained at 28°C. Behavioural assays on larvae ca rrying mutations for Nf1a^p301 (ZDB-ALT-130528–1) and Nf1b^p303 (ZDB-ALT-130528–3) were conducted blind to genotype. Subsequent genotyping was performed with the KASP method with proprietary primer sequences (LGC Genomics). This method was validated using previously described PCR genotyping[32]. All animal protocols were approved by the University of Pennsylvania Institutional Animal Care and Use Committee (IACUC).

METHOD DETAILS

Behaviour recording

Larval behaviour was recorded in multiwell plates fabricated from 6.3mm thick clear acrylic sheets (US Plastics). The acrylic was cut with a laser cutter into 8mm diameter wells with a volume of ~ 300 uL, arranged in a 20×15 grid for a total 300-well plate. 3.2mm white acrylic (US Plastics) was bonded to the cut wells (SciGrip 4), acting as both the bottom of the plate and the light diffuser. To minimize evaporation and to maintain a consistent ~28°C temperature in the beha viour wells, the 300-well plate was placed under a 29–31°C water bath that acted as a heated lid for the plate.

Larvae were illuminated from below with IR LEDs (890nm, Vsiahy.com part number TSHF5410) driven by a 1A Buckpuck driver (Luxdrive). Images were recorded from above with EoSens 4CXP Monochrome Camera (Mikroton), an 85mm 1.8 AF D lens (Nikon) with a IR long-pass filter (LP780–62, Midwest optics), and a Cyton Quad Channel CoaXPress Frame Grabber (Bitflow). The camera was triggered at 560hz using a Teensy 3.2 microcontroller (PJRC).

Due to the symmetrical design of the behavioural assay (1 hr training blocks, 1 hr rest between blocks), we were able to double the throughput of the rig by alternatingly imaging between two separate 300-well plates during the experiment. The larvae in plate 1 were recorded during training, and during the rest period the camera view was switched to plate 2, which was trained and recorded during the rest period for plate 1 (and vice versa). Therefore, the experimental time for the first and second plates are offset by one hour. Therefore, for the comparisons of trained and untrained larvae (Figure S1), the untrained larvae are always tested 1 hour after the trained larvae, and have been inside the rig for one hour longer. Switching the camera views was done by placing the camera at a 90 degree angle above the behaviour plates and using two 4” × 5” 45-degree incidence hot mirrors (43–958, Edmund Optics) to direct the camera view towards the two behaviour plates. The mirrors were attached to Nema 17 stepper motors (ROB-09238, Sparkfun), driven by an EasyDriver (ROB-12779, Sparkfun), a Teensy 3.2 microcontroller (PJRC), and the Multi-fish-tracker software. In this way they could be rotated in and out of place to view each plate. Light cross talk between the behaviour plates was minimized using blackout hardboard (TB4, Thorlabs).

For visual stimuli, we made a rectangular ring of 115 RGB LEDs (WS2812B 5050 RGB LED Strip 1M 144LED/M, ebay.com) to border the 300-well plate, diffused by 3.2mm white acrylic (US Plastics). LEDs were controlled using a Teensy 3.2 microcontroller and the fastLED Animation library (http://fastled.io/). The LEDs were set to white color with a brightness value of 50, yielding an intensity of approximately 130uW/cm² at the behaviour plate. During a dark flash, the LEDs were turned off for 1 second and video of larval responses was recorded at 560 Hz. After this, the light intensity increased linearly to the original brightness over 20 seconds. To induce the optomotor response, a translating stimulus was generated by illuminating every 8th LED along the top and bottom of the plate. The position of the illuminated LED was progressively shifted down the strip by ramping down the intensity of the illuminated LED, while ramping up the intensity of the adjacent LED. This results in a stimulus that is approximately sinusoidal in space, 5.5 cm peak to peak, translating at 5.5 cm per second. In this way, the motion stimuli were translated in the leftward and rightward directions relative to the plate, moving with a constant speed. The direction of motion was switched every 30 seconds, for a total testing period of 1 hour. The orientation of the zebrafish larvae was tracked online using the Multi-fish-tracker (see below) at 28 Hz and was used to quantify the optomotor response, which follows the direction of perceived motion. Acoustic tap stimuli were delivered using a Solenoid (ROB-10391, Sparkfun) that delivered a single tap to the top of the water bath and induced acoustic escape responses.

Multi-fish-tracker

The code to track individual zebrafish larvae in a multi-well format was custom written in C# (Microsoft, USA) using Intel’s integrated performance primitives (IPP, Intel, USA) for fast image processing. Specifically, a running average background was kept for each plate that was updated with an exponential decay time of 2 minutes. This was done to flexibly adapt to different lighting conditions. The plate was subsequently divided into two sections, which were tracked on separate threads to increase throughput. Individual wells were identified using user-defined masks. The background was subtracted from each image and the resulting absolute difference was thresholded. Subsequently, the biggest object in each well, physically close to a previously identified larval position, was designated as the larval object, and relevant parameters such as position and heading angle were extracted using image moments. At baseline, tracking was performed at 28 Hz. For one second after each dark flash (or tap), all frames at the full camera frame-rate of 560 Hz were written to disk, for detailed offline kinematic analysis of behaviour.

Offline video tracking

Offline tracking on recorded videos was performed in Matlab (Mathworks). The image was background subtracted and thresholded to identify the centroid of the larvae in each well. The background subtracted image was convolved with a disk filter with a 3 pixel radius, and the maximum intensity pixel was used to identify the head coordinate between the two pigmented eyes. To track the points along the body axis, we calculated a search direction vector defined by the head-to-centroid direction, and searched in an pi/3 sized arc placed at a radius of 5 pixels away from the head coordinate. The brightest point on this arc was considered the 2nd point along the fish. The search direction vector was then updated to the 1st to 2nd point direction, and a second arc was calculated 5 pixels from the 2nd point and the brightest pixel on this arc was assigned as the 3rd point. This process was iterated until 8 points were placed along the larvae. If no pixels above an intensity value of 4 were identified on an arc, tracking on this frame was stopped. The head coordinate was used to calculate displacement of the larvae, the head-to-centroid vector was used to calculate the heading orientation of the larvae, and the cumulative angle between the tail points was used to calculate the bend amplitude of the larvae.

Pharmacology

Stock solutions of 100uM Haloperidol (H1512, Sigma), Pimozide (P1793, Sigma) and Clozapine (C6305, Sigma) were prepared in DMSO. 10× solutions in 1% DMSO in E3 media were then prepared, and 30uL of these 10× solutions were directly pipetted into the wells containing the larvae, which have a total volume of ~300uL, yielding 10uM Haloperidol, 1uM Pimozide, or 10uM Clozapine, in 0.1% DMSO vehicle. 30uL of 1% DMSO in E3 solution was pipetted into the vehicle control wells, yielding 0.1% DMSO vehicle control treated larvae. Larvae were treated with drug for between 30 and 90 minutes before the first dark flash was delivered. Vehicle control and drug treated larvae for each comparison were from the same clutches of larvae, and were assayed in different wells in the same behavioural plate.

QUANTIFICATION AND STATISTICAL ANALYSIS

Behavioural quantification

Analyses of larval behaviour and statistical analyses were performed in Matlab (Mathworks). For each dark flash or tap stimulus, the offline tracked videos were used to score behaviour during the 1 second of recorded video. Dark flash responses were identified as movement events that had a bend amplitude greater than 3 radians (172 degrees). Responses to taps were identified with a bend amplitude greater than 1 radian. Compound responses (Figure S2C) were classified as dark flash responses which had at least two local maxima in the bend amplitude trace during the initial bend before the bend amplitude trace crossed 0.

The proportion of the larval population that performed a response at each dark flash was used to quantify habituation performance for “Probability” (Figure 1C, 2A). To generate habituation curves for the other behavioural components (Figure 2B–I, S2A), larvae that did not perform an response for a given stimulus were excluded from the analysis at that stimulus. To fit exponential curves to each training block of 60 flashes (x), we used Matlab’s ‘fit’ function, with a ‘fittype’ formula of:

$$y=a\times 2^{-b\times x}+c$$ (1)

To quantify the recovery of the response in the population (Figure 2J), we averaged the response at each flash in Block 4 and the Re-test block across all larvae, and divided the Re-test block responses by the responses in Block 4. These distributions were plotted with ‘violin.m’[33], with a bandwidth of 0.15. The comparison in dark flash responsiveness in trained and untrained larvae (Figure S1A) was performed by averaging the response for each population at each of the 60 dark flash stimuli, dividing the trained mean vector by the untrained mean vector, and plotting a histogram of the result. Values below 1 indicate suppressed responsiveness in the trained larvae.

Percent habituation was calculated for each larva as the decreased mean responsiveness for the 60 flashes in training Block 4, relative to the mean response for the 60 flashes training Block 1, using formula (2). To make the distributions comparable across the different components of behaviour, the minimum observed mean value across all larvae was subtracted from the block 1 and block 4 mean responses. This ensures that the responses can scale all the way to 0 regardless of behaviour component -for example “Curvature”, which by our definition of dark flash responses, must be at minimum 3 radians. Except when calculating for the “Probability” behaviour component, larvae that did not respond to a given stimulus were excluded from the analysis at that stimulus.

$$\%Habituation=100\times \left(1-\frac{Block4}{0.5\times (Block4+Block1)}\right)$$ (2)

For the correlational analyses of habituation performance for different dark flash response components across fish (Figure 3B–E), the percent habituation scores for each component were assembled into a vector across the 3120 fish, and the spearman correlation coefficient was calculated using Matlab’s ‘corr’ function. The same was done for the naive response (Figure S2D), using the response to the first flash rather than percent habituation. This analysis can not be done comparing to the “Probability” component, since the other components only manifest if a response actually occurs, as is reflected in the missing data in the matrix.

Analysis of stimulus-free swimming behaviour (Figure S1E–G) was done using the online tracked larval coordinates from the Multi-fish-tracker. This was done for a 30 min period beginning one hour after the fourth dark flash habituation training block. Analysis of the optomotor response (OMR, Figure S1A–D), was done using the heading angle from the Multi-fish-tracker for one hour, beginning 3 hours after the fourth dark flash habituation training block. To calculate OMR performance, we isolated each 30-second period where the larvae were either being stimulated with leftward or rightward motion. We isolated the left-right component of the orientation by calculating the arcsin of the sin of the larval orientation. We then reflected the traces in time during the rightward stimulus presentation, such that each 30-second period would have an increasing slope if the larva were to reorient to follow the direction of motion (Figure S1D). All the 30 second periods were then averaged for each larva, and these averaged traces were fit with linear regression using Matlab’s ‘polyfit’ function. The slope of this fit (OMR slope) was taken as the measure for OMR performance for each larva (Figure S1E).

To calculate the cumulative difference in habituation performance for the Nf1 mutants, pharmacological treatments, 80% flash, and the circadian experiments (Figure 4), we calculated the average response across larvae at each dark flash. This was done for the treatment and control groups, yielding a mean vector for each group. These two vectors were normalized by dividing them by the initial response to the first flash for each group, and they were then subtracted, yielding a mean difference vector between stimulus and controls flash. To generate statistical confidence in these vectors, we used bootstrapping of 2000 replicates, and calculated the 99.5% confidence intervals using Matlab’s ‘fitdist’ and ‘paramci’ functions. If the two groups are habituating similarly, then these difference vectors will have a mean of approximately 0, and thus the cumulative mean distribution would remain near 0. However, treatments that affect habituation will show strong increasing or decreasing cumulative mean distributions, reflecting increased or decreased habituation performance throughout training, respectively. We confirmed this by comparing larvae in even and odd numbered wells, which showed little divergence from 0 in the cumulative distributions (Figure 4I), and based on this dataset, we set a magnitude threshold of +/− 0.05 reflecting the expected variability in this analysis.

Modeling

We modeled two habituation loci acting in series in Matlab. We began with the upstream locus, which follows learning kinetics of the exponential fit for the first training block of 60 flashes when measuring “Probability” (Figure 2A, Block 1, Equation (1)). In each model run, gaussian noise was added to the coefficients of the fit, resulting in variable habituation curves. Each learning curve was normalized such that the maximum value is equal to 1. The downstream locus follows the same learning kinetics, but the learning curve was truncated based on how much habituation occurred at the upstream locus. This was measured as the area under the habituation curve at the upstream node. If little habituation occurs, this will be close to 60, for the full 60 flashes. However, if habituation is profound this will approach 1. Therefore, the opportunity for habituation at the downstream node is negatively dependent on habituation performance at the upstream node. The variability in habituation across runs (axes in Figure 3G) was controlled by varying the standard deviation of the gaussian distribution from which the noise added to the coefficients was derived. Learning performance was defined as in Equation (2), replacing Block 1 with the initial value of the habituation curve, and Block 4 with the final value of the curve. For each value of habituation variability (‘sigma’) at each locus, 10000 iterations of the model were run. The correlation in learning performance at each node across model runs (Spearman’s rho) was calculated using Matlab’s ‘corr’ function.

Supplementary Material

1. Video S1. The dark flash response. Related to Figure 1.

The same larva performing a response to three different dark flashes. The responses are characterized by an initial high-amplitude “O”-shaped bend, followed by a swim in the new heading direction. The videos are synchronized to the beginning of the movement, and last for 197 msec.

2. Video S2. Habituation of the dark flash response. Related to Figure 1.

Videos of the same 300 larvae comparing the response to the first dark flash (Naive Response, left), and the 240^th flash (Habituated Response, right). Larvae are recorded at 560hz in 300-well plates. The images are background subtracted so that the wells are not visible and the larvae appear white. The cumulative path taken by the larvae becomes visible as orange streaks in the videos to highlight the responses. Related to Figure 1

3. Video S3. Compound vs Simple Responses, and double responses to a single dark flash. Related to Figure 2.

Recordings of two larvae, cropped from videos of the 300-well behaviour plates. Time is displayed relative to light-offset, which remains off for the entire video. The larva in the bottom well performs a Simple Response, characterized by a single high-amplitude bend, followed by a swim forward in the new heading direction. The larva in the top well performs a Compound Response, characterized by two separate high-amplitude bends in the same direction, followed by a swim forward. After a period of immobility, this larva then performs a second response, this time a Simple Response. The larva on the bottom also moves a second time, but performs a “swim” without any characteristic high-amplitude bend.

Highlights:

• High-throughput behavioural analysis of dark flash habituation in larval zebrafish

• Multiple components of the response adapt with different habituation kinetics

• Controlled by multiple circuit loci with different molecular requirements

• Modular habituation selectively tunes behavioral components based on context