A central question in systems neuroscience is how neural circuit assembly and function relate to animal behavior. Genetic screens in invertebrate models, such as Drosophila melanogaster and Caenorhabditis elegans have begun to unravel the genetic basis of circuit function and behavior (Chalfie et al., 1985; Moore et al., 1998; Scholz et al., 2000). Automated methods have recently been developed in these species to track the position of individuals alone or in a group (Branson et al., 2009; Swierczek et al., 2011) and to categorize behavior (Dankert et al., 2009; Kabra et al., 2013). The zebrafish has emerged as an important vertebrate model organism for developmental biology, neurobiology, and human disease models, and is now used as a genetic model organism for the study of the mechanisms modulating complex behaviors in vertebrates such as depression and anxiety (Blaser et al., 2010; Lee et al., 2010; Cachat et al., 2011; Vermoesen et al., 2011; Zakhary et al., 2011; Ziv et al., 2013), sleep (Zhdanova et al., 2001; Appelbaum et al., 2009), or addiction (Petzold et al., 2009; Khor et al., 2011). The permeability, small size, genetic tractability, transparency, and low cost of zebrafish make them highly suitable for large-scale genetic and chemical screens (Driever et al., 1996; Granato et al., 1996; Haffter and Nusslein-Volhard, 1996).

Although simple for a vertebrate, the locomotor patterns of the zebrafish larva bring technical challenges to automated analysis. Larvae spontaneously swim in discrete bouts in a manner often described as “beat and glide,” which can be classified as individual maneuvers, including slow forward swim, routine turn, or escape. These short movements are characterized by a large range of tail-beat frequencies (15–100 Hz), which require high-speed imaging to capture accurately and can be separated by long resting periods of up to a few seconds. Manual tracking via frame-by-frame analysis has formed the basis of contemporary knowledge and has enabled initial characterization of the larval zebrafish locomotor repertoire (Budick and O'Malley, 2000; Borla et al., 2002; McElligott and O'Malley, 2005). However, manual techniques are both laborious and limited in scope for high-throughput screens (Driever et al., 1996; Granato et al., 1996; Haffter and Nusslein-Volhard, 1996). The currently available automated tools have limitations in either refinement or time-scale. Recent chemical or genetic screens have relied on commercial software that estimates an index of mobility of the larvae, usually measured as the distance traveled during a recording session or the amount of time spent moving (Rihel et al., 2010; Elbaz et al., 2012; Rihel and Schier, 2012). These approaches for high-throughput screens provide information about average velocity and distance traveled by tracking the animals' center-of-mass over minutes to hours. Previous studies have either focused on analyzing movement duration and speed at low frequency over long periods of time or on fine analysis of kinematics at high frequency but for very short acquisition (typically 1000 ms, Burgess and Granato, 2007; Liu et al., 2012). Accurate categorization of maneuvers for each individual in a group requires novel methods to record behavior with high temporal resolution and over long durations, automatically tracking and categorizing thousands of maneuvers.

Here we developed a new program, ZebraZoom, to track the full body position over a multiple-minute timescale of 56 larvae simultaneously recorded at high frequency and to finely characterize each maneuver. To identify core and tail positions for large datasets, videos were obtained on multiple larvae simultaneously over long periods of time and at high resolution using a high-speed camera run in a streaming-to-disk interface (Methods). Typically 500–1000 movements from seven larvae were recorded per dish in four minutes and eight dishes were monitored in parallel. To simplify tracking, we placed larvae in conditions that reduced overlapping in the z-plane during swimming (Methods; overlaps occurred on average once every 145 s per larva). We developed an offline 2D tracking method for identifying and separating each animal even when in close contact (Methods, Figure 1). For each larva several features were identified, a core position that included the head and swim bladder (Figure 1A) and ten points along the tail (Methods and Figure 1B; Video S1). As movements occurred as discrete episodes, ZebraZoom detected movements based on the tail-bending angle over time (Methods and Figures 1C–D). To validate the accuracy of movement detection, one trained experimenter manually identified all movements occurring in a subset of videos. In three videos representing a total of 189 events, movements occurred with a false negative rate of 2.7% and a false positive rate of 3.7%.

To quantify movements in a consistent manner, we used the location of the head, the position of the tail, the heading direction and the tail-bending angle to estimate global parameters of locomotion (Figure 2A, Methods). We observed that movements for 5–7 dpf wild-type (WT) larvae occurred every 2.22 s on average per larva (at 0.4495 ± 0.0117 Hz). For all movements identified, larvae performed on average 3.19 ± 0.01 oscillations per movement, had a 24.29 ± 0.03 Hz tail-beat frequency (TBF), lasting 189.5 ± 0.0004 ms with a 51.14 ± 0.18° heading direction range, 2.49 ± 0.008 mm traveled distance and 13.35 ± 0.04 mm/s speed per maneuver. We illustrated the use of ZebraZoom for quantifying the effects of a known glycinergic receptor antagonist, and for analyzing a blind genetic mutant. Glycine is responsible for reciprocal inhibition in the spinal cord that permits left-right alternation to sustain oscillations (Dale, 1985; Grillner et al., 1995; Granato et al., 1996; Drapeau et al., 2002; Li et al., 2004). In zebrafish, mutants for glycinergic receptors or transporters have been associated with defects in motor pattern generation (Granato et al., 1996; Odenthal et al., 1996; Hirata et al., 2005; Masino and Fetcho, 2005). We measured the effect of bath application of 75 μM strychnine on spontaneous locomotor activity in larvae and compared to control siblings that were not exposed to the drug (Figure 2B, Methods). For control larvae, we did not observe a significant change in the occurrence of movements over time (0.35 ± 0.05 movements per larva/s before and 0.27 ± 0.03 movements per larva/s after), or on any of the global parameters (Figure 2B; all p > 0.15). However the locomotor behavior of larvae treated with strychnine was significantly impacted (Figure 2B). Overall, movements occurred less frequently (0.30 ± 0.04 Hz before and 0.12 ± 0.02 Hz after, p < 0.0002). Although the average TBF during a movement did not change (p > 0.81), the number of oscillations decreased (3.52 ± 0.17 before and 2.89 ± 0.16 after; p < 0.0078), an effect that was associated with a decrease in movement duration (p < 0.0001), distance traveled (p < 10⁻⁵), and average speed (p < 10⁻⁵). Strychnine application also resulted in a decrease in the range of heading direction (p < 10⁻⁵). atoh7 mutant larva lack retinal ganglion cells, rendering them blind (atoh7^−/−, Kay et al., 2001). Considering the importance of vision for zebrafish larvae, analyzing their locomotor output could reveal corresponding behavioral differences. Overall atoh7^−/− mutants generated episodic movements less frequently than control siblings (0.33 ± 0.02 Hz vs. 0.51 ± 0.02 Hz, 112 larvae for each condition). Quantitative analysis of global parameters of the blind mutants showed no difference in the average TBF or the average speed per larva (Figure 2C; p > 0.85 and p > 0.83, respectively) but there were small but significant decreases in the number of oscillations, duration, heading direction range, and distance traveled (Figure 2C; all p < 10⁻³). These defects were observed systematically in four clutches. atoh7^−/− mutants thus display small but substantial differences in basic motor behavior when compared to control siblings.

Zebrafish larvae display a variety of locomotor maneuvers that are often grouped into discrete categories. In these experimental conditions, three types of movement occur in groups of larvae at early stages: slow forward swims (S), routine turns (T, also referred to as slow turns), and escapes (E, including C-turns or burst swims). Figure 3 shows examples of these movements reported by ZebraZoom. For each maneuver, we superimposed a succession of images (Figures 3Ai–Ci), the tail-bending angle over time (Figures 3Aii–Cii) and the curvature along the rostro-caudal axis and as a function of time (Figures 3Aiii–Ciii). The three types of maneuvers included a series of slow left-right alternation; high values of curvature were confined to the caudal tail (Figures 3Aiii–Ciii). While high values of curvature of the tail were confined to the caudal end for slow forward swims (Figure 3Aiii), high values of curvature were distributed from head to tail for routine turns and escapes (Figures 3Biii,Ciii). Stereotypical routine turns and escapes differed by the frequency of left-right alternation in the tail bend (Figures 3Biii,Ciii). As larvae did not always exhibit a canonical slow forward swim, routine turn or escape, some movements were ambiguous. To estimate the percentage of these movements, four experimenters subjectively classified 390 movements distributed over eight videos. Overall about 82% of all movements were classified uniformly by at least three out of four experimenters (Methods) indicating that 18% of movements were difficult to categorize.

Using knowledge of stereotypical locomotor events, we designed a multiclass categorization approach with supervised machine learning to automatically sort each movement into one of the three categories. To implement the multiclass categorization, we used two successive support vector machine (SVM) classifiers: the first classifier sorted S vs. all other maneuvers, and when necessary the second classifier sorted T vs. E. Locomotor events were segregated subjectively in the training set (n = 201). This machine learning approach relied on associating dynamic parameters extracted from the tail-bending angle over time with each maneuver type identified in the training set (Figure 4A and Methods). To reduce the dimensionality of the data, we performed Principal Component Analysis (PCA). Based on the selection of a trained experimenter on the learning set, we validated the multiclass categorization to sort maneuvers by comparison with the subjective classification performed by a trained experimenter for a recognition set (n = 189; Figures 4B,C). We observed that ZebraZoom agreed with the trained experimenter 82.5% of the time for the recognition dataset (85% for S, 82% for T, and 79% for E; Figures 4B,C; Table 1 and Methods). When compared to four independent experimenters, ZebraZoom reached 91% accuracy for categorization of stereotypical maneuvers that all experimenters had unanimously identified and 76.4% on average for all maneuvers (73.2–82.5%, Table 1). Once validated, we applied the ZebraZoom categorization algorithm on a large dataset of 44,688 movements of WT larvae (Figure 4D). We identified 14.911 S (33.36%), 21,432 T (49.96%), and 8,345 E (18.67%). The distribution of global parameters for the three classes of maneuvers were similar in terms of number of oscillations and duration, but they differed in terms of mean TBF, heading direction range, distance traveled and speed (Figure 4D).

The investigation of interactions between individuals leading to coordinated motion in animal groups has been a long-standing challenge that is central to elucidating the mechanisms and evolution of collective behavior. Most studies have focused on the analysis of speed or directionality to reflect the interaction between animals (Katz et al., 2011; Gautrais et al., 2012). We availed ourselves of ZebraZoom's features to accurately identify each larva and categorize their maneuvers to study how larvae interacted. In comparison to juvenile and adult zebrafish that swim continuously, larval zebrafish swim episodically with maneuvers that occur in a beat-and-glide manner. Each movement can be regarded as a discrete event, therefore we were motivated to explore how local perturbations of a single individual could impact the group. The program switched identity of larvae once every 109 s (once every 49 movements on average), allowing us to track single larvae. We modeled sequences of maneuvers performed by larvae within a group as Markov chains. Utilizing the classifier, we described larva–larva interactions in a group and intrinsic properties of individuals. We calculated a transition index (I) for each sequence of two maneuvers as the transition probability between first and second maneuvers divided by the probability of random occurrence of the second maneuver (Figure 5; Table 2 and Methods). When the two successive maneuvers were the same, a higher transition index indicated the probability of repetition of this maneuver was greater than chance. The transition index was equal to one when the order of sequential maneuvers was random. Overall I was greater than one for repetition of the same maneuvers (Table 2). We sorted the data into interactions between different animals and the repetition within the same animal. We analyzed how the transition index for a given succession of maneuvers depended on the distance between the two larvae's core positions at the onset of the movement and the time between the onset of each movement (Figure 5 and Methods). Individual larvae often performed the same type of maneuver sequentially (maximal values I_{S max (same)} = 1.43, I_{T max (same)} = 1.37, I_{E (same)} = 2.38, all p < 0.002; Figure 5A, Table 2, and Methods). Although slow forward swims or routine turns were not frequently repeated between larvae (I close to 1: maximal values I_{S (diff)} = 1.09 and I_{T (diff)} = 1.01, p > 0.05; Figures 5Bi,ii, Table 2, and Methods), we found that recurrent escapes were very frequent between different larvae (maximal value I_{E (diff)} = 3.6, p < 0.002; Figure 5Biii, Table 2, and Methods). Five to seven dpf larvae do not show evidence for social interactions (Buske and Gerlai, 2011). By taking advantage of the algorithm for identifying single larva and categorizing simple maneuvers, we reveal that larva–larva interactions primarily occurred for escape responses. These series of escapes occurred after direct collisions (in one third of the cases) or via long distance interaction (two third of cases). Blind atoh7^−/− larvae showed a similar profile of interactions for escapes (data not shown); these interactions were most likely mechanically triggered.

Large-scale chemical and genetic screens would benefit from a quantitative approach to analyze fine locomotor patterns over long periods of time. Compared to other genetic models, zebrafish locomotion is difficult to analyze because larvae initiate maneuvers intermittently and during these short events, the larvae swim at a high speed with TBFs ranging from 15–100 Hz. The quantitative analysis of motor behavior for large-scale screens requires solving the problem of recording multiple animals simultaneously at high frequency (above 200 Hz) and for long periods of time (minutes). Here we implemented a reliable method for quantifying global parameters of movements based on stream-to-disk recordings acquired at high frequency and over long periods of time, limited only by data storage. Next we developed a robust method for tracking the full body position of zebrafish larvae swimming in groups. We first manually validated that the tracking accurately detected discrete movements, and then used the global parameters obtained to characterize the locomotion of WT larvae. Quantification of the global parameters describing larval movements corroborates previous observations based on fewer samples (Budick and O'Malley, 2000; Danos and Lauder, 2007; Liu et al., 2012). Similar estimates of the duration of movements, distance traveled and speed were obtained from the recent application of C-trax (designed originally for Drosophila) to zebrafish larvae [Lambert et al. (2012) based on Branson et al. (2009)]. In these conditions, recordings at low frequency over long periods of time, typically 60 Hz for minutes or hours, revealed the global level of activity over time but no information on fine kinematics during individual maneuvers (Elbaz et al., 2012). When recordings were performed at high frequency to capture the dynamics of motion, they usually lasted 1000 ms (Burgess and Granato, 2007).

We illustrated the benefit of ZebraZoom to quantify global parameters of movements by analyzing the effect of a drug to block glycinergic neurotransmission, which has been known to be involved in motor pattern generation and alternation between the left and right side of the spinal cord across vertebrate species (Grillner, 2003; Korn and Faber, 2005; Nishimaru and Kakizaki, 2009). Most studies relied on ventral nerve root recordings where muscles were dissected out or paralyzed in order to record the activity of motor neurons at the level of a few segments at most. Our automated quantification of locomotor events enabled identification of effects induced by bath application of the glycinergic antagonist strychnine on locomotion in intact animals. As predicted, bath application of strychnine dramatically reduced the occurrence of movements and the number of oscillations per movement, that was correlated with a reduction of the duration of movement and of the distance traveled. While mean TBF was not affected, we observed a reduction in the heading direction range and in speed. Our approach pinpointed effects of glycinergic blockade, including a reduction in the number of oscillations per movement, a kinematic feature not estimated in commercially available software. The analysis of the mutant atoh7 revealed that although TBF and speed were not affected in the blind mutant, there was a small but significant decrease in the number of oscillations, heading direction range, distance, and duration of each bout compared to their control siblings. These effects were systematically observed on four clutches suggesting that visual feedback may impact some global parameters of locomotion. However since the pattern of expression of atoh7 has not yet been fully characterized, it cannot be excluded that the gene may be expressed in cells other than retinal ganglion cells.

The originality of ZebraZoom lies in categorizing all maneuvers performed by individual larvae in a group. The subjective analysis of maneuvers based on four independent experimenters revealed that locomotor maneuvers were not obvious to categorize. Based on subjective estimates, 18% of all movements corresponded to ambiguous maneuvers. By using a machine-learning paradigm, we trained ZebraZoom to categorize all maneuvers over tens of thousands of movements with 82.5% accuracy, a similar value to the 72% agreement rate of all four experimenters measured over a few hundreds of movements. The approach we developed here could be expanded to include directionality of the turns, sequences of maneuvers such as those occurring during prey tracking, and subcategories of escapes.

This study constitutes an important first step for accurate tracking of multiple larvae in groups over long periods of time and for categorizing maneuvers. Some improvements could be implemented in the future. While our tracking method currently relies on a simple “blob” approach solely based on raw image analysis, a model-based approach may be more reliable in particular when animals are in close contact (Fontaine et al., 2008). We show here that ZebraZoom can achieve an accurate categorization of maneuvers, comparable to experimenters' estimates, based solely on the dynamics of movement of head and tail. An interesting avenue of exploration to address this could be investigation of novel dynamic parameters for the learning and recognition process of the classifier to yield subtler methods for detection of defects. Quantification of motor patterns in C. elegans is based on a description of all possible positions of the animal over time (Stephens et al., 2008). In order to fully understand larval zebrafish behavior we need to identify a minimal set of parameters sufficient to describe all motor patterns. All together this work brings new insight to the complexity of behavior determination in zebrafish larvae and could be applied to investigation of the mechanisms of addiction, arousal, feeding, social interaction and aggression in larvae and juveniles (Gahtan et al., 2005; Bianco et al., 2011; Buske and Gerlai, 2011; Miller and Gerlai, 2012; Ziv et al., 2013). The observation of complex interactions in juveniles raises the hope that it will soon be possible to investigate the neuronal circuits and molecular pathways underlying social interactions. The fact that we can track individual larva and analyze their interactions is a major advance over existing methods. Our approach that systematically quantifies and categorizes thousands of motor patterns was designed to bring efficiency and reliability to drug screening and forward genetic screens. ZebraZoom can detect, quantify, and categorize movements to provide a quantitative description of global parameters as well as a qualitative description of all maneuvers performed by individual larvae.