Multi-dimensional social relationships shape social attention in monkeys

Abstract

Social relationships guide individual behavior and ultimately shape the fabric of society. Primates exhibit particularly complex, differentiated, and multidimensional social relationships, which form interwoven social networks, reflecting both individual social tendencies and specific dyadic interactions. How the patterns of behavior that underlie these social relationships emerge from moment-to-moment patterns of social information processing remains unclear. Here, we assess social relationships among a group of four monkeys, focusing on aggression, grooming, and proximity. We show that individual differences in social attention vary with individual differences in patterns of general social tendencies and patterns of individual engagement with specific partners. Oxytocin administration altered social attention and its relationship to both social tendencies and dyadic relationships, particularly grooming and aggression. Our findings link the dynamics of visual information sampling to the dynamics of primate social networks.

Keywords: individual engagement; neuroscience; oxytocin; rhesus macaque; social attention; social engagement; social relationships.

Figure 1.. Different behavioral dimensions of social relationships among a group of four monkeys within a colony room.

(a) The diagram depicts the procedure of acquiring social behaviors in various pairs of monkeys in the study. (b), Different social interactions exhibited by an example pair throughout the course of two experimental days. (c), The percentage of time engaged in three dimensions of behavior was quantified as a fraction of the total recording duration, in the given example pair. (d-f), Three dimensions of interaction behaviors in all dyads were evaluated across the total recording duration using a dichotomous framework with values of 1 and –1. The diameter of a given circle presents the cumulative scores obtained by that monkey in aggression (d), grooming (e), and proximity (f). The number on each arrow indicates the proportion of individuals engaged in the corresponding social behavior. (g), Social interactions among the group of four monkeys were quantitatively assessed based on three dimensions of social relationships, with an equal weight of 1 for each dimension. The thickness of lines corresponded to the interaction score in dyads. The monkey who receives an arrow is considered a negative score. Conversely, the monkey who sends an arrow earns a positive interaction score. (h), Social engagement index (SEI), which quantifies general social tendencies within the group, was computed by aggregating the scores of three social interactions between a specific monkey and the other three groupmates. (i), Individual engagement index (IEI; color-filled circles) evaluates the social interactions between a given subject and his specific counterparts, compared to their related SEIs (gray-empty circles).

Figure 1—figure supplement 1.. A multi-stage automatic detection system designed to recognize social interaction between monkeys.

(a), Detection of monkeys' postures using the YOLOv5 model: Two neighboring monkeys are recorded simultaneously by camera systems. Frames were randomly selected from the films and identified with individuals’ postures, including information about their heads, bodies, and buttocks. These labeled images were applied for training the object detection framework YOLOv5s. Then the trained YOLO model was employed for inference, in conjunction with the records. This method can yield the identification of each individual monkey, the coordinates of the bounding box surrounding the object, and the category of the bounding box. Based on this information, an individual’s postures can be assessed as a function of time. (b) Detection of social interactions within pairs of monkeys by the analysis of time sequences. The posture sequences of dyadic monkeys were combined and compared frame by frame in order to provide a mapping of social behavioral modes, such as grooming, aggression, and proximity. (c, d) During the YOLO training process, the confidence loss of the object detection box (which measures the grid confidence), the localization loss (which quantifies the discrepancy between the predicted and annotated bounding box coordinates), and the classification loss (which measures the loss of category error between anchor boxes and their corresponding annotations) were tracked across epochs. (e) The precision and recall curve represent the average curve for all categories. The area enclosed by the curve corresponds to the average precision (AP) value, when the IoU is 0.5. (f) The mean average precision (mAP) value at IoU = 0.5 is 0.97, and the average of all mAP values, which are calculated by increasing the IoU threshold from 0.5 to 0.95 in steps of 0.05, is 0.89.

Figure 2.. Individuals’ social attention in response to the social interference task.

(a) The experimental design encompasses both social interference trails and visual motor trials. (b) The distractor stimuli consist of seven monkeys’ neutral face photographs and the related scrambled images. Four monkeys among them were raised in a colony group (in-group) and served as subjects in the study. The left three ones were from outside the group (out-group). (c, d) The eye position (top) and the horizontal eye velocity (bottom) on sample trials from Monkey A. The response times (arrows pointed) were determined utilizing a threshold of eye velocity at 50 deg/s and eye acceleration at 500°/s² (black dash lines). (e, f) The response time distributions were shown based on social interference trials from subject Monkey A’s example experimental day, including data to an intact face picture (blue bars) and its scrambled image (gray bars) under congruent (e) and incongruent (f) conditions. The arrows present the corresponding average values, and the black dash lines show the average response time in visual motor trials. (g) The distractor of intact monkey faces had a notable impact on the saccade latencies of all four subjects. Each gray line represents the behavioral responses to a certain monkey face picture and its scrambled image on a given day. (h) The drifting time in the response time caused by the presence of both intact faces and their scrambled images. (i) To synthesize the interference effects of social attention on monkey intact face pictures, the difference in the drifting times between the incongruent and congruent conditions was analyzed as the distractor bias. The distractor bias varies significantly among the four subjects. Student’s t-test, ***p<0.001; *p<0.05. MA and MK: p=2.32 × 10^–17, MK and ML: p=6.77 × 10^–22, ML and MC: p=7.73 × 10^–14, Error bars: Mean ± SD. (j), Distractor biases of four monkeys in the group are negatively correlated with their SEIs_(1,1,1) (r=–0.87, p<0.001), as determined by a two-sided Pearson’s correlation coefficient. Error bars: Mean ± SD.

Figure 3.. Three dimensions of social relationships distinctly shape attention towards in-group and out-group members.

(a), Methodological schematic defining correlations between social engagement indexes (SEIs) and distractor biases towards the in-group and out-group members. (b) The distribution of distractor biases on an experimental day for the example Monkey A. Arrows indicate the related mean values (Congruent: In-group: 113.32±37.05 ms, Out-group: 102.21±33.34 ms; Incongruent: In-group: 216.94±40.26 ms, Out-group: 226.08±41.51 ms). (c) Individuals in the same group exhibited more social attentions in monkey faces from the out-group, except MC (paired Student’s t-test, MA: p=2.90 × 10^–3, MK: p=1.49 × 10^–2, ML: p=1.01 × 10^–3, MC: p=0.14). Error bars: Mean ± SEM. (d, e) Quantification of weights for three behavioral dimensions computed from the correlations between SEI and distractor bias for in-group (d), and out-group monkeys (e). X, Y, and Z axes represent the weights of proximity, grooming, and aggression. The colormap illustrates the correlations between distractor bias and SEIs (r values), which were estimated by varying weights in the model from 0 to 1 with a step of 0.02. The white dot indicates the maximized correlation with the smallest negative r value. Each slice represents the r value surface along the X, Y, and Z axes, respectively. (f, g) Pseudo-colored heatmaps (f) and polar graph (g) show the ‘best’ weights for three social dimensions (in-group: ω_a = 0.34, ω_g = 0.54, ω_p = 0.58; out-group: ω_a = 0.48, ω_g = 0.62, ω_p = 0.92). (h) The liner relationships between monkeys’ distractor biases and their SEIs with the ‘best’ weights. Error bars: Mean ± SD. (i) The differences in the SEIs between any two monkeys in H were scaled hexagonally in the circle, illustrating the magnitude of general social tendency in the colony. Compared to data towards out-group (green lines), the impact of social engagement was smaller when monkeys viewing in-group members (orange lines).

Figure 3—figure supplement 1.. Statistical analysis on the initial saccade directing towards the distractor in the incongruent condition.

(a, b) In the incongruent condition, monkeys occasionally made their initial saccades to the distractor, and then quickly produced a corrective saccade towards the target. (c) The choice probability of the initial saccade directing towards distractors in the incongruent condition. The monkeys exhibited a higher inclination to direct their initial saccades towards the distractor of intact monkey face (blue cycles), indicating a notable influence on attentional capture, except Monkey C who is scored by the highest SEIs (paired Student’s t-test, MA: p=1.37 × 10^–34, MK: p=2.58 × 10^–2, ML: p=1.41 × 10^–5, MC: p=0.18). Error bars: Mean ± SD. (d) The differences of choice probability between intact monkey faces and scrambled images was calculated separately for the in-group and out-group. The change towards distractors from the in-group were significantly less than those towards distractors from the out-group across all four monkey subjects and all experimental days (paired Student’s t-test, p=8.30 × 10^–6, n_in = n_out = 48). Data points lie well above the dashed unity line. (e) The differences of choice probability in individual monkeys towards in-group and out-group distractors. Only Monkey A and L exhibited significantly higher changes towards out-group (paired Student’s t-test, n_in = n_out = 12, MA: p=5.28 × 10^–5, ML: p=2.40 × 10^–3; MK: p=0.54, MC: p=0.44). Error bars: Mean ± SD. (f) The differences of choice probability exhibited a negative correlation with individuals’ SEIs calculated in Figure 3h (two-sided Pearson’s correlation coefficient, in-group: r=–0.71, p=1.21 × 10^–23; out-group: r=–0.73, p=1.12 × 10^–25). Error bars: Mean ± SD. (g), The administration of oxytocin failed to significantly modulate the differences of choice probability in monkeys (n=48, paired Student’s t-test, in-group: orange bars, p=0.93; out-group: green bars, p=0.41). Error bars: Mean ± SEM.

Figure 3—figure supplement 2.. The effect of a single facial stimulus on distractor biases.

(a) A monkey face distractor was randomly withdrawn from both the in-group and out-group stimuli for each monkey subject. This resulted in three combinations for each subject and a total of 81(3⁴) combinations of subjects and distractors. (b) The datasets were aligned using all p-values calculated for each combination of in-group and out-group. The color map on the right side indicates the negative logarithm (base 10) of the p-value. (c) The pie plot summarized that 99.30% of 6561 combinations had significantly smaller distractor biases towards in-group monkey faces, compared to out-group faces, with p-values of less than 0.05.

Figure 3—figure supplement 3.. The correlation between distractor bias and three dimensions of social relationships in frameworks of social engagement and individual engagement.

(a-c) Correlations between distractor bias for in-group monkeys and social engagement scores for aggression, grooming, and proximity (two-sided Pearson’s correlation coefficient, aggression: p=3.40 × 10^–37, r=–0.83; grooming: p=2.73 × 10^–13, r=–0.56; proximity: p=9.43 × 10^–5, r=–0.32). (d-f) The correlations between distractor bias for out-group monkeys and social engagement scores for aggression, grooming, and proximity. (two-sided Pearson’s correlation coefficient, aggression: p=5.10 × 10^–38, r=–0.83; grooming: p=8.86 × 10^–10, r=–0.48; proximity: p=1.29 × 10^–7, r=–0.42). (g-i) Correlations between distractor bias towards in-group monkeys and individual engagement scores for aggression, grooming, and proximity (two-sided Pearson’s correlation coefficient, aggression: p=3.83 × 10^–21, r=0.68; grooming: p=3.01 × 10^–6, r=0.38; proximity: p=3.31 × 10^–3, r=0.18).

Figure 4.. Impact of oxytocin on social attention regulation by multi-dimensional social relationships.

(a), Experimental schematic (top) illustrating the administration of saline and oxytocin. Oxytocin greatly enhances the social attention towards distractors from outside the group (green bars, n=48), while there was no significant change in the group’s attention towards faces from the in-group (orange bars) (*p<0.05; **p<0.01; ***p<0.001, paired Student’s t-test. Error bars: Mean ± SEM). (b-e), The ‘best’ weights of three social dimensions by analyzing the correlations between social engagement indexes (SEIs) and distractor bias towards in-group and out-group monkey distractors following administration of saline (b, c) and oxytocin (OT) (d, e). (f), Change ratios of weights after OT and saline administration for in-group and out-group distractors. (g, h) Pseudo-colored heatmaps represent ‘best’ weights relating three social dimensions to in-group and out-group distractor bias under OT and saline. (i, j) Relationship between monkeys’ SEIs defined by ‘best’ weights and their distractor biases towards in-group (i) and out-group (j) (the slopes of in-group: saline: –26.07, OT: –60.47; the slopes of out-group: saline: –19.75, OT: –34.84). Error bars: Mean ± SD. (k), The differences in the SEIs between any two monkeys following administration of saline and OT (i, j) were scaled hexagonally in the circle. Compared to data under saline (empty circles), OT (filled circles) reduced the impact of social engagement on social attention under both conditions: viewing in-group and out-group monkey faces.

Figure 4—figure supplement 1.. Saccadic eye movements observed during visual motor trials were unrelated to social relationships within the group.

(a) The design of visual stimuli in visual motor trials. When comparing social interference trials, there was an absence of distractor images given during these trials. (b) Mean horizontal eye velocity traces of the four monkey subjects recorded on an example experimental day. (c) The distribution of saccade latencies in individuals in B. The arrows symbolize the mean values of individual monkey subjects, presented separately. (d) The saccade latency in visual motor trials failed to demonstrate linear associations with the social engagement indexes (SEIs) calculated from the in-group and out-group settings in Figure 3h (two-sided Pearson’s correlation coefficient, in-group: p=0.08; out-group: p=0.18). Error bars: Mean ± SD. (e) Oxytocin does not have a significant effect on the saccade latency in visual motor trials across all four monkeys, as compared to saline (paired Student’s t-test, n=48, p=0.31).

Figure 5.. Individual engagements on three dimensions of social behavior distinctly shape social attention towards in-group.

(a) Methodological schematic defining the correlations between the individual engagement index and the distractor bias towards in-group. (b) Quantification of weights for three behavioral dimensions computed from correlations between individual engagement indexes (IEIs) and distractor biases. The intersection point of the three slices in the graph signifies the position of the maximal r value (white point), with each slice representing the r value surface along the X, Y, and Z axes, respectively. (c) Polar graph shows the ‘best’ weights for three social dimensions (ω_a = 0.54, ω_g = 0.04, ω_p = 0). (d), Relationship between IEIs with ‘best’ weights and distractor biases (two-sided Pearson’s correlation coefficient, p<0.001, r=0.69). (e, f) The ‘best’ weights of three social dimensions by analyzing the correlations between IEIs and distractor bias towards in-group following administration of saline (e) and OT (f). (g) Pseudo-colored heatmaps represent the ‘best’ weights of three dimensions of social interaction behavior in (e) and (f). (h) Relationship between IEIs and distractor biases (two-sided Pearson’s correlation coefficient, saline: p<0.001, r=0.63; OT: p<0.001, r=0.64. The slope of saline: 48.63; the slope of OT: 23.21).