Investigating social and independent behavior structure in early life is critical for understanding development and brain maturation in social mammals. However, this investigation necessitates monitoring animals over weeks to months often with subsecond time resolution creating challenges for both lab studies focused on brief observation periods and field studies in which animal tracking can be imprecise. Here we used machine vision and two-week long continuous behavior recordings of families of gerbils, a highly social rodent, in large, undisturbed home environments to quantify the behavioral development of individual pups. We discovered that individual pups exhibited complex social behaviors from the first day they left the nest including a preference for interactions with siblings over parents. Critically, independent behaviors such as foraging for food and water emerged several days later, each with a stereotyped temporal trajectory. Analysis of individual animal development confirmed the quality of our tracking methods and the stability and distinctness of each behavioral measure. Our work supports a model in which early and sustained social interactions may be supportive of solitary exploration for physiological needs. This model suggests that understanding the development of behavioral independence as well as maturation of sensory and motor systems in social rodents such as gerbils may require integration of social behavioral knowledge earlier than typically considered.

Tracking behaviors of individual gerbil family members in home-cage environments

The emergence of autonomous behaviors

Using the processing pipelines and heuristics described above, we focused on tracking the development of autonomy, i.e., the development of capacities to regulate and sustain oneself independently of other animals’ behaviors, including the parents.

Structured social behaviors precede the development of autonomous behaviors

We next sought to characterize pair-wise and group social behaviors to capture dependencies between individual family members (Fig 3). We defined social behaviors as those that involved another animal such as time spent huddling in the nest or approaches between two animals. As we were interested in the maturation of behavior, we again evaluated behaviors relative to the average adult behavior recorded when pups were P29 (as in Fig 2).

The development of multi-animal social state dynamics

We next evaluated the development of multi-animal social behaviors by focusing on the complexity of social states and configurations in the home cage. We were particularly interested in quantifying the complexity of multi-animal social configurations in their home environment. We partitioned the environment floor into a 5 by 4 grid and defined a social state at each recording time point as the number of animals at each location (Fig 4a, 4b) (see also “Methods”). This procedure enabled us to generate a network-based analysis where each time point in our recording generated a 20 Dimensional (20D) social state vector and we could generate a network connectivity matrix by connecting all social states that occurred sequentially (Fig 4c; see also “Methods”). The network connectivity matrix enabled us to generate graphs that represented social state dynamics in our cohorts (Fig 4d).

The development of circadian rhythm based behaviors

We next evaluated the correlation between the animal behaviors and the day:night cycles. In particular, we sought to quantify significant changes at the transition between light cycles (Fig 5a). For these analyses, all values are presented separately for both pups and adults as absolute values.

Pup behaviors develop along distinct time courses.

Across both autonomous and social behaviors, we observed stereotyped trajectories in development. Furthermore, the time courses tracking the development of single behaviors for pups from different cohorts were more similar to each other, than when compared with the time courses from the same pup but for different behaviors (Figs 2c and 3b, 3d, 3f, 3h, 3i). This pattern suggests that these classes of behaviors emerge on distinct timelines rather than being driven by random and/or individual processes. For example, foraging-based metrics such as time spent exploring and distance traveled per day each displayed a gradual increase between P16 and P26, at which time pups reached adult-like levels. In contrast, food proximity behavior displayed two phases, increasing gradually at first, but escalating dramatically from P22–P26, and reached 200% of the adult average. Similarly, water proximity behavior emerged slowly at first, but increased substantially on P24 and reached approximately 80% the adult average by P26. Taken together, these results suggest a slow and prolonged development of exploration behaviors with a rapid value-driven emergence of food and water exploratory behaviors around P22–P24. In fact, the rapid increase in pup food and water exploration coincides with a decline in adult female time spent at the food hopper (26% decrease after P24) and waterspout (41% decrease after P24).

Pup open exploration time increased from <1 hr at P16, to >5 hrs by P25. However, pup social behaviors predominated as they first began to leave the nest. As compared to adults, pups preferentially socialized in groups of 2 or 3, and preferentially associated with one another, rather than a parent, from P16–17 onwards (Fig 3d, 3h, 3i). For example, pups spent up to 7 times more time in groups of 3 than did their parents in the first few days after leaving the nest. This observation leads us to speculate that the early bias towards social interactions may prepare pups for later solitary activity.

Early social interactions support the transition to independence.

Juvenile and adolescent animals often display a rich array of social interactions with their peers, and this may serve as the substrate for the emergence of adult skills [36–38]. With few exceptions (e.g., [39,40]), the long-term effects of experimentally manipulating early social experience have been carried out in rodents. For example, depriving pre-adolescent rats of social play for two weeks leads to decreased social activity in adulthood [41, 42]. Furthermore, juvenile play during a developmental critical period can induce a preference for the environment in which the social interaction occurred [42–45], and this effect is diminished when pups are treated with a drug that interferes with social play [46]. Our descriptive data is broadly consistent with a constructive role of early social experience. When gerbil pups begin to leave the nest, they preferentially retain sibling interactions (Fig 3d, 3h, and 3i), which then leads to a dramatic increase in open exploration time (Fig 2b).

Behavior development is plausibly supported by the maturation of spatial-cognition and motor systems.

A systematic finding in our study is that all social and foraging behaviors commence at near zero levels relative to adults at P16, and mature with distinct time courses between P16 to P26. This period coincides with the maturation of neural pathways that subserve spatial memory, especially in the hippocampal and medial-entorhinal-cortex axis (HPC-MEC). In mice, there are several systems that develop during this period: formation and maturation of head direction cells (P10–P20), place cells (P14–P40), boundary responsive cells (P16–17) and grid cells (P20–P21) (Tan and colleagues [47]). In rats, hippocampus dependent learning has also been shown to commence around P18 for water maze learning (e.g., visual platform – P17–20 – and hidden platform – P19–21 – learning), and T-maze learning (P19–P42) [14]. In fact, social encoding of family membership may, itself, reside in the hippocampus [48]. The behaviors that we observed are not closely correlated with the onset of sensory transduction. The onset of transduction displays a similar sequence in most vertebrates, progressing from somatosensory to vestibular, olfactory, auditory and, finally, to the visual system [49]. However, a prolonged development of central sensory processing has been demonstrated for many mammals [50]. For example, as compared to adults, gerbil auditory perceptual skills and auditory cortex processing remain immature during early adolescence [51,52]. Therefore, one plausible interpretation is that the highly dynamic and specific periods of development that progress from a minimal state to near-adult levels between P16 to P29 are driven, in part, by maturation of HPC-MEC axis and central processing of sensory information.

A second plausible interpretation is that the changes and stabilization we observe arise as the motor cortex matures. In particular, the motor cortex is still relatively immature (e.g., as compared to sensory cortices), with an estimate that M1 starts producing a significant outflow only after postnatal day 20 [13]. Cortical motor outflow appears to begin at approximately P24 in rats [53–55], and there is a significant increase in stimulation-related motor activity after this age. For example, M1 activity recorded in vivo in P19–23 rats has yet to exhibit activity before wake or sleep movements [56]. Therefore, an increase in exploration may also be tied to the maturation of the motor cortical pathway.

While the maturation of the HPC-MEC axis, sensory, and motor cortices may contribute to behavioral changes during P16 to P29, it will be essential to consider other regions involved in development and threat-related behaviors. The BLA (basolateral amygdala), septum, and PAG (periaqueductal gray) play crucial roles in exploration and avoidance of threats during this time [57]. Additionally, regions involved in motivation, such as the nucleus accumbens (NAc), likely influence behavior, especially in relation to hunger, growth spurts, and pubertal changes, which may affect exploratory behaviors [58]. It is also worth noting that non-rod or cone-mediated phototaxis in young rodents [59], may wane during development, potentially influencing behavioral responses to environmental visual stimuli.

Circadian rhythms in adults.

Our paradigm enabled us to track the circadian features of adult behaviors and their emergence in pups. Studies on gerbil circadian rhythms indicate that light conditions can affect behavior [35,60,61], with a crepuscular pattern for most behaviors except drinking [35]. We found adults were active during both light and dark cycles, with more behavioral activity during the day time and a substantial increase of activity prior to dark cycle onset. This pattern is reminiscent of crepuscular behaviors in field studies, although not identical. Two significant differences with field studies is that our lab paradigm contained no risks, such as the presence of predators, and offered ad libitum food and water resources. We also observed diurnal behavior patterns only in the amount of distance traveled with other behaviors showing weak or no crepuscularity. We note that increased activity during the day is consistent with findings of greater maternal attention to pups during the day than at night [18,19].

Development of a diurnal pattern in pups.

Beginning at approximately P26, pups began to display diurnal patterns in all behaviors with peaks at night cycle onset. However, in contrast to adults, these older pups also displayed substantial diurnal changes in activity as light-change switches approached across most behaviors (note: pup behavior was weak in the pup–pup approach behavior – similar to adults). This suggests that pups, unlike adults, had stronger drives to increase their activity before light onset than adults. Considering all periods of development, however, we found that some behaviors such as open exploration time and pup–pup time spent together changed much less between early and late development as compared to food hopper region and waterspout proximity. We sought to use the heterogeneity in food hopper proximity in particular – as it showed the largest differences over time – to agnostically detect different developmental stages for circadian rhythms, and identified P16–19, P20–27, and P28–29 as the most likely partitioning of development into three stages.

Limitations of our work and directions for future studies.

Our enriched and enlarged home cage environment enabled the tracking of several types of behaviors. However, although the cage area was approximately 5–6 times the size of animal facility cages, it was significantly smaller than gerbils’ natural habitat, and lacked an underground burrow. In particular, providing access to an underground compartment during development permits the maturation of normal escape behavior [62], suggesting that foraging area size alone is insufficient for the emergence of a full behavioral repertoire. Therefore, developmental trajectories may differ with orders of magnitude larger areas to explore. Additionally, we did not carry out highly specific behavior classifications such as grooming, fighting, playing, and sexual interactions, which would require significantly more complex machine vision tools than those developed in this study. We also substituted one day of our study due to cage cleaning with the average behaviors on neighboring days, and this may have affected some of our metrics which relied on stable or rapidly increasing behaviors over days. Our environment did not capture other behaviors observed in the wild such as gerbils’ increased tendency for exploration towards novelty [63] and towards “open-arms” of mazes (compared to rats and mice; [64]). Gerbils also show increased unpredictability in their escape behaviors as compared to other rodents [65], and it would be interesting to observe this escape behavior dynamics by simulating predators within our environment. Beyond home cage environments, we propose that contemporary machine learning methods are improving substantially and it may be possible to carry out similar studies in “limited size” environments such as scientifically-purposed barns [66,67].

Understanding development from the micro to the macroscale.

Our work builds on a series of technically and computationally advanced paradigms that quantify behavior over days to months. A small group of adult mice have been monitored continuously for 3–4 days in 50–70 cm² chambers using video recordings, permitting the acquisition of multiple behaviors that can be modeled to reveal group dynamics [68–70]. In contrast, groups of 10–40 mice have been monitored continuously for up to 3 months within 1–2 m² environments using radio frequency identification transponders, thereby revealing the emergence of individuality [71,72].

Similar to these studies, our approach supports the viability of studying rodent social behavior across orders of magnitudes of temporal resolution during highly dynamic periods of behavior and development in family groups of rodents. Our findings reveal novel time courses for development in rodents, including dynamic periods of behavior acquisition and the development of circadianity. We view that increasingly ethological (e.g., group and family housed animals) and naturalistic (e.g., larger home enclosures with increased enrichment opportunities) behavior paradigms in the laboratory environment is an adequate and complementary method for the characterization of behavior and CNS development in health and disease.

Experimental animals

Three gerbil families (Meriones unguiculatus, n = 6 per family: 2 adults, 4 pups) were used in this study (Charles River). All procedures related to the maintenance and use of animals were approved by the University Animal Welfare Committee at New York University, and all experiments were performed in accordance with the relevant guidelines and regulations (protocol 2020-1112).

Audio recording

Four ultrasonic microphones (Avisoft CM16/CMPA48AAF-5V) were synchronously recorded using a National Instruments multifunction data acquisition device (PCI-6143) via BNC connection with a National Instruments terminal block (BNC-2110). The recording was controlled with custom python scripts using the NI-DAQmx library (https://github.com/ni/nidaqmx-python) which wrote samples to disk at a 125 kHz sampling rate. In total, 13.084 TB of raw audio data were acquired across the three families. For further analyses, the four-channel microphone signals were averaged to create a single-channel high-fidelity audio signal.

Home cage enclosure

We used an enlarged home enclosure (W × L × H: 35.6 cm × 57.2 cm × 38.1 cm) with ad lib access to a food hopper and a water source.

Video recording

We recorded video using an overhead camera, FLIR USB blackfly S (FLIR USA), at 25 frames per second. We recorded individual periods of approximately 20 min (cohort 1) or 60 min (cohorts 2 and 3) following which video and metadata were saved for offline use. The saving period resulted in approximately 10% downtime between recordings.

Night:day cycle

Daytime began at 08:00 and night began at 20:00 during which an infrared light was activated.

Animal shaving patterns

We used a distinct shaving pattern across all our cohorts (as described in Fig 1b). We opted for use of shaving patterns instead of fur bleaching, implanted trackers or body piercing as we viewed shaving as a less intrusive method.

Tracking—Features

We manually labeled 1,000 frames using the SLEAP (Pereira and colleagues [30]) labeling workflow. The skeleton consisted of 6 nodes: nose, spine1, spine2, spine3, spine4, and spine5; and 5 edges: nose to spine1, spine1 to spine2, spine2 to spine3, spine3 to spine4, and spine4 to spine5. Additionally, we labeled each instance with the animal’s identity class as female, male, pup1, pup2, pup3, and pup4 to enable tracking unique identities. Due to idiosyncratic differences in shaving patterns, lens distortion and light conditions, each cohort and light condition (i.e., day and night) was labeled and trained individually, resulting in 6 feature models. The final results were trained using SLEAP bottom-up ID models that achieved 97% accuracy of adults and 91% for pups.

Track clean up—features

We developed an ID-switch error detection method to detect and correct some of the SLEAP ID-switch errors. We defined ID switch errors as those for which track segments were swapped between two animals. These errors were identified by evaluating sequential segments of tracks produced by SLEAP and determining whether (i) a large jump in track location occurred; and (ii) whether another segment fit better (usually within <5 pixels of distance). We additionally implemented an interpolation method that connected SLEAP segments <3 cm apart that belonged to the same animal. A custom python pipeline was developed to carry out this correction.

Tracking—Huddles

We manually labeled 1,000 frames using the SLEAP (Pereira and colleagues, [30]) labeling workflow. The skeleton consisted of 1 node: huddle; and no edges. We identified a huddle as a group of 2 or more animals grouped together in a stable location, which were usually places that animals huddled in for the entire recording period or several sequential days. Due to idiosyncratic differences in shaving patterns, lens distortion and light conditions, each cohort and light condition (i.e., day and night) was labeled and trained individually, resulting in 6 huddle models. We additionally implemented an interpolation method that connected SLEAP segments for huddles that were less than 30 s apart.

Bedding changes and exclusion of time points

We provided an interpolated measure of behavior on days of bedding changes. Briefly, on cohorts 1 and 2 we averaged the behaviors between P23 and P25 instead of using P24. On cohort 3 on P18 we averaged the behaviors between P17 and P19 instead of using the P18 day as we observed a substantial difference in the behavior just for that day.

Cross-validation of behavior tracking

We implemented 10-fold cross-validation to evaluate the precision of SLEAP identity tracking against a human annotator. Briefly, for all feature NNs we split our human labeled frames dataset (1,000 labeled frames) into 950 – training and tested on the hold out of 50 frames. We repeated this 10 times by randomizing the training set frames and calculated the average identity accuracy errors across adults and pups.

Convex hull overlap metric

We defined overlap volume between behaviors as the intersection between the convex hull generated by individuals in each type of behavior relative to the other behaviors. We calculated the overlap using principal component analysis (PCA) as the intersectional volume of the first 3 principal components (using the python vedo package) or – when this was not possible due to insufficient points to form a polyhedron – the first 2 principal components into the PCA. The data input into PCA was [n_animals, n_time_points] = [12,15] where the values were normalized to adult levels. We viewed the overlap volume as better at capturing and representing the inter-behavior similarity by providing a metric that quantifies when some of the individual behaviors were not distinguishable between behavior types. We note that comparing the individual behaviors at each time point was not practical due to only 3 samples in the behavior of the adult female and male and male pup. We also note that we did not use more than 3PCs (or the full dimensionality of the distributions) to compute these similarity metrics as algorithms to compute multidimensional overlapping volumes are not available. Furthermore, the dimensionality in our study would likely yield vanishingly small overlaps.

Rapid development and stabilization periods

We defined a rapid development period as days that had distributions that were significantly different than the previous 2 days. We indicated these periods with triangles in the plots in Figs 2 and 3. We thus pooled behaviors in groups of 2 days and compared them following groups. We used two days as a more conservative way to account for higher variability – than applying smoothing or other filtering techniques on the raw data. We used a 2 sample Kolmogorov-Smirnov test as a more conservative measure of significance. For clarity, at each day, we pooled data from individuals (e.g., 12 pups or 6 adults) and calculated the statistics across these values. We used the python package statsmodels method sm.stats.multipletests for the Benjamini–Hochberg multiple hypothesis testing adjustment and reported p values < 0.05 as statistically significant.

Heuristics-based behavior classification

We implemented heuristics to compute the location of animals near the food hopper or waterspout and each other. For the food hopper we used an ROI of size 19.50 × 10.50 cm. For waterspouts we used an ROI of size 8.50 × 8.00 cm. For inter-animal proximity we used a distance of <5 cm separation and a minimal proximity time >200 ms. Animal locations were calculated as the median centroid over (x, y) locations. For control and comparison to ROIs in food and waterspout regions we also selected ROIs in random locations of the cage. We found that ROIs that were randomly placed had little occupancy (limiting statistical analysis) and did not have the dynamics of ROIs in the food and waterspout regions.

Computation of group socialization time

We calculated the amount of time pups and adults were socializing as the time they spent within 5 cm of each other. This method was used for both pups and adults, for example, to compute how much time each age group spent near groups of 2 or 3 other animals (i.e., <5 cm from nearby gerbils). The relative time computation was calculated similar to other steps, namely, by computing the socialization time for adults near other gerbils including other adults and pups and using this as a baseline.

SimBA-based behavior classification

We defined an approach as an animal moving towards another – beginning when the initiator starts moving and ending when it stops moving (at most one gerbil length away from the other) or makes physical contact. We manually annotated 300 approaches (4,578 frames) in 22 videos and implemented a modified version of SimBA [31] for classifying pair-wise animal approach behaviors. We calculated the threshold for approach behavior as the 99% percentile of the prediction distribution. We verified that detected approaches matched with qualitative approaches.

Descriptive statistics of animal movements, angles, and distances in sliding time-windows were calculated using runtime optimized methods available through the SimBA API and used in a downstream random forest classifier. We randomly sampled annotated frames from separate events in training and test set to reduce time-series data leakage. Within the training set, we further balanced the data by under-sampling the number of non-event observations to a 1:1 ratio of the event observations. Classifiers were validated and optimal discrimination threshold determined by evaluating performances on novel held-out videos excluded from the training and test sets.

Multi-animal social state networks

We developed an approach to quantify unique social configuration states and their dynamical transitions. We first discretized the cage space into 5 × 4 squares of approximately 11 cm × 9 cm each – yielding a 20 Dimensional state space. At each time point in our recordings we calculated the occupancy of each location as the number of animals at that location. This yielded an N_videoframes × 20D matrix which in general had a few hundred or thousand unique 20D vectors per day. Each 20D vector thus contained integers from 0 (no animal in the location) up to 6 (i.e., all animals were at the grid location). Connectivity matrices for each day were then calculated by connecting each of the videoframe nodes to the one occurring subsequently. We used the connectivity matrices to generate 14 networks for each day of our study and for a variety of animal configurations. For clarity, we calculated 5 types of networks: single pups (i.e., 1 animal only), pairs of animals (the 2 adults and all pairwise combinations of pups), all-pups (i.e., all 4 pups) and all-animals (all 6 animals). (Fig 5a, 5b; see also “Methods”). This enabled us to generate a network-based analysis where each time point in our recording generated a 20 Dimensional (20D) social state vector and we could generate a network connectivity matrix by connecting all social states that occurred sequentially (Fig 5c; see also “Methods”). The network connectivity matrix enabled us to generate graphs that represented the social state dynamics at each day in our cohorts (Fig 5d).

Automatic detection of circadian rhythms development.

We developed a computational method to evaluate whether pups had more discrete transitions from non-circadian to circadian rhythms-driven behaviors. The method was applied to food hopper region proximity data – and an assumption of three developmental periods based on qualitative observations in our study defined as: an initial period D1, a middle period D2 and a final period D3. Based on our qualitative observations we initialized these periods as: P16–P20, P21–P25 and P26–P29. We next sought to determine whether the boundaries on these periods, i.e., t1 and t2 corresponding to P20 and P25, respectively, could be quantitatively defined via an optimization function. The optimization function sought to maximize the difference between the individual animal circadian trajectories in each of these periods. Our algorithm identified t1 = 19 and t2 = 27 as the dates that create the maximum differences between pup presence in the food hopper region, i.e., D1: P16–P19, D2: P20–P27, and D3: P28–P29. (Fig 5h). We noted that different behaviors or priors, e.g., a single transition may provide slightly different developmental windows.

Cohort and animal group sizes power analysis

To achieve a power of at least 80% with a significance value of 0.05 we estimated the required sample sizes for our work are 3 individuals in each group of pups and adults – whereas we had 12 and 6, respectively. Briefly, we estimated these values based on existing literature that shows P16 gerbil pups having little to no movement (they are still in the nest for most of the time) and a small variance of 10% relative to the adult (which was our baseline) and the adult having approximate variance of 20%. Thus, for P16 and early development, 3 individuals in each group were sufficient, with 12 and 6 animals in each group providing power of approximately 90% or better for both 0.05 and 0.01 significance. As we trained our neural networks to identify pups and adults across all days of our study (P16 to P29) our cross-validation analysis applies to all age groups – not just early development – and we found strong evidence that the machine-vision tools generalize sufficiently to keep the same animal identity over the entire window of development in our study. In sum, our data and analysis support that groups of approximately 3 individuals each would provide sufficient power for this study.

Computation of standard deviation

Standard deviation reported in all results in brackets is for all pups, i.e., n = 12. The values for each pup was obtained by averaging the behavior of the 2 adults within each cohort.

S1 Fig. 10-fold cross-validation evaluation of accuracy of animal tracking algorithms.

We labeled 1,000 frames per cohort under both day and night conditions (total 6,000 human labeled frames across three cohorts), and studied the accuracy of the animal tracking algorithm (SLEAP) as a function of the number of frames (x-axis) trained. We performed this analysis 10 times per training set size by shuffling the human labeled datasets (except the 1,000 – dataset where we held out only 50 frames and trained on 950). Accuracy of unique animal identification improves asymptotically with the number of human labeled frames in the dataset.

https://doi.org/10.1371/journal.pbio.3003348.s001

(TIF)

S2 Fig. Real versus identity shuffled development trajectories.

To visualize the uniqueness of the developmental trajectories, we plotted water (a) and food (b) consumption behaviors when pup versus adult identities for correct or “real” identity tracking versus identity shuffled condition. Shuffling identities results in behaviors appearing closer to adults and with higher variance (shading is SEM).

https://doi.org/10.1371/journal.pbio.3003348.s002

(TIF)

S3 Fig. Shuffling identity decreases behavior trajectory uniqueness.

Similarity of intra- versus inter-behavior trajectories over time visualized with principal component analysis (PCA). The individual developmental trajectories (14 Days = 14 Dimensional) for all pups (n = 12) is visualized using the first two PCs of these datasets. (a) When taking into account unique identities developmental trajectories (points) and convex hull (enclosed areas) for water and food proximity in PCA space show significant separation. (b) When the analysis shown in panel (a) is repeated, but with unique animal identity shuffled, substantial overlap is observed.

https://doi.org/10.1371/journal.pbio.3003348.s003

(TIF)

S4 Fig. Correlation of real versus shuffled development trajectories.

Distributions for all pair-wise individual pup developmental trajectories over 14 days of development are shown for proximity to water and food sources. Within each behavior class, the correlations are much higher for real versus identity-shuffled data. Between behavior correlations, such as water versus food proximity, were lower than within behavior correlations, but significantly higher than those computed for shuffled conditions. We note that correlation between food and water development trajectories was relatively high because both behaviors are generated by the same causal factors, including the increased drive to seek independent sustenance, which unfold on a similar time scale.

https://doi.org/10.1371/journal.pbio.3003348.s004

(TIF)

S5 Fig. Inter-behavior pairwise correlations.

Comparison of food and water proximity to other behaviors yields positive or negative correlations. This result is due to the Pearson correlation between a coarse measure of similarity that relies on the relative value of points in a time series to the mean. As many of the behaviors increased or decreased monotonically in similar ways, we did not view these correlations as a sufficiently sensitive method of analysis in our study.

https://doi.org/10.1371/journal.pbio.3003348.s005

(TIF)

S6 Fig. Cohort similarities show partial within group idiosyncrasies.

We performed a reanalysis of the data plotted in Fig 2C, but with members of the same cohort displayed as different symbols. There is some grouping of gerbil pup behaviors by cohort number (we used a 2D PCA and a different projection angle to improve visualization). This partial clustering is indicated by the proximity of trajectories (i.e., individual points) for gerbils from similar cohorts (denoted by symbol). However, our primary conclusion is preserved: pups from the same or different cohorts still preserve distinct behavior trajectories (in this higher-dimensional space) when compared across behaviors – and the findings are not drive by single cohort idiosyncrasies.

https://doi.org/10.1371/journal.pbio.3003348.s006

(TIF)