Long-term tracking and quantifying individual behaviorWe tagged individuals with color markers and employed automated tracking to determine each rat’s movement over time (Fig. 1A–D). Over the course of the experiment, we performed manipulations to alter the group composition and the living area available for the group to use. We used two different breeding lines of Wistar laboratory rats, denoted A and B (see “Methods” section for details), with associated individual labels \(a*\) or \(\alpha *\) and \(b*\) or \(\beta *\), respectively (see Fig. 1E). We initially divided the rats into four groups of seven, with A rats in groups A1 and A2 and B rats in groups B1 and B2. The rats remained in these groups for the first observation period, which lasted a total of 21 weeks—we denote this time as phase 1. Following this, in phase 2, we merged groups A1–A2 and B1–B2 for three weeks and then merged all for three weeks by opening portals between their compartments. For the final series of group experiments in phase 3, rats from each original group were mixed together to create four new groups. The reshuffling in phase 3 was done according to body mass at the end of phase 2 (mean = 480 g; min = 364 g, max = 613 g; Q1 = 423 g, median = 481 g, Q3 = 532 g) , allocating rats to new groups by ensuring that each group had the full range of masses and included members from every previous group (\(G1_{min}\) = 394 g, \(G1_{max}\) = 541 g, \(G2_{min}\) = 372 g, \(G2_{max}\) = 587 g, \(G3_{min}\) = 400 g, \(G3_{max}\) = 613 g, \(G4_{min}\) = 364 g, \(G4_{max}\) = 555 g). Figure 1E shows the experimental structure and the associated measurement periods.We calculated automated trajectory-based behavioral metrics to quantify behavior over the duration of the experiment. We calculated and averaged each metric over successive time periods of 3 weeks (denoted as Pd), with associated numbers 1–12. We use the summary metrics to ask how behavior changes over time, how individuals differ, how groups differ, and how previous individual and group behavior predicts changes when new groups are formed.Fig. 1Experiment setup and timeline. (A) Photo of the rats with color-codes for individual identification and tracking. (B) Still image from the video that was used for tracking (from group G1, during Pd 10) taken by a light-sensitive camera at low lighting conditions. Image overlaid with labels indicating the important objects (water, nestbox, etc.). (C) Continuous tracking allowed for the reconstruction of each individual’s space use. The heatmap shows the space use of two rats during a 3-week period at the beginning of phase 3. Areas used only by a3 are shown with red, only by \(\beta 1\) with green, and areas visited by both (e.g. at the water and the feeder) are shown with yellow. (D) Trajectories were used to identify dominance interactions in the form of approach-avoidance events, where one individual approaches another, but the other moves away (by backing up or fleeing). Shown is an example of trajectories from group G3 in period 10. Lines show locations for 60 seconds, with the semitransparent circles of increasing size showing the more recent positions. (E) Overview of experimental manipulations. We calculate behavioral metrics over each 3-week “period” (abbreviated as Pd). Phase 1 had rats in original breeding line-sorted groups A1–A2 (line A), B1–B2 (line B), for a total of 7 periods. Each rat is labeled with lowercase letters a/\(\alpha\) or b/\(\beta\) according to breeding line. Individual numbers within each group are sorted in ascending order according to rank as determined by Elo score at the end of phase 1, i.e. a1/a7 were the highest/lowest ranking individuals in A1 during Pd 7, \(\alpha 1\)/\(\alpha 7\) were the highest/lowest in A2, etc. In phase 2, the groups were mixed together by breeding line during Pd 8, and then all together for Pd 9. At the beginning of phase 3 (Pd 10), new groups were formed (G1–4). During Pds 11 and 12 in phase 3, the compartment area sizes were changed (see “Methods” section and Fig. S8). At the end of the experiments, individual behavior was assessed by traditional individual and pairwise assays.To assess dominance-related interactions and social structure, we tabulated approach-avoidance events between all pairs of rodents in each group. This automated method defines “events” as when a pair of rats come close to each other: the “displacer”, i.e. the dominant rat in an event, subsequently stays in place or continues moving forward, while the other (the “displaced”, i.e. subordinate rat) move away6. This type of approach-avoidance interaction can also be dynamic, such as when one individual chases another. We use the matrix of approach-avoidance events to calculate metrics that describe the dominance structure of each group and each individual’s position in this structure.Breeding line and group differencesWe use automated measures of space use and pairwise interaction events to characterize individual and group behavior. We first examine general differences between breeding lines.In the beginning, the rats were juveniles and were growing rapidly, as shown by the large increases in body mass during this time period. The A rats were, on average, significantly larger than the B rats during each period (T-test comparing average mass of A rats to B rats yields p < 0.001 for each period). All rats had approach-avoidance events during the experiment, and there were no consistent significant differences among the breeding lines. However, there was an increase in the number of events per rat in phase 3 compared to phases 1 and 2 (Mean number of events per rat in phases 1, 2, 3, respectively: 447, 494, 976; T-test mean of phase 1 to phase 2, p = 0.64; mean of phase 1 to phase 3, p = 0.0044; mean of phase 2 to phase 3: p = 0.0176).The metrics of time at feeder, distance from wall, home range46, time at top of nestbox, and time on wheel describe space use. While the breeding lines did not have general differences in time at feeder or time on wheel, line A rats tended to be farther from the wall, visited more parts of the living compartment (larger home range), and spent less time on top of the nestbox in comparison with B rat groups. However, while these differences were clear during phase 1, the differences in distance from wall and home range decreased when the lines were mixed, with home range no longer significantly different from Pd 9 onward, and distance from wall no longer significant in Pd 12. Breeding line differences in time spent at top of nestbox showed a large increase when group membership was changed in Pds 9 and 10, but subsequently decreased and were not significant in Pds 11 and 12 (Fig. 2). Note that one group (G1) in phase 3 displayed a different pattern of wheel usage than other groups, with several rats spending a very large amount of time on the wheel at the same time and thus unable to use it for running (Figs. S3, S4); however, there were no breeding line differences in this behavior. Overall these metrics suggest that the different breeding lines differed in their space use tendencies, but differences decreased when rats were placed in mixed groups in phase 3.Fig. 2Breeding line comparison and correlation. (A) Per-line body mass, average number of events, and space use metrics. Significant differences between breeding lines for a designated period, as determined with a T-test for difference in means, are denoted as follows: p < 0.05 with *, p < 0.01 with ** and p < 0.001 with ***. See also Fig. S3 for space use compared according to group, and Fig. S4 for space use metrics for each individual rat. (B) Correlation with the previous period, calculated across all rats with respect to a particular metric. Shaded area shows confidence interval calculated via bootstrapping. Note that values significantly different from zero are when the confidence intervals do not contain zero. (C) Correlation with Pd 7 (the last measurement in phase 1). Shaded area shows confidence interval calculated via bootstrapping.We quantify changes in individual behavior using the correlation coefficient across periods. This shows that individuals have consistency in number of events and space use, as demonstrated by the generally positive correlations during the entire observation period (Fig. 2B). However, while there is consistency from one period to the next, Fig. 2C shows that small behavioral shifts over time can accumulate. Moreover, we see that the re-groupings facilitated changes in behavior. This is demonstrated by the sharper decrease in the correlation of behavioral metrics with Pd 7 in phases 2 and 3 compared to that in phase 1 preceding Pd 7. In particular, while the correlation coefficient for home range and time at top of nestbox during Pds 11 and 12 showed high correlations (Fig. 2B), the correlation of these 3 measurements with Pd 7 values was lower (Fig. 2C). For example, for Pd 12, the correlation with the previous period for home range was 0.77 (95% CI [0.68 0.87] and for top of nestbox was 0.89 (95% CI [0.71 0.95]), while the correlation values with Pd 7 were 0.26 (95% CI [− 0.13 0.57]) and − 0.17 (95% CI [− 0.5 0.16]), respectively. This indicates that the new behavioral routines of phase 3 differed from those of phase 1.Metrics for group social structuresWith the pairwise approach-avoidance interaction matrices for each period, we use multiple metrics to characterize different aspects of group social structure and an individual’s placement in this structure. The metrics to characterize individual social placement include Elo score, David’s score, local reaching centrality, and fraction of events dominated, and those to characterize group social structure include Elo score steepness, David’s score steepness, global reaching centrality, directional consistency index, and triangle transitivity index. In this section we use idealized networks (shown in Fig. 3) to illustrate what the group social structure metrics represent. Note that while other work has used similar idealized or artificial networks as “categories” to label group social structure47, here we use the ideal networks (including connected hierarchy, line, layered hierarchy, layered-half, non-transitive, single dominant, single out, and symmetric) not as categories, but rather to give intuition for how the different metrics describe different aspects of the social structure. In the following section, we report the metrics for each group and use them to describe the experimentally observed structures.The Elo score steepness (ESS) is a measure of the spread of the distribution of Elo scores across the group. It is calculated by converting the Elo score to a success probability, summing normalized values across group members, and calculating the slope of a linear regression fit to the resulting values45. The David’s score steepness (DSS) (often referred to simply as hierarchy ‘steepness’, or ‘classic steepness’42,45) is calculated as the slope of a linear regression fit to the normalized David’s scores among group members48. Individual local reaching centrality (LRC) uses the directed network of excess pairwise event outcomes (positive entries for rats in a pair that was dominant in more events, and zero for the other rat—see “Methods” section) in order to assign higher scores to individuals in higher positions within a group hierarchy. For an unweighted directed network, LRC is the fraction of nodes reachable by any given node; a generalization of the metric accounts for weighted connections35. Global reaching centrality (GRC) is the average difference of nodal LRC with that of the highest LRC of any node in the graph, and a higher GRC indicates a more hierarchical network35.The directional consistency index (DCI) is the fraction of events dominated by the more dominant individual of each pair, with 1 corresponding to perfect predictability in the outcome of a pairwise event (i.e. one individual is always dominant), and 0 representing an exchange of approach-avoidance outcomes (i.e. each individual dominates the same number of events)42,49. The triangle transitivity index (TTRI) is the fraction of triad relationships that show transitivity in pairwise event dominance outcomes (i.e. if \(a\rightarrow b\) and \(b \rightarrow c\), then \(a \rightarrow c\) for a transitive triad)43,50.From Fig. 3 we note that the ESS and DSS, which both aim to measure the steepness of hierarchy within a group, show similar trends at times and differ at others; both have high values for the connected hierarchy network but differ for the line network. We also note that the aspects of the network structure described by ESS/DSS versus GRC are different (c.f. differences in the connected hierarchy, layered-half, and single dominant networks); the former is maximized when a well-connected structure exists (i.e. the hierarchy shows a clear distribution that lends itself to a linear regression fit), while the latter is maximized when more extremes in hierarchical structures exist (for example, the single dominant). Although a comprehensive evaluation of these metrics is beyond the scope of this study (see, for example45), here we calculate and examine multiple metrics to ensure a robust interpretation of the data, as well as to facilitate comparison of our findings with other assessments of group social structure found in the literature.Fig. 3Idealized networks and group social metrics. The table at the top shows the scores calculated: Elo score steepness (ESS), David’s score steepness (DSS), global reaching centrality (GRC), directional consistency index (DCI), and triangle transitivity index (TTRI). Note that the GRC is not defined for the symmetric network, and the TTRI is not defined for networks that do not contain any dominance triads. The different idealized networks have 7 nodes, and individual entries are either 100 or 0 (for the layered-half network, 100, 1, and 0 are used). The connected hierarchy network has a non-symmetric structure. The line network has a single “line” of pairwise interactions, where each individual only interacts with one other. The layered hierarchy network has a single individual who dominates all others and two other sub-dominant individuals who only dominate the four others below them. The layered-half network has the same structure but lower values for the subordinate individuals. The non-transitive network has individual 1 dominating 2–3, 2–3 dominating 4–7, but 4–7 dominating 1.The single dominant network only has events with the dominant individual. The single-out network only has events with the subordinate individual. The symmetric network has equal event dominance probability among all pairs. In the table, the highest score for each metric (rows) is in bold, and the lowest score is in italics. The ESS is highest for the connected hierarchy and line networks, the DSS is highest for the connected hierarchy network while low for the line network, and the GRC is highest for the layered-half (i.e. structured but nonlinear hierarchy) and single dominant networks. For the symmetric network, the DCI is 0 because there are no consistent dominance relationships; for other networks, dominance is one-sided and the DCI is 1. The TTRI is 0 for the non-transitive network, and 1 for other networks where dominance triads are predicted. In addition to matrix plots, each network is visualized by showing connections in the direction of the more to less dominant individuals in each pair (note that no connections are shown for the symmetric network because, in this case, there are no differences in event dominance probability). We show both a circular layout and a layout based on Elo scores, where individual nodes with higher Elo scores are shown higher up on the y-axis.Group social structuresWe find that groups differ in their social structure, but within-group structure shows consistency when group membership remains unchanged. We compare phase 1 and phase 3 group social structures because the associated periods all featured groups of 7 rats. We show results for all group social structure metrics, as well as the mean number of events and fraction of events with the dominant rat, and note instances where the trends for the metrics are similar versus contrasting.In phase 1, in contrast to space use, which showed clear breeding line-based differences (Fig. 2), we do not see clear differences between lines A and B in terms of overall group social structure. While the fraction of events with the dominant rat was higher for the A groups in comparison to the B groups, other metrics do not show large or consistent differences (Fig. 4). Within phase 1, each group showed consistency in the social structure over time, with the exception of Pd 6 for group A2, where a large change in the individual rank ordering in the social structure of the group took place. This is seen in the network visualizations, as well as in the positive correlation of Elo scores from one period to the next (Fig. 5A,B).Fig. 4Measures of group social structure. The mean number of pairwise events and the fraction of total events with the most dominant rat are shown in addition to the hierarchy-related metrics of Elo score steepness (ESS), David’s score steepness (DSS), global reaching centrality (GRC), directional consistency index (DCI), and triangle transitivity index (TTRI). The fraction of events with the dominant rat (where “dominant” is defined as the individual with the highest Elo score) is analogous to the measure of “despotism” used in other work22; the dashed line shows the expected value if all pairs of rats have the same number of events. The metrics are calculated for each period and are shown as boxplots for each phase 1 and phase 3 group. See also Fig. S7 for values for each group over time.Fig. 5Social structure network visualizations. (A,C) A visualization of group networks during (A) the last 3 periods of phase 1, and (C) phase 3. Columns correspond to different groups and rows for each period. The position of each individual on the y-axis is set according to their Elo score. The direction of each connection indicates which individual dominated more events in the pair (e.g. a connection \(a4\rightarrow a6\) indicates that a4 more often displaced a6 than vice versa), the color indicates the fraction of events dominated, and the transparency is proportional to the total number of pairwise events relative to the mean for that group and period. (B,D) The correlation of individual Elo scores with the previous period, for (B) phase 1 groups, and (D) phase 3 groups. Dashed line shows baseline correlation value calculated by shuffling groups and periods during phase 1 (B) or phase 3 (D). See also Fig. S5 for individual metrics (including num. events, fraction dominated, Elo score, David’s score, and reaching centrality) for each individual rat plotted for each period.In general, we saw larger differences between the groups in phase 3 in comparison to those in phase 1. In phase 3, G1 and G3 each had a single consistent dominant individual, G2 had ongoing changes in social structure, and G4 had a stable hierarchy but with ongoing events. All groups had consistency in structure, but the correlation of individual scores with the groups was higher for groups G1, G3, and G4 in comparison with group G2 (Fig. 5C,D).Compared to other phase 3 groups, G1 and G3 had a relatively low number of events and a high fraction of events with the dominant individual. Each of these groups had a single individual that was consistently ranked as most dominant (see Fig. 5C). However, in comparison to G1, G3 had on average higher DSS, ESS, and GRC. This and the higher DCI index suggests that group G3 had a steeper hierarchical structure than group G1.In contrast to groups G1 and G3, group G2 did not have a single individual that remained dominant during each period. Group G2 had many events, the lowest fraction of events with the dominant rat, and the lowest transitivity (TTRI) of the groups. This and the lower correlation coefficient in Elo scores compared to other phase 3 groups suggest an ongoing struggle for position within the social network where ongoing events maintained pairwise relationships. However, we note the differences obtained in the hierarchy steepness measures for group G2: the David’s score steepness suggests a weak hierarchy, while the Elo score steepness and GRC suggest hierarchies definitely exist.Group G4 had similar patterns of metrics to Group G3, but with several distinct differences: these include lower magnitudes of ESS, DSS, GRC, and DCI, a lower fraction of events with the dominant individual, and overall many more events (although the mean number of events decreased dramatically from Pd 11 to Pd 12—see Fig. S7). With this, we can describe G4 as having a middle-steep hierarchy that was maintained by many ongoing events among pairs. This differs from G1 and G3, where the high fraction of events with the dominant individual suggests that the hierarchy was maintained mostly by these events.During phase 3, the area available to each group was changed during Pds 11 and 12 by moving the compartment borders that separated the groups. In Pd 11, G1 & G4 had a larger area and G2 & G3 a smaller area. Pd 12 had these sizes switched. These manipulations did not have a consistent effect on space use or group social structure metrics (Fig. S8).Individual social rankings, changes over time, and body massWe found that previous social status in phase 1 did not predict an individual’s placement in the new group social structures of phase 3 (Fig. 6A). This result holds if instead of using absolute Elo score values as shown in Fig. 6A, rank scores of subordinate and dominant are used for the lowest two and highest two Elo scores, with other assigned as middle ranking (see Fig. S6). Because the individual ranking metrics are correlated (Fig. S5), for clarity we focus on showing results with the often-used Elo score32,33,34; however, we also note that the other individual social metrics (including faction of events dominated, David’s score, and local reaching centrality) showed similar trends (Fig. S6) with respect to predicting individual placement.Because social ranking can be used to regulate access to resources such as food, we further examine the relationship between social rank and weight gain/loss. We compare average individual social rankings to weight gain or loss during phase 3. At the beginning of phase 1, all rats were young and gaining weight. However, by the end of phase 1, the average weight gain from the previous period was small, and not all rats were still gaining weight. When the groups were merged in phase 2, the average change in body mass (\(\Delta\) body mass) continued to decrease and was negative for the last period of phase 2 and the first period of phase 3. Specifically, in the new groups of phase 3, the variance in the distribution of \(\Delta\) body mass increased, with one rat (rat \(\alpha 4\), which was subsequently permanently excluded from the experiment) losing nearly 100g relative to the previous period (Fig. 6B).Figure 6C shows that although the two most dominant rats during phase 3 gained the most weight (rats a3 and a1 from groups G1 and G3, respectively), average social dominance rank was not a robust general predictor of body mass changes across all rats. Including these dominant rats, the relationship between the average Elo score during phase 3 and the change in body mass during phase 3 has a significant correlation with the values shown in Fig. 6C. However, this result is not robust: if these two rats are removed, the correlation drops greatly to \(r=0.16\) and is no longer significant (p = 0.44). This result also holds if subordinate-middle-dominant ranks are used instead of absolute Elo score values (\(r=0.48\) and p = 0.013 including all individuals; \(r=0.26\) and p = 0.23 removing rats a3 and a1). This demonstrates the complex relationship between dominance and body size in rodent social groups51,52. While there was likely a feedback regarding social rank and body mass for the two dominant rats in groups G1 and G3, respectively, it is difficult to link weight gain/loss to social rank in a general sense.Fig. 6Phase 1 to phase 3 individual rankings, body mass changes and social ranking. (A) Comparison of individual rat ranking metrics at the end of phase 1 (Pd 7) to those at the start of phase 3 (Pd 10), after the new groups were formed. (B) Distribution of body mass changes over time for all individuals. The middle line is the mean, and rugged curves indicate the maximum/minimum across all individuals. (C) Average Elo score during phase 3 compared to body mass change during phase 3 for each rat. Change in body mass during phase 3 is calculated as the body mass in Pd 12 minus body mass in Pd 10.Individual metrics compared to behavioral assaysWe used individual and pairwise assays performed after the group experiments to test the behavior of each rat. The individual black and white box, canopy, and elevated plus-maze results were used to define a composite boldness score. A pairwise social test with an unfamiliar rat, where two individuals are placed together and various behaviors characterizing interactions, such as sniffing the other, are scored (see “Methods”; Fig. S10), was used to define a social interaction score. This social test, which is also referred to in the literature as the “reciprocal social interaction test”, has been widely employed for behavioral phenotyping related to anxiety and autism53,54,55,56.In general, we find low and/or inconsistent relationships between behavior in groups and behavior in the assays (Fig. 7A). This is shown by the comparison of behavioral metrics from the last period in phase 3 with the individual boldness and social interaction scores. In particular, the social metrics measured in a group setting, including the number of events and the Elo score, do not exhibit consistent or significant correlations with the pairwise social interaction score. Although we see positive correlations for the boldness score compared to the related metrics of distance from wall and home range, and a negative correlation with top of nestbox, these correlations are not significant (p > 0.05), with a notable remark that the 2 most dominant individuals (a3 of G1, a1 of G3) have high boldness scores within their group. However, when considering the individual assays separately, we do find a significant correlation between top of nestbox and time spent in the open area during the elevated cross assay (Fig. S11A). We also find that breeding line and group membership do not consistently predict differences in individual test scores (Fig. 7B).The social interaction score shown in Fig. 7A,B was obtained via pairwise behavioral assays performed with an unfamiliar rat. We repeated these tests with a second unfamiliar rat in order to test repeatability (tests were also performed with a familiar rat from the respective phase 3 groups—see Fig. S13). The composite scores from the tests with the second unfamiliar individual show a low correlation with the scores obtained with the first unfamiliar individual (Fig. 7C, p > 0.05).Other work has noted that individual behavior in assays may depend not only on an individual’s social dominance status, but also on the nature of the social hierarchy of the group to which the individual previously belonged47. To test this, we fit a linear regression model predicting individual behavior assay results based on a combination of individual metrics and the Elo score steepness (ESS) of the group where the individual was located during Pd 12. While including this additional information increased explanatory power, we did not observe consistent significant patterns (Fig. S12).Fig. 7Behavioral metrics at the end of phase 3 compared to assays. (A) Pearson correlation values for space use and social behavioral metrics from the final period in phase 3 (Pd 12) with individual assay scores. Labels and color scales denote correlation values. Note that none of the correlations are significant (all p-values \(>0.05\), calculated using t-distribution). (B) Individual score distributions according to breeding line (left), and by phase 3 group membership (right). Scores are normalized by the mean and standard deviation of values measured for all rats. (C) Comparison of social interaction scores calculated from tests with a first unfamiliar rat (x-axes; values shown in A,B), with scores calculated from tests with a second unfamiliar rat (y-axes). See also Figs. S11 and S13.
