ABSTRACT
Zebrafish lateral line is an established model for hair cell organ damage, yet few studies link mechanistic disruptions to changes in biologically relevant behavior. We used larval zebrafish to determine how damage via ototoxic chemicals impact rheotaxis. Larvae were treated with CuSO4 or neomycin to disrupt lateral line function then exposed to water flow stimuli. Their swimming behavior was recorded, and DeepLabCut and SimBA software were used to track movements and classify rheotaxis behavior. Lateral line-disrupted fish performed rheotaxis, but they swam greater distances, for shorter durations, and with greater angular variance than controls. Further, spectral decomposition analyses demonstrated that lesioned fish exhibited toxin-specific behavioral profiles with distinct fluctuations in the magnitude, timing, and cross-correlation between changes in linear and angular movements. Our observations support that lateral-line input is needed for fish to perform rheotaxis efficiently in flow and reveals commonly used lesion methods have unique effects on behavior.
INTRODUCTION
The lateral line is a sensory system used by fishes and amphibians to detect water flow. The functional units of the lateral line are neuromasts; bundles of sensory hair cells located externally along the head and body that mechanotransduce low frequency (≤ 200 Hz) water flow stimuli into electrochemical signals for interpretation by the central nervous system (reviewed in 1). The lateral line is known to partially mediate rheotaxis, a multimodal behavior (2) that integrates input from visual (3, 4, 5 vestibular (6, 7), tactile (3, 8–10), and lateral line systems (8–10) to facilitate fish orientation and movement with respect to water flow (9, 10).
Although contribution from the lateral line is well established, there is conflicting evidence on whether it is essential for rheotaxis in fish (2 , 5, 9–12). Inconsistent methodologies used on several species that differentially rely on the lateral line to mediate behavior obfuscate the relationship between the lateral line and rheotaxis. However, a recent review hypothesized that differences in the spatial characteristics and velocity of the flow stimuli used to assay rheotaxis likely resulted in the disparate results reported in these studies (see Table 1 in 13).
Zebrafish hair cells demonstrate unique regenerative features that make them an important model for hearing loss research (14–19). Many studies have focused on mechanistic disruption of hair cell activity, but few have explored the association between this disruption and behavior. Previous researchers have developed several different assays to study rheotaxis in larval fishes (e.g. 5, 20–22). However, we sought to develop a behavioral assay and analytical methodology that is sensitive, spatiotemporally scalable, and robust enough to be used on a variety of species, ontogenetic stages, sensory modalities, and behavioral responses that are pertinent to biomedical and ecological research.
We therefore investigated how the lateral line contributes to rheotaxis in larval zebrafish and developed a standardized assay that could identify previously observed subtle differences in rheotaxis behavior (12). We compared the rheotaxis response of fish with an intact lateral line to those with lateral lines ablated by two commonly used compounds: copper sulfate (CuSO4; 17, 23–25) and neomycin (5, 15, 26). We hypothesized that different toxins produce distinct changes in rheotaxis behavior because they injure lateral lines via different cellular mechanisms (27), and that these changes may be quantified objectively with machine vision and learning techniques.
RESULTS & DISCUSSION
To determine the contribution of the lateral line to rheotaxis in fishes we used CuSO4 and neomycin to ablate the lateral line neuromasts of larval zebrafish (6-7 days post-fertilization (dpf)), then video recorded the swimming behavior of individual larvae in a microflume under no-flow and flow stimulus conditions (Fig. 1). Although it was not possible to selectively block tactile cues, young larval zebrafish are not known to detect horizontal angular velocity cues (i.e. yaw) and our procedures eliminated linear acceleration and visual cues (see methods). We used machine vision and learning software to annotate videos for positive rheotaxis events where fish were oriented to 0° ± 45° and actively swam into the oncoming flow (Fig. 2). Under no flow conditions, we determined that each treatment group of fish was randomly distributed (Fig. 3; Supplementary Table 1) and had no natural proclivity to orient their bodies to 0° ± 45° (Supplementary Table S2). We standardized our analyses by comparing rheotaxis data acquired during flow to the swimming behavior of fish when they were randomly oriented at 0° ± 45° under no flow. A subset of all fish that had undergone behavioral testing were fixed and immunolabeled to confirm lateral line organ ablation (Supplementary Fig. 1).
Lateral Line Ablation Alters but Does Not Eliminate Rheotaxis Behavior
We predicted that fish treated with the minimum dosage of CuSO4 or neomycin necessary to completely ablate lateral line hair cells of the lateral line would not perform rheotaxis as well as control fish with an intact lateral line. Surprisingly, fish lacking intact lateral line organs could still orient into oncoming flow similar to control fish (Fig. 3; Supplementary Tables 1-2), indicating that the effects of lateral line ablation were subtle and not essential for rheotaxis behavior.
When analyzing individual components of rheotaxis, stark behavioral changes were observed among lesioned fish. Comparisons between lateral line-ablated and control groups showed significant differences in the mean body angle or angular variance (Fig. 3, Supplementary Table 3), mean duration of rheotaxis events, mean number of rheotaxis events, total distance traveled, and latency to the onset of rheotaxis (Fig. 4, Supplementary Tables 4-7). These observations suggest that rheotaxis occurs but is altered in lateral line-ablated groups, contrasting with previous reports on larval zebrafish (5, 22, 28) and supporting the idea that the lateral line is not required for rheotaxis in fish (2, 11, 12, 29, 30). We believe that our focus on acute rheotaxis behavior and use of fine scale spatiotemporal sampling and machine vision allowed us to detect subtle changes in rheotaxis behavior that were overlooked in previous studies (2, 5, 12, 29, 30). Our methods eliminated water motion, optic flow (3), and certain vestibular cues (yaw: 31, 32; linear acceleration: see methods). However, tactile cues were not eliminated because we could not prevent fish from contacting the substrate and our flow rate was sufficient to displace substrate coupled fish backward and against the rear mesh of arena. Therefore, we propose that lateral line ablated fish predominantly used tactile cues to gain an external frame of reference and perform rheotaxis (8, 10).
Toxins Differentially Influence the Distribution of Mean Body Angles During Flow
To identify differences in body orientation, the mean body angle for fish in the presence or absence of flow stimuli were compared among treatment groups (Fig. 3). Since flow originated at the top of the flume (located at 0°), mean body angle was determined relative to the oncoming water stimulus. Angular variance of each group is represented by the inverse of the mean resultant vector length (i.e. short vectors = high variance, and vice versa; Fig. 3). Under no flow, fish from each treatment group swam with a random orientation, indicated by a grand mean body angle that was statistically different from 0° and resultant vector length close to zero (Fig. 3a, Supplementary Table 2). When flow was applied, all groups exhibited a grand mean body angle clustered around 0° and resultant length close to one (Fig. 3b-c), thus demonstrating significant alignment into the oncoming flow stimulus.
During rheotaxis, post-hoc comparisons within treatment groups showed a significant difference in the homogeneity of distributions between initial and final portions of the flow stimulus, indicating that overall orientation behavior within groups was not consistent for the duration of flow presentation (Fig. 3b-c, Supplementary Table 3). Comparisons among groups within the initial 10s portion of flow showed a significant difference in the distribution of mean body angles between control and CuSO4 fish, and between control and neomycin fish (Fig. 3b, Supplementary Table 3). There was also an interaction between treatment and stimulus where the distribution of mean body angle in CuSO4-ablated fish during the initial stimulus bin was different than that of neomycin fish during the final sequence (Supplementary Table S3), suggesting that these lesion methods differentially impacted the rheotaxis behavior of fish.
Lateral Line-Ablated Fish Performed Rheotaxis for Shorter Durations but Travelled Longer Distances
Although intact lateral line was not required to perform rheotaxis, lesioned fish behaved differently than non-lesioned fish in flow stimulus, indicating that lack of input from neuromast hair cells affected swimming behavior. To quantify differences in rheotaxis behavior, we accounted for the random effects of individual variation then compared the mean duration and mean number of rheotaxis events, total distance travelled, and latency between flow presentation and behavior onset among groups. We standardized the data by comparing events of rheotaxis to events when fish were randomly oriented at 0° ± 45° under no flow and hereafter refer to both conditions as “events”.
Without flow, the mean duration of random orientation events did not differ among treatment groups (Fig. 4a, Supplementary Table 4). With flow, the mean duration of rheotaxis events increased for each group, and was greatest in the controls, less in neomycin-, and least in CuSO4-treated fish (Fig. 4a, Supplementary Table 4). An interaction between stimulus and treatment was seen during the initial 10s of flow only in the CuSO4 treatment cohort, but during the final 10s the differences among all treatments became significant (Supplementary Table 4). Mean number of events had a significant effect of stimulus and an effect of treatment where the number of events was greatest in neomycin-treated, less in controls, and least in CuSO4-treated under no flow and flow conditions (Fig. 4b, Supplementary Table 5). The total distance travelled during events was influenced by presence of flow stimulus and type of lesion, where neomycin-treated fish travelled further than control and CuSO4-treated fish in the presence or absence of flow conditions (Fig. 4c, Supplementary Table 6). CuSO4-lesioned fish also demonstrated a greater delay in initiating rheotaxis compared to control and neomycin fish (Fig. 4d, Supplementary Table S7). Altogether, our data show that an intact lateral line allowed fish to sustain rheotaxis into oncoming flow for longer durations yet travel shorter distances, suggesting that toxins reduce zebrafish economy of motion against flow stimulus. These results support the idea that the lateral line allows epibenthic fish to hold their station with respect to the substrate (2, 33)
Interestingly, CuSO4 and neomycin affected swimming behavior against flow stimulus in different ways. CuSO4 exposure decreased activity, as evidenced by fewer rheotaxis events with greater latency between flow and behavior initiation. This contrasts with the neomycin-exposed fish, that exhibited frequent bursts of rheotaxis, and longer distances travelled. We propose that the distinct effects of CuSO4 and neomycin treatments on the spatiotemporal nature of rheotaxis may be due, in part, to their different mechanisms of neuromast ablation. Hair cells were completely ablated and supporting cells and afferent neurons damaged at the dosage used for CuSO4 (17), but on occasion a few hair cells remained in neomycin-treated fish (Supplementary Fig. 1). If residual hair cells retained some functionality in the neomycin group despite severe morphological damage, then it is possible that their sensitivity might have been amplified through efferent modulation (34) or previously silent hair cells were recruited to compensate for reduced sensory input (35). In addition, while zebrafish larvae have been shown to be relatively resistant to the concentration and exposure time of CuSO4 used (36), subtle differences in behavior may also be a consequence of nonspecific neural toxicity.
Overall Spatial Use of the Arena During Rheotaxis Differs Among Treatments
During flow, we observed that lateral line-lesioned fish were often pushed against the back of the testing arena, whereas intact fish were often swimming at the front of the arena near the flow source. To quantify this observation, the density of fish throughout the duration of the experiment was plotted in one- and two-dimensional (1D, 2D) space and then compared among treatments.
Along the X-axis, the spatial use was similar among treatment groups and flow conditions (Fig. 5a). All groups preferred the sides over the middle of the arena. Without flow, fish that were placed into the arena remained close to the walls as they explored the boundaries. With flow, preference for the sides became more pronounced due to gradients in the laminar flow field where the highest velocities occurred at the surface and left side and the lowest velocities at the bottom and right side (Supplementary Fig. 2). These boundary layers provided a refuge (e.g. 37) for intact and lesioned fish alike to reduce their energetic costs (38).
By contrast, the spatial use of fish along the Y-axis differed significantly among treatment groups and became more prominent during flow conditions (Fig. 5b). Under flow, the control group predominantly used the front region (Fig. 5b), demonstrating that an intact lateral line allowed these fish to maintain their position at the front of the arena where the flow was strongest. Conversely, the predominant location of toxin-treated groups was shifted toward the back region, suggesting an impaired ability to station hold and increased reliance on tactile cues provided by contact with the rear grate. CuSO4-lesioned fish exhibited a greater shift towards the back of the arena than neomycin-lesioned fish (Fig. 5b), but both groups demonstrated this abnormal distribution along the Y-axis compared to control fish.
The spatial use of the arena in 2D further illustrates differences among treatment groups under flow. During rheotaxis, the total density of fish in 2D space was parsed out over five regions of interest (ROI, Fig. 6a) set according to size of fish, their orientation during rheotaxis, and the dimensions of the flow field. Generalized linear models (Supplementary Table 8) indicate that the spatial use of fish during rheotaxis differed among all three treatment groups. (Fig. 6d). During flow, control fish frequently maintained position in the upper-left corner of the arena whereas lesioned fish primarily occupied the back-right corner (Fig. 6d), suggesting that a functional lateral line enables fish to occupy the areas of strongest flow. The propensity of lesioned fish to be swept backwards yet still perform rheotaxis behavior without visual cues likely demonstrates that they relied on tactile cues to provide the external frame of reference necessary to orient and swim against flow (2, 8, 10). Cumulatively, these observations support that the sensory information provided by lateral line organs, in conjunction with the tactile system, allowed intact fish to actively maintain their position within a non-uniform laminar flow field where the velocity of water is greatest (2).
Lateral Line Ablation Reduces the Proportion of Individual Fish that Perform Rheotaxis
During flow presentation, the control group had the greatest proportion of individual fish that performed rheotaxis during the experiment (as defined in Fig. 2) followed by neomycin-treated then CuSO4-treated groups (Fig. 7). Intact fish plateaued relatively quickly compared to lesioned fish, but the values for all three groups converged during the final few seconds of flow presentation.
Spectral Decomposition Shows that Lateral Line Ablation Impacts the Trend and Degree of Periodic Fluctuation in Linear and Angular Movement
To uncover the impact that lateral line ablation had on the swimming kinematics of fish, we analyzed how the magnitude of the linear (relative distanced moved, relative velocity, relative acceleration) and angular movements (mean body angle, mean length of the resultant vector) changed over time. We observed that all treatment groups swam in the burst-and-glide style that is characteristic of larval zebrafish with intact lateral lines, but with noticeable differences in the quality of their movement. Because the lateral line is thought to mediate station holding behavior (2, 33, 38), we postulated that, under flow, the magnitude of the oscillations in relative linear and angular movements of lateral line-intact fish would be smaller than those of lesioned fish. The observed time series data (Fig. 8a, d, g; Supplementary Fig. 3 a, d) had a “seismic” appearance where noise masked the underlying signal. Therefore, we removed the random noise (Supplementary Fig. 3g-k) and decomposed the observed datasets into their fundamental components: the large scale trends in the magnitude of movements during the entire experiment (Fig. 8b, e, h; Supplementary Fig. 3 b, e) and the small scale periodicity, which indicates the recurring fluctuations in movement magnitude that occurred during any given second of the experiment (Fig. 8c, f, i). Only relative movement data are shown in Fig. 8 to provide visual clarity because the relative velocity and acceleration periodicity data showed similar fluctuations (Supplementary Fig. 3 c, f).
During rheotaxis, the trend among groups was that CuSO4 ablation reduced the magnitude of relative movement (Fig. 8b) but not the relative velocity or acceleration of fish (Supplementary Fig. 3 c, f) compared to control and neomycin groups. For the trends in angular movements, CuSO4- and neomycin-treated fish showed a reduced ability to achieve and maintain orientation with flow compared to controls (Fig. 8e). Lesioned fish had longer mean resultant vector lengths and greater fidelity to a mean body angle compared to controls (Fig. 8h) which swam with more angular variance (1- rho) than lesioned fish. Consequently, an intact lateral line allowed control fish to detect minute changes in water flow and make regular course corrections of small magnitude, thereby increasing their angular variance (Fig. 8h) and enabling them to rapidly orient into flow with greater accuracy (Fig. 8e) and led to a greater proportion of non-lesioned fish performing rheotaxis compared to lesioned fish (Fig. 7).
Among treatment groups, there were clear differences in the periodicity of the changes in the magnitude of linear (Fig. 8c; Supplementary Fig. 3 c, f) and angular (Fig. 8f, i) movements. The linear periodicity data showed that CuSO4-treated fish had the greatest fluctuation in the magnitude of their relative distance moved, velocity, and acceleration compared to the control and neomycin-treated fish (Fig. 8c; Supplementary Fig. 3 c, f; Supplementary Table 9). Conversely, the neomycin-treated fish had the lowest fluctuations in the magnitude of their linear movements. The large magnitude of the periodic oscillations in CuSO4-treated fish could compensate for their delayed response to flow (Fig. 4d), reduced trend in relative movement (Fig. 8b), and explain the relative increase observed in total distance traveled during flow (Fig. 4c). Furthermore, CuSO4-treated fish had large amplitude peaks that occurred early during the random sampling periods, whereas neomycin-treated and control fish had smaller peaks that increased gradually (Fig. 8c; Supplementary Fig. 3 c, f). These empirical data support the qualitative observations that CuSO4 ablation resulted in fish that performed rheotaxis with delayed responses and erratic movements. Additionally, the maxima in control fish occurred at the end of the sampling period (Fig. 8c), indicating an intact lateral line allowed fish to exhibit a graded and more controlled response to flow compared to lesioned fish.
The range in amplitude of the angular periodicity data showed a different pattern where neomycin-treated fish had the greatest fluctuation in mean body angle, followed by the controls, and then the CuSO4-treated fish (Fig. 8f; Supplementary Table 9). Notably, all three treatment groups had a small dip and correction during the initial portion of the sampling period (Fig. 8f), which indicates that all fish had the same initial small change in mean body angle to the left of the flow. Control and CuSO4-treated fish had a subsequent series of minor heading changes to the right (positive values) and left (negative values) of the flow, whereas the neomycin-treated fish regularly overshot their initial and secondary responses before recovering. These large oscillations might explain why neomycin fish had difficulty locking onto the source of the water flow (Fig. 8e). However, the smaller fluctuations in mean body angle observed in control and CuSO4-treated fish (Fig. 8f) is not sufficient to explain why control fish could rapidly and consistently orient toward the flow better than CuSO4-treated fish (Fig. 8e). Perhaps the greater variety of temporal fluctuation combined with the small amplitude fluctuations allowed intact fish to more rapidly attune their mean body angle to the flow compared to CuSO4 fish (Fig. 9d; see Spectral Analysis). In both neomycin- and CuSO4- lesioned fish, the waveform structure for the mean resultant length per second shows a rapid succession of several small peaks, but that of control fish has two small peaks followed by two large peaks (Fig. 8i). These small initial changes in angular variance of intact fish might reflect greater capacity to sample flow stimuli from a variety of small angles to produce large correctional changes in angular variance (Fig. 8i) with better efficacy (Fig. 8e).
Spectral Analysis Reveals that Lateral Line Ablation Impacts the Temporal Fluctuation of Linear and Angular Movement
The fluctuations in periodicity data have a predictable and repeatable temporal structure that can be described by a series of fundamental sine and cosine functions that oscillate at different frequencies (Hz; cycles s-1) and periods (s cycle-1). Therefore, we performed a spectral analysis to decompose the periodicity data (Fig. 8c, f, i; Supplementary Fig. 3 c, f) and determine the fundamental frequencies at which the fluctuations in linear and angular movement occurred (Fig. 9a-e). For the power spectra of each parameter, the frequency and amplitude of three most dominant peaks were summed to calculate the net shifts in frequency and power among treatment groups (Supplementary Table 10).
Relative to controls, the overall trend in lesioned fish was a net downshift in the dominant frequencies of relative movement, velocity, and acceleration, a net upshift in the mean body angle, and a net increase in amplitude across these four parameters (Fig. 9a-d, Supplementary Table 10). This indicates that lateral line-ablated fish swam with less frequent changes in their linear movements, more frequent changes in their mean body angle, and greater numbers of movements at these frequencies relative to controls. The curves for mean resultant vector length showed a notable difference between the two ablation treatments where CuSO4-treated fish had a net frequency downshift and a net amplitude increase relative to controls (Fig. 9e, Supplementary Table 10), and vice versa for neomycin-treated fish (Fig. 9e, Supplementary Table 10). Therefore, CuSO4-lesioned fish swam with more low frequency changes in angular variance, while neomycin-lesioned fish swam with fewer high frequency changes in angular variance.
Compared to neomycin-treated fish, the power spectra of the linear movements in control and CuSO4-treated fish (Fig. 9a-c) show broad clusters of dominant peaks that gradually upshift with each subsequent time derivative. CuSO4-treated fish swam with the greatest number of low-frequency fluctuations in relative movement, but the fluctuations in their velocity and acceleration had greater temporal variety and occurred with increasing frequency in a manner similar to those of control fish. Interestingly, the curves for linear movement in neomycin-treated fish are dominated by a single peak at the primary frequency of 0.21 Hz that remained constant with each time derivative (Fig. 9a-c). Therefore, neomycin ablation reduced the temporal variation of the cyclic fluctuations in relative movement, velocity, and acceleration of fish during rheotaxis. This clustering of rheotaxis behaviors into fewer peak responses with discrete timing, or temporal contraction, was further seen in the angular power spectra of lesioned fish. For example, the power spectrum of the mean body angle in control fish is relatively flat with small peaks, whereas those of ablated fish are focused into distinct dominant frequencies with larger amplitudes (Fig. 9d; also see Fig. 8f). Likewise, the spectrum of the mean resultant length in controls shows five broadly spaced peaks, but those of the lesioned fish are clustered into fewer peaks focused at the dominant frequencies (Fig. 9e). Therefore, intact fish swam with fewer overall changes to their mean body angle, but their angular variance had greater temporal variety (Fig. 9e) and magnitude (Fig. 8h, i) compared to lesioned fish.
We interpret the relatively flat and generally lower peaks observed in the power spectra of control fish (Fig. 9a-d) as indicating that sensory data from the lateral line allowed fish to make fewer changes in linear and angular movement over a broader range of frequencies, thus increasing the efficiency and efficacy of their response to flow. Conversely, fewer peaks of greater amplitude observed in lesioned fish indicate that loss of mechanosensory input from the lateral line resulted in greater numbers of swimming adjustments that were temporally restricted, reducing their economy of movement within flow.
Cross Correlation Between Linear and Angular Movements Shows Distinct Rheotaxis Profiles Among Groups
Changes in linear movement may be temporally cross-correlated with changes in angular movement, depicting rapid turns in response to flow. As these movements receive input from the lateral line, we investigated the effect of toxins on cross correlation between linear and angular movements. The correlograms depict how an above average increase in relative movement (Fig. 10a-b), velocity (Fig. 10c-d), or acceleration (Fig. 10e-f) cross-correlates with above average changes in the mean body angle (Fig. 10a, c, e) or mean resultant length (Fig. 10b, d, f).
Focusing on the strongest cross-correlations (Fig. 10; Supplementary Table 11) shows three distinct patterns in the direction and relative timing of changes between mean body angle and linear movements (Fig. 10a, c, e) and between temporal changes in angular variance and linear movements (Fig. 10b, d, f). In controls, the changes in body angle direction and the timing between angular and linear movements were temporally cross correlated, producing smooth re-orienting movements characteristic of effective station holding during rheotaxis (Fig. 10a, c, e). In addition, control fish tended to reduce their angular variance prior to a large change in movement (Fig. 10b), velocity (Fig. 10d), or acceleration (Fig. 10f). Therefore, the relative timing and direction of changes in mean body angle in lateral line-intact fish depended upon the time derivative of the linear movement, but the relative timing between reduced angular variance was consistent across changes in all linear movements.
In lesioned fish, the timing and direction of the strongest cross-correlations shifted the rheotaxis profile of these fish in a predictable, consistent, and treatment-specific manner. CuSO4 fish tended to change their mean body angle to the right of the oncoming flow vector after large changes in movement (Fig. 10a), velocity (Fig. 10c), and acceleration (Fig. 10e). However, neomycin fish tended to change their body angle to the left of the flow after changes in movement (Fig. 10a), velocity (Fig. 10c), and acceleration (Fig. 10e). As in controls, lesioned fish always reduced their angular variance when changing their linear movements, but the relative timing of the cross-correlation shifted according to the toxin used and the time derivative of the linear movement. In CuSO4-treated fish, the strongest correlation between a reduction in angular variance tended to occur prior to changes in movement (Fig. 10b), after changes in velocity (Fig. 10d), and after changes in acceleration (Fig. 10f). Conversely, in neomycin-treated fish, reductions in angular variance occurred after changes in movement (Fig. 10b), before changes in velocity (Fig. 10d), and before changes in acceleration (Fig. 10f).
These data revealed that control fish with intact lateral line detected oncoming flow and adjusted their mean body angle prior to making a change in their relative movement, while lateral-line lesioned fish changed their mean body angle after a change in their relative movement. We posit that an intact lateral line allowed fish to rapidly detect flow and adjust their heading prior to swimming into the flow while, in the absence of visual and lateral line cues, lesioned fish had to rely on tactile cues, such as physical displacement along the substrate, to gain an external frame of reference necessary to orient and swim into flow. Intact fish also tended to change their body angle to the left of the flow vector, which might reflect lateral line handedness where larval zebrafish prefer to use the right side of their lateral line to detect flow stimuli in a manner similar to that of blind cavefish (39). This undoubtedly shifted the relative timing of the cross-correlations between angular and linear movement, perhaps impaired any potential lateral line handedness, and further contributed to the erratic rheotaxis behavior observed in these fish.
In summary, lateral line ablation of larval zebrafish resulted in distinct, treatment-specific rheotaxis profiles that differed from that of intact fish in the following ways: 1) delayed relative timing between changes in mean body angle and all linear movements, 2) changed mean body angle in response to flow, and 3) shifted timing between reductions in angular variance and changes in linear movement.
CONCLUSIONS
In this study, ablating the lateral line of larval zebrafish with two commonly used toxins impacted their ability to produce the fine adjustments required to station hold in response to water flow. Lateral line ablation inhibited the ability of fish to discriminate subtle distinctions in flow, resulting in more intense overcorrections and decreased economy of motion. Our data support the hypotheses that the lateral line mediates station holding behavior (2, 33, 38), but is not required for rheotaxis behavior (2, 11, 12, 29, 30) in larval zebrafish in non-uniform laminar flow. We posit that the physical displacement of lesioned fish along the substrate provided sufficient tactile cues (8, 10) necessary for lateral line ablated fish to perform rheotaxis.
We propose that the greater angular variance observed in intact fish might indicate that these fish were regularly sampling the velocity gradients of the flow stimuli (22) from a variety of body angles so that they could reduce their response latency, quickly orient with respect to fluctuating flow stimuli, and maintain their overall mean body angle with greater fidelity to the flow vector than lesioned fish. During flow, intact fish had recurring fluctuations in relative movement, velocity, and mean body of lower magnitude, and fluctuations in relative movement, velocity, acceleration, and mean body angle of controls that were fewer in number yet occurred over a wider range of temporal frequencies compared to lesioned fish. Thus, the sensory cues detected by the lateral line allowed control fish to respond to water flow with less intensity and greater temporal variation in their movements, resulting in greater economy of movement.
This is the first study to demonstrate that two toxins commonly used to ablate the lateral line impacts the behavioral profiles and mechanism of rheotaxis in zebrafish. We propose a novel functional assay for hair cell ototoxicity, which may be used to supplement future studies exploring lateral line injury, protection, and recovery. Furthermore, the simplicity of the equipment and precision of the machine learning analyses used in this assay make it amenable to adaptation for detecting subtle behavioral changes in a wide variety of animal models.
METHODS
Ethics Statement
This study was performed with the approval of the Institutional Animal Care and Use Committee of Washington University School of Medicine in St. Louis and in accordance with NIH guidelines for use of zebrafish.
Zebrafish
Adult zebrafish were raised under standard conditions at 27-29°C in the Washington University Zebrafish Facility. The wild type line AB* was used for all experiments unless otherwise stated. Embryos were raised in incubators at 28°C in E3 media (5 mM NaCl, 0.17 mM KCl, 0.33 mM CaCl2, 0.33 mM MgCl2; (40)) with a 14:10 hr light:dark cycle. After 4 days post fertilization (dpf), larvae were raised in 100-200 mL E3 media in 250-mL plastic beakers and fed rotifers daily. Sex of the animal was not considered in our study because it cannot be determined in larval zebrafish.
Lateral Line Ablation
At 6 or 7dpf, ∼15 larval zebrafish were placed into each well of a flat bottom 6-well polystyrene plate (#351146, Falcon) in 8 mL of E3 media. Treatment animals were placed into 50 µM neomycin trisulfate salt hydrate (N1876-25G, Sigma Aldrich) or 10 µM CuSO4 (451657-10G, Sigma Aldrich) solutions made in 8 mL of E3 media. The plate was placed into an incubator at 29°C and exposed to the treatment for 30 min (neomycin) or 60 min (CuSO4). After exposure, the fish were removed from treatment, rinsed 3X in media, then placed into 8 mL of clean media and allowed to recover for 120 min (CuSO4) or 150 min (neomycin) in the incubator. Recovery times for CuSO4 and neomycin were chosen to standardize the total time of treatment and recovery to 180 min. Control fish (n = 248) received no chemical treatments yet underwent the same procedures as the CuSO4 (n = 204) and neomycin (n = 222) fish. At the end of treatment, the larvae were removed from the incubator and immediately began rheotaxis behavior trials.
Immunohistochemistry
Lateral line ablation was confirmed via immunohistochemistry on a subset of control and lesioned fish. Larvae were sedated on ice (5 min, 0°C) then fixed overnight at 4°C in PO4 buffer with 4% paraformaldehyde, 4% sucrose, and 0.2 mM CaCl2. Larvae were rinsed 3X with PBS and blocked for 2 hr at room temperature in PBS buffer with 5% horse serum, 1% Triton-X, and 1% DMSO. Primary antibodies for Otoferlin (HCS-1, Developmental Studies Hybridoma Bank, mouse IgG2a, 1:500) were diluted in 1x PBS buffer with 2% horse serum and 0.1% Triton-X, then incubated with the larvae overnight at 4°C with rotation. Larvae were rinsed 5X with PBS solution, then placed in secondary antibody (ThermoFisher, goat anti-mouse IgG2a, Alexa 488, 1:1000) diluted in PBS with 2% horse serum and incubated for 2 hr at 22°C with rotation. Fish were rinsed 3X with PBS then incubated with DAPI (Invitrogen, 1:2000) in PBS for 20 min at 22°C to label cell nuclei. Larvae were rinsed 2X with PBS then mounted onto glass slides with elvanol (13% w/v polyvinyl alcohol, 33% w/v glycerol, 1% w/v DABCO (1,4 diazobicylo[2,2,2] octane) in 0.2 M Tris, pH 8.5) and #1.5 cover slips.
Confocal Imaging and Processing
Immunolabeled z-stack images were acquired via an ORCA-Flash 4.0 V3 camera (Hamamatsu) using a Leica DM6 Fixed Stage microscope with an X-Light V2TP spinning disc confocal (60 µm pinholes) controlled by Metamorph software. The region of interest tool in Metamorph was used to select specific neuromasts (∼700 x 700 px) from the surrounding area. Z-stack images of 100 ms exposure were acquired through with a 10X/0.3 N.A. dry lens in 2 µm slices or through a 63X/1.4 N.A. oil immersion lens 0.5 µm slices. Excitation for DAPI (405 nm) and Alexa 488 GFP was provided by 89 North LDI-7 Laser Diode Illuminator on the lowest power setting (20%) that could acquire images and minimize photobleaching. Confocal images were processed in FIJI (ImageJ, NIH) software to create maximal z-stack maximal projection composite images with minor exposure and contrast adjustments. Composite images were stitched together in Adobe Photoshop CC.
Experimental Apparatus for Rheotaxis Behavior
A microflume (220 x 100 x 40 mm; Fig. 1a) was constructed in two pieces from clear resin (RS-F2-GPCL-04, Formlabs) using a high-resolution 3D printer (Form 2, Formlabs) and joined with two-stage epoxy. A 1 mm-thick top layer of plastic was cemented with silicone sealant downstream of the flume pump (Fig. 1a) to prevent water spillage out of the apparatus. The low-pressure portion of the flume away from the flume pump remained open at the top to facilitate the addition and removal of water and fish when necessary. A square (30 x 30 x 10 mm) arena was 3D printed from clear resin and 25 µm mesh was cemented with epoxy on to the open upstream and downstream sides, allowing water to flow through the working section. The arena served to isolate larval fish at a specific location within the flume, enabling reliable video recording. A 6V bow thruster motor (108-01, Raboesch) was inserted into the flume (Fig. 1b) and used to pump water through the flume at a constant velocity. The water pump was connected to a 7.2 V power supply and the flow rate was modulated by an inline rheostat. An Arduino (UNO R3, Osepp) with a digital display using custom scripts was used to control and monitor the onset and duration of water flow in a consistent manner.
Rheotaxis is a multimodal behavior that is mediated by the visual, vestibular, mechanotactile, and lateral line systems. It was not possible to selectively block the detection of tactile cues in a non-invasive manner. However, evidence suggests that the angular motion in the horizontal plane (i.e. yaw) is not detectible in larval zebrafish at 6-7dpf (31, 32), and we reduced linear acceleration cues to the vestibular system by using a flow stimulus of that rapidly accelerated (< 250 ms) to a constant maximum velocity (V = 9.74 mm s-1). We further isolated the contribution of the lateral line to rheotaxis by performing the rheotaxis assay under infrared light to eliminate visual cues (Fig. 1a). The flume was placed onto a diffused lighting array of 196 LEDs that emitted infrared (IR) light at 850 nm upward through the translucent flume (Fig. 1b). A monochromatic high speed camera (SC1, Edgertronic.com) without an IR filter was placed on a weighted tripod directly over the flume to record behavioral trials at the following settings for optimal speed of data acquisition and machine learning analyses: either 200 or 60 frames s-1 (fps), ISO 20,000, and 1/1000 s shutter speed. The 60 mm Nikon macro lens was manually set to an aperture of f16 to ensure adequate depth of field. The live video feed was monitored on the PC and video recordings were remotely triggered by the Arduino. Videos of each trial were recorded onto 64 GB SD cards for subsequent archiving in duplicate onto a 10 TB RAID array (Mercury Elite Pro Dock, OWC).
Rheotaxis Behavior Pre-trial Preparation
The camera was connected to the local area network (LAN) and initialized according to the manufacturer procedure (http://wiki.edgertronic.com/index.php/Quick_start_guide). A layer of light diffusing material (several Kimwipes© in a sealed plastic sheath) was placed on top of the IR light array. The flume was placed on the diffusion material and filled with E3 media (28°C) and the arena was then placed within the flume. The camera height over the flume was determined by the maximum image size of the arena within the video frame that allowed a ∼10-20 pixel border for subsequent cropping during video pre-processing. Because IR light is a significant source of heat, a thermometer was placed into the open portion of the flume and mini frozen ice packs (2 x 2 cm; -20°C) were used to maintain a consistent temperature range of 27-29°C.
Rheotaxis Trials
Under IR illumination, a single larval zebrafish was transferred by pipette from the 6-well plate to the arena in the flume. The swimming activity of the fish was monitored for ∼10 s to ensure that it exhibited the burst-and-glide behavior indicative of normal larval zebrafish swimming (41). The Arduino was activated, and the trial began by recording 10 s of baseline swimming behavior on video. After 10 s had elapsed, the flume pump was automatically activated for 20 s and delivered a constant mean (± SD) water flow stimulus of 9.74 ± 1.43 mm s-1 (approximately 2 larval fish body lengths s-1) while the camera continued to record swimming behavior. After a total of 30 s (no flow = 0-10 s; initial flow = 10-20 s; and final flow = 20-30 s) had elapsed, the pump turned off, the camera stopped recoding, the file was saved to the SD card, and the larvae was removed from the arena. Cohorts of five individual fish from each group (control, neomycin, or CuSO4) were tested before switching to a new group of fish. After five fish from each group had been tested, the process was repeated for up to four cohorts for the entire experiment.
3D Markerless Pose Estimation (DeepLabCut)
Equipment
We used a computer dedicated to behavioral data acquisition and analysis that was based on a Dell Precision 3630 workstation with the following specifications: Windows 10 operating systems, Intel Xeon E-2246G processor, 64 GB RAM, multiple 2TB SSD hard drives, EVGA GFORCE RTX 2080Ti video card, dual 24” 4K monitors, Dell Thunderbolt 3 PCIe Card, OWC Mercury Elite Pro Dock (TB3RSDK24T) -24TB Thunderbolt 3 Dock and Dual-Drive RAID configured as 12TB RAID 1.
Installation
Our GPU equipment required the installation of Tensorflow 1.12 with the NVIDIA CUDA package prior to installing multi-animal DeepLabCut2.2b8 (maDLC; 42Mathis et al. 2018, 43Nath et al. 2019), Python 3.6, and all dependencies in an Anaconda virtual environment according to (https://github.com/DeepLabCut/DeepLabCut/blob/master/docs/installation.md).
Detailed tutorials for using maDLC with a single animal are available online (https://github.com/DeepLabCut/DeepLabCut/blob/master/docs/maDLC_AdvUserGuide.md, https://www.youtube.com/channel/UC2HEbWpC_1v6i9RnDMy-dfA); however, the pertinent details of our procedure are as follows.
Dataset Curation
To reduce computational load, all video files were pre-processed by down-sampling to 1000 x 1000 px then cropping out dead pixels (i.e. black spots) on the camera sensor and extraneous portions of the video by using the video editor function of maDLC.
Project Creation
A new single animal maDLC project was created and the project config.yaml file was modified to include seven unique body parts that were easily identifiable on a larval zebrafish (left_eye, right_eye, swim_bladder, tail_1 (near the base), tail_2, tail_3, tail_4 (tail tip), Fig. 1a). The skeleton was initially created by forming individual connections between the seven body parts in the config.yaml file. We then extracted 20 frames per video from a curated set of ten videos that contained representative examples of the target behaviors and experimental treatments. The seven body parts were labeled on the fish in each frame using the graphical user interface (GUI). The annotated frames were checked for accuracy and multiple additional skeletal connections between the seven body parts were added using the skeleton builder GUI pop up window, helping maDLC learn faster and create accurate models.
Pose Estimation
The training dataset was created using cropped images (400 x 400 px) to reduce computational loading and the default setting of Resnet-50 pre-trained network weights, and imgaug data augmentation. We set the display iterations to 100, save iterations to 1000, and trained the network until the all of the loss parameters plateaued at 100,000 iterations. The network was evaluated (PCK values close to 1, RMSE values low) and cross-validated using the default parameters.
Identity Tracking
Curated videos were analyzed and the detections (pickle files) were assembled into tracklets (h5 files) using the box method, providing superior results for these data compared to the skeleton method (https://github.com/DeepLabCut/DeepLabCut/blob/master/docs/maDLC_ UserGuide.md). The interactive tracklet GUI was launched and the original videos with the newly created labeled body parts were used to correct for outliers in the tracklet data files. Jitter and wholesale misidentification of the body part data were changed and saved.
Post Processing
The results were plotted for each video and a new labeled video was created using the original curated videos and the refined tracklet data. The labeled output videos were checked for labeling accuracy and there was no need to extract additional frames to augment the data set. Novel videos were batch processed beginning at the video analysis step and ending with the plot trajectories and label videos step. Once several videos of different conditions and behavioral responses could be accurately labeled, the create labeled videos step was omitted to reduce computational load, increase analysis productivity, and reduce hard drive space storage requirements.
Supervised Behavioral Annotation, Classification, & Analysis (SimBA)
Definition of Rheotaxis Behavior
We defined positive rheotaxis as when the larvae swam into the oncoming water flow at an angle of 0° ± 45° for at least 100 ms. The movement component was used to distinguish active swimming behavior from passive drift in cases where larvae would orient into the flow, stop swimming, and drift backward without rotating in the X-Y plane.
Installation
SimBAxTF-development version 68 (44Nilsson et al. 2020), Python 3.6, Git, FFmpeg, and all necessary dependencies were installed in a separate Anaconda virtual environment: (https://github.com/sgoldenlab/simba/blob/master/docs/installation.md).
Dataset Curation
The pre-processed video files used in maDLC analysis were converted to the AVI format using the video editor function of SimBA.
Build Classifiers
Project Creation
A new SimBA project was created according to Scenario 1 as shown here: https://github.com/sgoldenlab/simba/blob/master/docs/Scenario1.md. Before the project config file was generated, we selected the user defined pose configuration option and the DLC-multi animal options. The user-defined pose configuration nomenclature must match the body parts and individual animal labels used in the maDLC config file. Next, several curated videos of fish from different treatments that displayed ideal examples of rheotaxis were imported into the project folder. Then the h5 files (final tracklet data) generated by DLC for these videos were imported. Last, the frames of each video were extracted into the project.
Load Project
The project was loaded into SimBA and the video parameters were set for each of the curated videos. We made certain to set the correct frame rate (fps), resolution, and pixel measurements (px/mm), as these settings affected downstream analyses. Outlier correction was achieved by selecting two body parts that were reliably labeled by DLC but not too close together on the animal (e.g. swim_bladder, tail_3); movement criterion was set to 0.7; location criterion was set to 1.5 (44). Features were extracted using a custom R script, which is only possible within the development version of SimBA. Rheotaxis behavior was labeled (i.e. annotated were created for predictive classifiers) for each of the curated videos, but one was not annotated and was set aside for validation. The model was created by training it according to the default machine model settings, hyperparameters, and model evaluation settings listed in Scenario 1 (44). Model validation was done using the aforementioned curated video, its associated csv file that was previously set aside without annotation, and the model file that was just created. Throughout the video, the interactive plot was used to evaluate the probability threshold (i.e. the smallest peaks) where rheotaxis was accurately predicted by the model and the minimum duration of each predicted rheotaxis event. These values became the discrimination threshold and minimum behavior bout length settings when we ran the machine model. All options were chosen during the analysis of the machine results and output files were saved in the project_folder/ logs subdirectory. Subsequent analyses were performed in R using the csv data files located in the machine_results subdirectory of the project/csv directory.
At this point the model was created and was used to predict rheotaxis on novel videos as outlined in Scenario 2: (https://github.com/sgoldenlab/simba/blob/master/docs/Scenario2.md #part-3-run-the-classifier-on-new-data). Videos and data files were archived after they were processed and analyzed to avoid reanalyzing them when new videos and data were added for analysis.
Data Analysis
Because rheotaxis data were limited to instances when fish swam at angles 0° ± 45° under flow conditions, they were compared to random swimming data of fish with a body angle of 0° ± 45° under no flow. The only exceptions to this procedure were for the angular comparisons among groups where the total data set of orientation angles was used to determine if the fish were randomly distributed under no flow (Fig. 3) or had a natural proclivity to orient toward the front of the arena, and the X-Y (2D) spatial use density plots under no flow (Fig. 6a) and flow (Fig. 6b). Time series data were quantized into 100 ms bins because it was the lowest common bin size for videos shot at 200 and 60 fps.
Data wrangling and cleaning was performed in R (45) with the packages tidyverse (46), dplyr (47), plyr (48), and readbulk (49). Figures and graphs were created with packages circular (50), ggplot2 (51), and viridis (52). The Rayleigh statistical tests (V-test) of uniformity for circular data in a specified mean direction (mu = 0°) and the Watson-Wheeler tests for differences in the grand mean body angle or angular variance (the test does not specify which parameter differs) were performed with the package CircMLE (53). The mean duration, number, total distance travelled and mean latency to the onset of rheotaxis events were calculated for 10 s bins (no flow, initial and final flow) using the SimBA developer version. The generalized linear mixed models with post hoc t-tests using the Satterthwaite method were performed in R with the packages lmerTest (54) and lme4 (55), and the significance values for fixed effects were done using stats (45) for type III ANOVAs. The packages zoo (56) and spectral (57) were used to convert data into times series and perform spectral decomposition analyses, respectively.
Declaration of interests
The authors declare no competing financial or non-financial interests.
SUPPLEMENTARY INFORMATION
Acknowledgments:
We would like to thank David Lee for editorial support and Mark Warchol for feedback on the manuscript. This work was supported by the National Institute on Deafness and Other Communication Disorders R01DC016066 (L.S.), NIDA R00DA045662 (S.A.G.), NIDA P30 DA048736 (S.R.O.N. and S.A.G.), NARSAD Young Investigator Award 27082 (S.A.G.).