Deep phenotyping of open-field recordings

We recorded over 700 trials in an open-field arena from 162 mice (Table 1). Each mouse was recorded for 20 min every day for a period of four days. We imaged the mice from below a transparent floor to allow observation of the paws and base of the tail. Movies were analyzed with a semi-supervised behavioral classification and behavioral labeling scheme (Fig. 1 and Additional file 1: Fig. S1). In the first step, we used a deep neural network to measure the posture of each animal over time. Recent advances in body part tracking allow for the use of a small set of user-generated labels to train a neural network for labeling body parts [21]. We trained the LEAP network with 660 human-annotated images and achieved position estimates with median confidence probabilities ranging from 0.91 to 0.98 for the snout, chin, inner and outer limbs, and base and tip of the tail. Ears, body center and sides, and the tail center point were excluded due to lower estimation accuracy (median confidence probability of 0.81). This pose estimation step resulted in a time series of the two-dimensional position of each body landmark, \(}_i(t)\). The next step in our analysis was a semi-supervised clustering of postural dynamics. Our previous work on behavioral clustering analyzed the dynamics observed in whole images of an animal but did not specifically probe the dynamics of individual body parts [19, 20, 22]. We adapted this method for use with the body part position time series, \(}_i(t)\) [21] and clustered the dynamics across timescales to define eight major classes of behavior in the open field (Fig. 1, Additional file 1: Fig. S1). The resulting ethogram revealed details not visible from centroid tracking alone (Additional file 1: Fig. S1a), namely the structure of spontaneous behavior across mesoscopic timescales.

Fig. 1

Processing of video recordings in the open field produces multi-scale quantitative descriptions of behavior. The pipeline takes virtual markers from pose estimation with LEAP to find behavior clusters and generate wavelet signatures. I. A visual representation of the embedding/clustering steps. Distance matrix calculated between virtual markers per each time frame transforms into the frequency domain and clustered using k-means. II. Raw joint trajectories are used to create wavelet signatures or ’behavioral fingerprints’ by finding the mean power spectrum during each behavioral cluster found in part I

Table 1 Summary of experimental strains and number of mice recorded

The time series \(}_i(t)\) was rotated and translated into the reference frame of the animal by aligning with the anterior–posterior axis defined by the snout and tail base body parts and transformed into the frequency domain using wavelet decomposition. This produced a high-dimensional output that captured multi-scale (0.25–20 Hz) changes in posture over time (Additional file 1: Fig. S1c). We then performed k-means clustering of this high-dimensional data on a balanced selection of samples from recordings across all male mice. We initially clustered the sampled data into \(k=100\) clusters, enough to achieve fine-grained clustering that in some cases was not distinguishable by eye. Data from all experiments were then clustered by assigning each time point to the cluster of its nearest neighbors in the k-means training set.

An exploration of all 100 clusters revealed several broader classes of behavior exhibited by mice in the open field (Table 2). Similar to recent findings from fruit flies [38], we found that mouse behavior could be organized into coarse classes with recognizable subclasses representing variable aspects of behavior. For example, the 21 clusters that made up the ’locomotion’ class represented the diversity of locomotion movements recorded, differing in velocity, amplitude and speed of limb movement, and the coordination of the limbs.

Table 2 Eight coarse behavioral classes are used to describe the 100 fine-grained clusters obtained from a k-means clustering of the posture dynamical signal Fig. 2

Spatial–temporal structure of C57BL/6J male mice behavior in the open field. a Behavioral cluster frequency across mouse models on Day 1. (left) Heatmap of the total behavioral occupancy of each behavioral cluster; rows are individual animals and columns are behavioral clusters. All rows within the 100-cluster matrix sum to one and describe behavior used over 20 min. Clusters are ordered by median centroid speed for each of the eight behavioral classes (bottom). (right) Heatmap of the behavioral occupancy for each of the eight behavioral classes. b Total occupancy of behavioral classes in a 20-min recording quantified by human annotators and by the algorithm. c Log ratio differences for all 8 behavioral classes between selected groups. The error bars show the bootstrapped 95% confidence interval (N = 5000, percentile bootstrap). d The spatial distribution is shown for each of the eight behavioral classes. e Behavioral usage for each of eight coarse categories plotted for C57BL/6J male mice for each of four observation days. All individuals are shown as points, colored traces correspond to the median fraction of time spent in the behavior for each day. f The usage of grooming and locomotion behaviors during the 20-min observation period for Days 1 and 4. Shaded regions represent the 95% confidence interval (left). The mean usage (right) for each of eight behavioral classes is shown for Days 1 and 4

We categorized each of the 100 clusters into one of eight behavioral classes (Additional file 1: Fig. S1 and Fig. 2a). This manual curation step revealed that time spent in the open field was composed of commonly studied behaviors such as locomotion and grooming, but also less distinct movements made for a large fraction of the time including spatial exploration and ambling. We defined the classes fast exploration and slow exploration to characterize periods that are usually unaccounted for in traditional measurement paradigms. Fast exploration, for example, included quick turns and sniffs that are characteristic of alertness or anxiety, whereas slow exploration included head sweeps and extensions during calm periods.

We validated the accuracy and interpretability of these classes by visually inspecting randomly sampled movies of behavior from each of the 100 clusters and the eight classes. Movie segments were sampled from all C57BL/6J male movies (300 20-min movies overall) and limited to examples where the behavior in question was performed for at least 250 ms or 20 frames at our frame rate of 80 Hz. Examples from the 100 behavioral clusters and eight behavioral classes are available in the Supplementary Materials. We also compared the eight classes with manual annotations by three researchers who were instructed to label each frame after reading the descriptions from Table 2 and watching selected video snippets. Annotators showed high agreement with one another, but annotators and the algorithm agreed less often (Additional file 3: Fig. S3a, c). Human-vs-algorithm mismatches may have arisen from the difficulty of conceptually capturing a class with a category name. In addition, the algorithm uses spectrograms whose power is concentrated at fast timescales, whereas human annotators may use longer timescale information. This possibility is supported by the fact that the algorithm captured the beginning and end of locomotion bouts to match the limb kinematics, whereas human annotators would often classify beginnings and ends into other behaviors. Human annotators also tended to merge bouts of behavior into a single category (Additional file 3: Fig. S3b); for example, human annotators might combine a series of ‘rear’ and ‘climbing’ episodes as a single bout of ‘climbing’ where the algorithm made subdivisions based on high-movement hindlimb lifts from the arena floor. Altogether, differences between manual annotations and semi-supervised behavioral classifications were reflected in the fractional occupancy by each behavior (Fig. 2b).

Behavioral differences upon introduction to the open-field arena

The fraction of time spent in each cluster in the experimental arena on the first day of exposure (day 1) is shown in Fig. 2. Clusters in Fig. 2a are ordered horizontally according to their behavioral class and then by median centroid speed within each class. Mice spent the most amount of time on day 1 locomoting after placement in the novel arena, followed by turning, climbing and rearing, but almost no time in the idle state.

Inspection of the occupancy of locomotion modes across individuals revealed features that set the Cntnap2 KO and L7-Tsc1 mutants apart from their WT littermates, as well as from the large number of C57BL/6J mice, which had the same genetic background. C57BL/6J mice were most likely to use the fourth- and fifth-fastest locomotion modes at around 0.2 m/s on Day 1. The Cntnap2 KO was distinct from WT and heterozygote littermates by the enriched usage of the several fastest locomotion clusters and a decrease in more moderate locomotion (Fig. 2a). This result was in agreement with findings that Cntnap2 KO mice are hyperactive [27] and accounted for the large distance these mice covered on the first day in the open field (Additional file 2: Fig. S2). Full mutants from the L7-Tsc1 group showed the opposite trend from Cntnap2 KO mice, spending less time in the fast locomotion clusters than any other group. The L7-Tsc1 mutants often used several clusters of slow locomotion that were uncommon in either their control littermates or the baseline C57BL/6J mice.

To investigate overall differences in the time spent in the eight behavioral classes, we performed a compositional data analysis [36]. After transformation of the fractions into isometric log ratio coordinates, differences between groups were analyzed using a nonparametric multivariate test (Wilks’ lambda-type test statistics). For day 1, we found significant differences at \(\alpha = 0.05\) between female and male C57BL/6J mice. The behavior of male and female C57BL/6J mice was significantly different from Cntnap2 KO mice and their littermates as well as L7-Tsc1 mutants and their littermates, except for between female C57BL/6J and Cntnap2 KO mice. L7-Tsc1 mutants behaved differently than their littermates and Cntnap2 KO mice. No significant difference in the time spent overall the eight behavioral classes between Cntnap2 KO mice and their littermates was observed. To gain further insights into the behaviors that caused the differences between the groups, we calculated log ratio differences between groups for each behavior (Fig. 2c) [36]. Female C57BL/6J mice spent reduced time in slower behaviors (idle, groom, slow explore) and an increased time in locomotion compared to male C57BL/6J mice. A similar observation was made when comparing L7-Tsc1 mutants with their WT littermates on day 1: they spent less time being idle, grooming or slow exploring, but more time rearing, turning and in locomotion.

Spatial habituation is reduced in L7-Tsc1 mutant mice

Mice tend to avoid open spaces in natural environments in an attempt to avoid predators. Traditional analyses of open-field experiments divide the arena into zones to quantify time spent in the open space, near the edges and in the highly confining corners. Mice spend most of their time at the edges and corners (thigmotaxis). Previous researchers have sometimes interpreted thigmotaxis in terms of the emotional state of the mouse, often anxiety [39]. We found that in these corner positions, mice performed a variety of behaviors including grooming, rearing, climbing and exploration (Fig. 2d).

When first introduced to a new environment, mice explored the full space with some preference for corners. With time and repeated exposure to the same environment, mice habituated and spent less time crossing through the center of the arena and more time in the corners (Additional file 2: Fig. S2c, d). We measured the time an animal spent in the corner zones of the arena for each of the 4 days of observation to identify the development of thigmotaxis. C57BL/6J males, as well as WT and heterozygote littermates of L7-Tsc1 mutants and Cntnap2 KO mice, demonstrated a gradual increase over days in time spent in the corners (d range from 0.87 to 2.11 for day 1 compared to day 4, repeated-measures ANOVA for each group, \(p < 0.01\)), while C57BL/6J females, L7-Tsc1 mutants and Cntnap2 KO mice all failed to show a significant increase of time spent in the corners (repeated-measures ANOVA, \(p = 0.13\), \(p = 0.28\), \(p = 0.58\)) (Additional file 2: Fig. S2c). Compared to males, female C57BL/6J mice spent on average 15% less time in the corners on day 1 (d = 0.83, one-way ANOVA, groups-by-day 1, \(F(7,154) = 8.83\), post hoc \(p = 0.008\)) and 26% on day 4 (d = 1.43, groups-by-day 4, \(F(7,154) = 11.15\), post hoc \(p = 0.0002\)). In summary, a propensity toward corners after multiple days was sex-dependent in C57BL/6J mice and failed to occur in L7-Tsc1 mutant and Cntnap2 KO mice.

Reduced spatial habituation was also seen in the number of center crossings for L7-Tsc1 mutants (repeated-measures ANOVA \(F(3,24) = 4.24\), \(p = 0.12\), \(F(3,48) = 47.76\) and \(F(3,48) = 34.36\) for littermates \(p < 0.001\)). L7-Tsc1 mutant mice showed similar initial values of center crossing as WT littermates on day 1 (one-way ANOVA, p = 0.39). But after four days of habituation, mutants showed a tendency to cross to and from the center zone more frequently than WT littermates (one-way ANOVA, groups-by-day 4, post hoc \(p = 0.07\)) (Additional file 2: Fig. S2d).

Behavioral habituation is reduced in L7-Tsc1 mutant mice

Mice also modulate specific behaviors as they habituate to a new environment. All experimental groups traveled significantly less on the second day compared to the first day (\(d = 1.5, 2.1, 2.8, 1.8\) for female, male C57BL/6J mice, Cntnap2 KO and L7-Tsc1 mutant mice, repeated-measures ANOVA for each group, \(p < 0.05\)) (Additional file 2: Fig. S2b). We found that both female and male C57BL/6J mice increased the time they spent idling, grooming, exploring and rearing on day 5 over day 1 (\(d = 0.9, 1.2, 1.6, 0.5\) for females and \(d = 0.9, 1.6, 1.6, 0.7\) for males; repeated-measures ANOVA for each group, \(p < 0.01\)) and decreased the time they spent locomoting (\(d = 1.9\) for females and \(d = 3.2\) for males, \(p < 0.01\)) in the same arena. The largest day-on-day shift occurred between the first and second day in both female and male C57BL/6J mice (Fig. 2e and Additional file 4: Fig. S4a). Male C57BL/6J mice in addition spent significantly less time climbing and turning (\(d = 0.5\), repeated-measures ANOVA, post hoc \(p < 0.01\) and \(d = 2.0\), post hoc \(p < 0.01\)) in contrast to female C57BL/6J mice (repeated-measures ANOVA, \(F(4,72) = 1.36\), \(p = 0.51\) and \(F(4,72) = 2.47\), \(p = 0.17\)).

Behavior also changed within the 20-min observation period of each experiment (Fig. 2f). The pattern of behavioral change exhibited by C57BL/6J mice within an observation period was similar to the across-day change: locomotion, climbing and turning decreased over time while idling, exploring and grooming increased (Fig. 2e). A comparison of within-day habituation curves for each of eight behavioral classes over five days of recording revealed this trend across all behaviors, with females performing active behaviors more than males for the entire duration of each experiment (Additional file 4: Fig. S4).

Cntnap2 KO mice were unaltered in their ability to habituate within-day and their behavior across days was not significantly different from their littermates (two-way mixed ANOVA idle: \(F(2,36) = 2.96\), \(p = 0.07\); slow explore: \(F(2,36) = 1.92\), \(p = 0.16\); fast explore: \(F(2,36) = 1.56\), \(p = 0.22\); rear: \(F(2,36) = 0.09\), \(p = 0.92\); climb: \(F(2,36) = 2.24\), \(p = 0.12\)), although they did groom significantly more often on day 2 compared to WT littermates (one-way ANOVA, \(F(2,36) = 6.07\), post hoc \(p = 0.02\)) (Fig. 3 and Additional file 5: Fig. S5). In contrast to Cntnap2 KO mice, L7-Tsc1 mutant mice showed reduced within-day and across-day habituation (Fig. 3 and Additional file 6: Fig. S6).

Fig. 3

Behavioral usage over time for Cntnap2 KO and L7-Tsc1 mutant mice. a, b Behavioral usage for each of eight coarse categories plotted for a Cntnap2 KO and b L7-Tsc1 mutant mice for each of four observation days. All individuals are shown as points, colored traces correspond to the median fraction of time spent in the behavior for each day. c The usage of turning and locomotion behaviors during the 20-min observation period for Days 1 through 4 for Cntnap2 knockouts and littermates (left) and L7-Tsc1 mutants and littermates (right). Colors indicate the strain (blue—WT, green—heterozygote, orange—homozygote). Shaded regions represent the 95% confidence interval

Fig. 4

Habituation of grooming behaviors a The usage of grooming behaviors during the 20-min observation period for Days 1 through 4 for Cntnap2 and L7-Tsc1 mice. Colors indicate the strain (blue—WT, green—heterozygote, orange—homozygote). Shaded regions represent the 95% confidence interval. b Stacked bar plots showing the mean frequency across mice for each of five grooming behaviors across Days 1 through 4 with corresponding stacked bar plots below showing the normalized mean frequency for each of five grooming behaviors (c)

The defect in L7-Tsc1 mutant habituation was most apparent in locomotion and turning behaviors (Fig. 3 and Additional file 6: Fig. S6). These mice did not show the expected reduction in turning either over days or within the 20-min observation period. Turning was used more often on day 1 (one-way ANOVA, \(F(2,40) = 5.62\), post hoc \(p = 0.03\)) and did not decrease to the same degree over the observation time as WT and heterozygote littermates. Locomotion decreased over time but to a lesser degree and the level of locomotion in the L7-Tsc1 mutant mice was significantly higher than in WT littermates for days 1, 2 and 3 (one-way ANOVAs, groups-by-day, all post hoc \(p < 0.05\)).

Grooming behaviors vary in a complex manner in neurodevelopmental mouse models

Grooming refers to a variety of repetitive self-touching behaviors, and mouse self-grooming has been used as an animal model for the self-stimulating behaviors observed in autism [40]. Previous reports have limited quantification to 10 min of observation and have not distinguished different types of self-grooming or reported sex differences [29, 41, 42]. In our four-day recording period, we observed that all mouse groups groomed more frequently as the days passed (Figs. 2e, 3a, b, 4, Additional file 4: Fig. S4a, Additional file 5: Fig. S5a and Additional file 6: Fig. S6a), with larger increases in males compared with females (males \(d = 1.6\) and females \(d = 0.6\) day 1 compared to day 4, repeated-measures ANOVA, \(F(4,236) = 32.61\) and \(F(4,72) = 9.60\), post hoc \(p < 0.01\)). We furthermore found that within-model contrasts (male v. female, Cntnap2 KO v. Cntnap2 WT, L7-Tsc1 mutant v. L7-Tsc1 WT) were modest on day 1 but grew considerably over the next three days. The two neurodevelopmental mouse models showed opposite trends: Cntnap2 KO mice starting from day 2 groomed more frequently than their littermates (one-way ANOVA day 1: \(F(2,36) = 0.07\), \(p = 0.93\); day 2: \(F(2,36) = 6.07\), \(p = 0.02\)), whereas L7-Tsc1 mutants mice groomed much less frequently than littermates cross all days (one-way ANOVA day 1: \(F(2,40) = 5.28\), \(p = 0.04\); day 4: \(F(2,40) = 5.27\), \(p = 0.04\)) (Fig. 4).

The grooming class could be further divided into several modes of grooming: body grooming, face/paw grooming/licking, small quick movements, foot scratching and a mixture mode to account for time spent readjusting between modes. To validate grooming modes, human annotators were presented with movie segments that included grooming based on the algorithm and then were asked to annotate videos frame by frame (Additional file 3: Fig. S3d). A comparison indicated that the algorithm emphasized shorter timescales and identified more mixed bouts of grooming than human annotators (Additional file 3: Fig. S3e). The total number of episodes spent grooming was normalized to 1 to obtain a relative usage of these grooming modes (Fig. 4b). Despite differences in total grooming frequency between C57BL/6J males and females, their relative grooming-mode frequencies were similar and showed the same trend over time. Across days, body grooming increased in relative frequency (repeated-measures ANOVA, \(F(1,77) = 8.62\), \(p = 0.02\)) while face grooming and quick grooming movements became less frequent (repeated-measures ANOVA, \(F(1,77) = 8.74\), \(p = 0.02\) and \(F(1,77) = 6.96\), \(p = 0.048\)).

Both neurodevelopmental mouse models showed altered distributions of the grooming behaviors. Cntnap2 KO mice exhibited more body grooming compared to their WT littermates on day 2 (one-way ANOVA \(F(2,36) = 7.11\), post hoc \(p = 0.02\)). L7-Tsc1 mutant mice showed the opposite effect with less body grooming on all days of observation (one-way ANOVA, \(F(2,40) = 5.53\), each day, post hoc \(p < 0.05\)) and more face/paw grooming on days 3 (one-way ANOVA \(F(2,40) = 7.59\), post hoc \(p = 0.002\)) than WT littermates (Fig. 4c).

In all groups except for L7-Tsc1 mutants, grooming frequency also increased dramatically within each day’s observation period. Cntnap2 KO mice and their littermates groomed with similar frequency at the start of each day’s observation, but knockouts increased the time spent in grooming at a greater rate during each 20-min period (Fig. 4a). L7-Tsc1 WT and heterozygote mice likewise groomed more within each day’s observation period, but L7-Tsc1 mutant mice did not. These measurements show that systematic patterns of variation in grooming within each day of observation typically exceeded the variation across days.

L7-Tsc1 mutant and Cntnap2 KO mice show travel differences and gait defects

All experimental conditions showed reduced amounts of locomotion after day 1. However, the distribution of locomotion speeds depended on condition. In C57BL/6J mice, female mice moved on average 11% faster (\(d = 1.21\), one-way ANOVA \(F(7,154) = 33.47\), \(p < 0.001\)) than their male counterparts. Female C57BL/6J mice more often locomoted at speeds above 0.3 m/s than males (Fig. 5a). Cntnap2 KO mice did not differ from their littermates in the median velocity (\(p = 0.11\)) or distance traveled (\(p = 0.12\)) (Additional file 2: Fig. S2), although they did locomote at speeds above 0.3 m/s significantly more often on day 1 (one-way ANOVA, \(F(2,36) = 11.58\), post hoc \(p < 0.01\)). In contrast to all other experimental groups, L7-Tsc1 mutant mice kept their median velocity across days (repeated-measures ANOVA \(F(3,24) = 1.26\), \(p = 0.31\)) compared to WT littermates (\(F(3,48) = 9.46\), \(p < 0.001\)). The distance that L7-Tsc1 mutants traveled on days 1 to 4 was not significantly different from their WT littermates (one-way ANOVA, group-by-days, post hoc \(p = 0.60\), \(p = 0.33\), \(p = 0.66\), \(p = 0.13\)). They did so by moving more often (one-way ANOVA, group-by-day 1, post hoc \(p = 0.02\)) but at 19% slower median velocity (\(d = 1.55\), one-way ANOVA, group-by-day 1, post hoc \(p = 0.001\)) (Fig. 2c and Additional file 2: Fig. S2a): L7-Tsc1 mutant mice significantly increased the number of locomotion bouts at 0 - 0.2 m/s (one-way ANOVA, group-by-day, each day post hoc \(p < 0.001\)), a difference that persisted across all four days of testing (Fig. 5a).

Fig. 5

Locomotion kinematics are altered in the neurodevelopmental mouse models. a Stacked bar plots showing the mean frequency for each of four speed bins on Days 1 through 4. b Left: Examples of the motion of the limb and tail points during 5-s bouts of locomotion for C57BL/6J, Cntnap2 KO and L7-Tsc1 mutants. The trajectory of the centroid of the mouse is color coded by time. Time series of the position of each body part projected onto the anterior–posterior axis is plotted. The start of each gait cycle is marked with a dashed black line and defined using the front right paw (FR, red). Center: An example image of a mouse is labeled with colors used for each body part, and the axis of body alignment used when segmenting strides is shown as the axis between the nose and tail base point (TP). Right: Polar plots of the phase \(\theta\) of the gait cycle in which each paw reaches a minimum along the alignment axis in the body frame. The front right paw is used to define \(\theta =0\). The radial axis indicates the speed of locomotion, which was used to bin the phase results, and the size of the circles indicate how frequent that particular speed was for a given condition. For Cntnap2 KO and L7-Tsc1 mutants, the colors of dots correspond to the four labeled paw points

Nearly all children with autism show motor impairment [43, 44], with gait following a locomotor pattern resembling cerebellar ataxia [45, 46]. To explore the kinematics of gait, we identified all locomotion bouts (Figs. 1g, 5b) and analyzed individual body part trajectories along the anterior–posterior axis. As expected, the front and hind paws moved in an oscillating pattern in the frame of reference of the mouse (Fig. 5b) [47]. We also found that the tail often oscillated as well, matching the frequency of the paw oscillation.

Mouse locomotion could be broken into two different gaits depending on the speed. At very slow speeds (\(v<0.1\) m/s), C57BL/6J mice used a walking gait with paws moving in sequence, each for approximately one quarter of the walking cycle (Fig. 5b). The order of movement and coordination in the walking gait was Front Right, Hind Left, Front Left, Hind Right. At higher speeds, mice transitioned to a trot gait with alternating diagonal pairs of legs moving together. Left–right pairs of legs were not in phase with each other within the gait cycle. We found that for speeds above \(\sim 0.1\) m/s, C57BL/6J mice had an average phase difference of \(\sim 0.5\) rad or about 8% of the gait cycle. This phase offset decreased with speed from about \(\sim 1\) rad at \(v=0.1\) m/s to \(\sim 0.25\) rad at \(v=0.4\) m/s.

The two types of neurodevelopmental mutants displayed differences in both the walk-to-trot transition speed and the phase difference between opposite-limb (FL/HR and FR/HL pairs) movements. Cntnap2 KO mice showed a similar transition speed of about 0.1 m/s as their littermates. In contrast, L7-Tsc1 mutants used the walking gait much more and transitioned to the trotting gait at a much higher velocity of \(\sim 0.2\) m/s. The phase difference during trotting was larger for both neurodevelopmental mouse models. Cntnap2 KO had a phase difference at 0.2 m/s of 0.7 rad compared to a difference of 0.5 rad for their littermates. L7-Tsc1 mutants had an even larger difference of 1.0 rad at 0.2 m/s compared to 0.5 rad for their littermates.