We began by generating a common coordinate framework (CCF) for the prairie vole brain by iteratively averaging tissue autofluorescence from 190 brains imaged using light-sheet fluorescence microscopy (LSFM) (Video 1). Each of the brains was co-registered into this coordinate framework for computational analyses. Next, we registered an LSFM-based CCF of the mouse brain onto the prairie vole CCF, enabling us to apply anatomical labels of the Allen Reference Atlas (ARA; Figure 1A, Figure 1—figure supplement 1, and Video 1; Dong, 2008). This alignment revealed that the prairie vole brain (n = 190, 94 males and 96 females) is ~30% larger than the mouse brain (n = 108 males, negative binomial test: z = 36.957, false discovery rate [FDR] corrected q < 0.0001) and distinct in shape, but the relative volumes of brain regions were consistent across species (q > 0.05 for all regions) and showed no evidence of sexual dimorphism (q > 0.05 for all regions) (Figure 1A, Figure 1—figure supplement 1, Supplementary file 1, and Supplementary file 2).

To validate and refine the Allen Reference Atlas (ARA) anatomical borders in the prairie vole brain, we performed whole-brain Nissl staining as well as iDISCO immunolabeling targeting the cell-specific markers tyrosine hydroxylase, parvalbumin, and somatostatin; we then aligned these images onto the prairie vole CCF (Figure 1B, C and Figure 1—figure supplement 1). Lastly, we adapted for the prairie vole our computational pipeline for whole-brain detection and statistical comparisons of c-Fos+ neurons in LSFM-imaged brains (Kim et al., 2015; Kim et al., 2016; Renier et al., 2016; Figure 1C). With a robust atlas and analysis tools in place, we conducted a detailed study of the time course of mating and bond formation to examine IEG induction across the prairie vole brain and test our hypotheses.

Sexually responsive voles typically mate within the first hour of pairing, and repeated mating over ~6 hr initiates partner preferences characteristic of attachment (DeVries and Carter, 1999). The bond stabilizes between ~12 and 24 hr, leading to prolonged changes in attachment and related behaviors (DeVries and Carter, 1999; Williams et al., 1992). With these milestones in mind, we precisely manipulated mating experience and examined how repeated sexual behaviors lead to a bond (Figure 2A, Figure 2—figure supplement 1, Figure 2—figure supplement 2).

To coordinate the timing of mating, all subjects were isolated for 4–5 days, females were brought into estrus, and both sexes were screened for sexual receptivity (see Figure 2A). Subjects assigned to the bonding condition were paired with a novel opposite-sexed individual; to control for non-sexual social affiliation, remaining subjects were re-paired with a same-sex sibling who had been a cagemate prior to isolation. Members of each pair were isolated on either side of a divider for 2 hr. Following this acclimation, the divider was removed, and the pair could interact freely (Figure 2A). Behavior sessions were terminated following a fixed timeline: animals were euthanized just before barrier removal (0 hr), following initial mating (2.5 hr), after initial bond formation (6 hr), or after bond stabilization (22 hr). Automated behavioral measures, such as proximity, vocalization, and relative movement were scored throughout the behavioral sessions. We performed detailed manual scoring for 1 hr focal intervals beginning 2 hr before euthanasia – a timeframe that reveals the behavioral states of animals during IEG induction (Figure 2B, C, Figure 2—figure supplement 1, Figure 2—figure supplement 2, Figure 2—figure supplement 3, and Supplementary file 3).

Mate and sibling pairs did not differ in locomotor activity or time spent in side-by-side contact (‘huddling’; two sample t-tests: velocity, n = 94 mates and 96 siblings, t = 1.592, false discovery rate [FDR]-corrected q = 0.170, huddling, n = 47 mate and 48 sibling dyads, t = 1.122, q = 0.3663; Figure 2D). Mate pairs progressed through known stages of mating behavior, showing elevated rates of anogenital investigation and vocalization (t-tests: investigation, n = 94 mates and 96 siblings, t = 4.184, q = 0.0003, vocalizations, n = 48 mate and 48 sibling dyads, t = 4.708, q = 0.0003); males moved more often toward females, while females moved away from their partners (paired t-test: net movement, n = 47 females and 47 males, t = 3.127, q = 0.0138), a behavior consistent with female copulatory pacing (Pfaus et al., 2001). Consummatory aspects of mating – mounting, intromission and ejaculation – were common in both the 2.5 and 6 hr focal windows; by 22 hr mating was rare, and levels of male–female huddling resembled those of same-sex siblings housed together since birth (Figure 2E). This profile is consistent with literature on the timing and process of prairie vole bond formation (DeVries and Carter, 1999; Williams et al., 1992).

Next we measured brain-wide c-Fos immunostaining, a common proxy for neuronal activity and plasticity (Sheng and Greenberg, 1990). To identify brain areas that differed between mate-paired and sibling controls, we used a generalized linear model (GLM) to compare two alternative models. A null model included terms for sex, time, and experiment block; a full model included these variables, as well as pairing status and two interaction terms (pairing × sex, pairing × time). Brain-wide comparisons using voxels or regions of interest (ROIs) were largely concordant (Figure 3A, B, Figure 3—figure supplement 1, and Video 2), revealing an extensive but specific network of brain regions active during mating and bonding. The brain atlas is organized hierarchically, and of 99 ROIs that differed significantly (permutation test: n = 189 animals, permutations = 10,000, FDR-corrected q < 0.1; Figure 3B and Supplementary file 4), 68 regions were anatomically distinct.

A hierarchical cluster analysis assigned these 68 ROIs into 8 groups (Figure 3C, D, Figure 3—figure supplement 2, and Figure 3—figure supplement 3). Each of these groups, or ‘clusters’, included multiple brain regions but often centered on a specific structure or group of structures. The purple ‘BST’ (bed nucleus of the stria terminalis) cluster contained regions of the preoptic area and periventricular nucleus of the hypothalamus, as well as multiple substructures within the BST. The blue ‘POA-VMH’ (preoptic area, ventromedial hypothalamus) and light blue ‘PVH’ (paraventricular hypothalamus) clusters contained multiple regions within the hypothalamus, including the paraventricular nucleus of the hypothalamus, as well as regions of the preoptic area. The green ‘PFC’ (prefrontal cortex) cluster was composed of regions within both the prefrontal cortex and olfactory areas. The orange ‘Amyg’ (amygdala) cluster contained the lateral hypothalamus and various olfactory related regions, but the majority of the cluster consisted of amygdalar nuclei such as cortical and medial amygdala and the anterior amygdala area. The light orange ‘AUD’ (auditory cortex) cluster contained some regions within auditory cortex, and the red ‘Thal’ (thalamus) cluster involved a variety of thalamic regions. The light purple ‘AOB’ (accessory olfactory bulb) cluster contained minimally correlated regions (nucleus y, accessory facial motor nucleus, and AOB glomerular layer), and so we do not consider this group to be one of the major clusters. Comparing the regions with connectivity reported in the ARA mouse ‘connectome’ (Knox et al., 2019) suggests that six major clusters in our dataset (BST, POA-VMH, PVH, PFC, Amyg, and AUD) are anatomically connected (permutation test with 10,000 shuffled datasets, p = 0.0001, Figure 3—figure supplement 4).

Although the nucleus accumbens did not survive multiple test corrections in our ROI analysis (n = 189 animals, F = 2.936, FDR q = 0.1747), it was significant in the univariate analysis (p = 0.0304), particularly when focused on the 2.5 and 6 hr timepoints (two sample t-test: n = 47 mates and 47 siblings, t = 2.530, p = 0.0138, Video 2). Furthermore, voxel-level comparisons revealed significant sites within the ventral striatum and the posterior nucleus accumbens (Figure 2A, Figure 3—figure supplement 1, and Video 2).

Canonical correlation analysis (CCA) reveals the latent correlational structure within two sets of variables, and so is well suited to compare the principal dimensions of behavioral variation to its neural counterparts (Figure 4, Figure 4—figure supplement 1, and Figure 4—figure supplement 2). We found that the first canonical correlate (CC1), which defines the largest axis of shared variation in brain and behavior, loaded highly on the BST, POA-VMH, and PVH clusters, as well as on mating-related behaviors (Figure 4A–D). CC1 scores captured responses to mating and bonding in the 2.5 and 6 hr timepoints (Figure 4A, B). The second canonical correlate (CC2) captured differences between animals who were isolated or paired and loaded particularly high on the limbic cortical cluster (PFC; Figure 4B, Figure 4—figure supplement 1).

The first two CC dimensions showed no evidence of sexual dimorphism, while the third dimension suggested subtle but non-significant differences between male and female mates (Figure 4—figure supplement 1). A formal two-model comparison isolating the effect of the sex × pairing interaction revealed no regions that differed following FDR correction (permutation tests: q > 0.1 for all tests). If we limit the comparison to just those 68 regions identified as responsive to pairing, we find 29 unique regions that exhibit evidence of a sex × pairing interaction, although these effects are much weaker than those identified in our above analysis of pairing (Video 2).

Although differences between females and males were modest, the CCA suggested substantial variation across mated pairs. We used these data to test whether pairing involved a coordinated change in neural activation across brain regions. We find that during mating and bond formation (2.5–6 hr), activity is strongly correlated across regions (Figure 4E, F and Figure 4—figure supplement 3). Similar to these patterns distributed across putative pair-bonding regions, CC1 brain scores show a strong correlation between female and male mates (Pearson correlation: df = 22, r = 0.93, p < 0.0001, Figure 4G).

Following bonding (22 hr), within-pair correlations are confined to the coupling of activity between the hypothalamus and amygdala, including between medial preoptic area and medial amygdala (Figure 4E). In siblings, who were housed together since birth, hypothalamus–amygdala correlations are relatively uniform across timepoints, while correlations outside of these regions are largely absent (Figure 4—figure supplement 3).

The principal nucleus of the BST (BSTp), a subregion of the posterior BST, loads highly on CC1 (Figure 4B), and exhibits a strong response to mating in the 2.5 and 6 hr timepoints (permutation test: n = 189 animals, permutations = 10,000, FDR q = 0.003, Figure 4C); we found that this brain region also shows strong correlations in heterosexual pairs (Pearson correlation: df = 21, r = 0.851, p < 0.0001, Figure 4G). Remarkably, the number of observed male ejaculations is strongly predictive of BSTpr c-Fos+ cells in both males and females (Pearson correlation: males: df = 22, r = 0.894, p < 0.0001; females: df = 21, r = 0.847, p < 0.0001; Figure 4H). To our surprise, but consistent with the CCA, male ejaculation rates were the strongest predictor of activity across brain regions of both sexes (Figure 4B, D). Statistically controlling for the number of ejaculations effectively abolished the correlation we observed in mated pairs (Figure 4F).

This brain-wide map of IEG induction provides a uniquely comprehensive perspective on the circuits that enable sexual behavior to become a bond. To help organize this enormous dataset, we looked to existing literature to compare the implicated circuits to those related to sociosexual behaviors in other species (Figure 5A, B). Overlaying the time course of behaviors and circuit activity suggest distinct stages of bonding and corresponding neural function. We divide those stages into mating, bonding, and ongoing bond maintenance.

The first canonical correlate captured the major dimensions of both brain and behavioral variation, and showed the largest group differences surrounding the 2.5 hr timepoint. Behaviorally, this difference mapped strongly onto mating behavior, and in particular to rates of male ejaculation. The neural circuits of rodent sexual behavior in general, and of male ejaculation in particular, have been well studied in rats, hamsters, and laboratory mice (Bayless et al., 2019; Pfaus and Heeb, 1997; Veening et al., 2014; Veening and Coolen, 2014). In this context, chemosensory signals are thought to drive activity in medial amygdala (MeA), medial preoptic area (MPO), and PVH. Copulation drives activity in ventromedial hypthalamus (VMH), MPN, PD, and BST. We found that elevated c-Fos induction levels occur in these areas early on in the bonding time course and then decline. Such findings support the idea that these core circuits orchestrate early social interactions that enable individuals to process chemosensory signals and then initiate lordosis, mounting, and ejaculation. Our study also revealed a set of regions that have yet to be emphasized in sociosexual circuits, including lateral hypothalamic area and lateral preoptic area. While research on these regions emphasize non-social functions, a number of studies suggest that these regions play a role in social behaviors, or participate in goal-directed behavior and reward (Chen et al., 2021; Nieh et al., 2016; Petzold et al., 2023).

Interestingly, one anomaly from the literature on sexual behavior circuits is that females also show IEG induction associated with male ejaculation – a finding evident in both rats and mice (Pfaus and Heeb, 1997; Veening et al., 2014). Similarly, recent work in female laboratory mice demonstrate that Vgat+ neurons in the BSTpr exhibit responses to male ejaculation (Bayless et al., 2019). The extraordinary pattern of coordination we observe between members of a pair across nearly 70 brain regions, and its prediction by male ejaculation, suggest that both females and males are experiencing similar and profound affective states. Moreover, the fact that repeated copulation is necessary for bonding is consistent with work in a variety of taxa suggesting that repeated copulation is a means of partner assessment (Dewsbury, 1988; Eberhard, 1996); in prairie voles, we suggest that both females and males are assessing the ability of a male to monopolize a female, a trait that would predict male paternity and ability to defend a nest against conspecifics in the field. Pfaus et al. have argued that females of other species, including laboratory rodents, exhibit orgasm-like responses (Georgiadis et al., 2012). Although our current data are unable to address this claim directly, the hypothesis offers a parsimonious interpretation of our data, and the topic merits further scrutiny.

In this circuitry, the ‘extended amygdala’ regions of the BST and MeA stand out for their extensive projections to and from the hypothalamus, and for their known roles in individual recognition. Dumais et al., 2016, for example, have documented that the BST of the rat is essential to individual recognition; in parallel, Knoedler et al., 2022 find that estrogen receptor alpha (ERa) in the BST coordinates sex recognition in laboratory mice. Most directly, Cushing et al. have shown that ERa in the BST is essential to the formation of pair bonds in prairie voles (Lei et al., 2010). Similarly, oxytocin actions in the MeA are known to govern social recognition in lab mice (Ferguson et al., 2001). We hypothesize that IEG induction in response to mating allows the identity of a mate to get access to the modulation of hormonal and behavioral states by the hypothalamus. Such identity representations may coordinate partner-specific patterns of neuropeptide release (PVN), selective aggression (ventrolateral VMH [VMHvl]), selective mating (VMHvl, MPO), and social reward (MPO).

After bond formation and stabilization, we find that that mated pairs are similar to co-housed siblings in terms of brain-wide IEG induction patterns (Figure 5C). More remarkable, is that after 22 hr both co-housed siblings and mated pairs show strong correlations in amygdala IEG induction, particularly in the MeApd. We propose that this correlated activity reflects a sensitivity to the behavioral state of a partner that emerges as a function of bonding. Recent work has revealed that in laboratory mice, tachykinin + GABAergic neurons project from the posterodorsal MeA (MeApd) to the MPO, where they contribute to both consolation grooming and social reward (Hu et al., 2021; Wu et al., 2021). These projections have not been implicated in pair bonding, but our current data suggest that this circuit may play a broader role in the ongoing function of relationships in both mates and siblings. Since prairie voles often live in larger family groups (Getz et al., 1993), particularly over winter, a common mechanism for the maintenance of familial and sexual bonds would be consistent with their ecology. In addition, the strong correlation of activity across a variety of related amygdala and associated regions suggest this circuit may be anatomically much broader than the interactions between MeApd and MPO documented in laboratory mice.

Overall, our data survey the brain to identify circuits active as mating produces a pair bond. These data allowed strong tests of opposing hypotheses about sexual dimorphism and coordination during bonding (De Vries, 2004; Hasson et al., 2012; Kingsbury et al., 2019; Kinreich et al., 2017; Long et al., 2020; Zhang and Yartsev, 2019), revealing a surprising absence of sexual dimorphism in structure and function, and an extensive coordination of neural activity among new pairs. We confirm the activity of known regulators of mating and bonding such as reward circuits, the paraventricular nucleus, and the bed nucleus of the stria terminalis (Coolen, 2005; Insel and Young, 2001; Lei et al., 2010; Pfaus, 2009); we map a path by which sexual activity finds its way to known regulators of bonding, and identify a variety of novel regions and circuits whose roles in bonding remain to be examined. Our results suggest a novel model in which the BST is a key node connecting sexual experience to the neuroendocrine functions of the hypothalamus and preoptic area, and that coupling between the preoptic area and the amygdala may play an unappreciated role in the maintenance of established bonds. Manipulations of these circuits and their behavioral consequences offer rich new opportunities for research into the mechanisms of bonding and their contributions to well-being.

Prairie voles (Microtus ochrogaster) used in behavioral and anatomical experiments were derived from wild-caught voles from Jackson County, Illinois and bred at The University of Texas at Austin. At weaning (PND 21), voles were housed in polycarbonate cages (R20 Rat Cage, Ancare Corp, NY) in groups of two to five same-sex littermates and provided with standard rodent chow (LabDiet 5001, Lab Supply, TX) and water ad libitum. The colony room was maintained at a controlled temperature (20–23°C), and the photoperiod was on a 12:12 light:dark cycle (lights on: 0600, lights off: 1800). Housing arrangements allowed animals to receive visual and olfactory cues, but not tactile contact, with conspecifics. Mice used for anatomical research were housed with ad libitum access to food and water in a controlled temperature (21–22°C) and light (12:12 light/dark cycle) room. All animal procedures were approved by the Institutional Animal Care and Use Committees (IACUC) at the University of Texas at Austin (protocols #2016-00247 and #2019-00228) and Cold Spring Harbor Laboratory (protocol #18-15-12-09-12).

Prairie voles underwent intracardiac perfusion ~20 min (mean = 21 min, range = 10–40 min) after the behavioral observation endpoints (2 hr acclimation or 2 hr acclimation in addition to 2.5, 6, or 22 hr cohabitation). Animals were anesthetized with open-drop isoflurane exposure, exsanguinated with 0.9% saline (2 min 30 s at 13 ml/min), and fixed with 4% paraformaldehyde (PFA) in 0.05 M phosphate buffer (PB; 5 min 30 s at 13 ml/min). All harvested brain tissue samples were post-fixed overnight at 4°C in 4% PFA in PB. After post-fixation, samples were washed 3× in 0.05 M PB and stored in 0.05 M PB 0.02% sodium azide at 4°C until immunolabeling processing.

Mice of 8–10 weeks old were anesthetized with ketamine/xylazine and transcardially perfused with isotonic saline followed by 4% PFA in 0.1 M PB (pH 7.4). Brains were extracted and post-fixed overnight at 4°C in the same fixative solution, and stored at 0.05 M PB until immunolabeling processing.

The right hemisphere of all brain samples were cut and immunolabeled for immediate early gene (IEG) c-Fos and posteriorly cleared using iDisco+ protocol (Renier et al., 2014; Renier et al., 2016). Briefly, samples were initially delipidated with methanol and later permeabilized and blocked with DMSO (dimethyl sulfoxide) and donkey serum, respectively. Thereafter, all samples were incubated with c-Fos antibody and Alexa Fluorophore 647 secondary antibodies. Samples were cleared with increasing concentration steps of methanol and dichloromethane as previously described (Renier et al., 2014; Renier et al., 2016). After clearing, samples were imaged sagittally on a light-sheet fluorescence microscope (Ultramicroscope II, LaVision Biotec). Samples were imaged continuously every 5 μm at 640 and 488 nm, for signal and background channels, respectively.

For the construction of the prairie vole reference brain, 190 brains from c-Fos+ cell counting analysis were co-registered as described below (3D registration of the vole brain) (Kim et al., 2015). For the construction of the prairie vole reference atlas, the Allen Reference Atlas (ARA) was initially used as a template. After initial registration, output transformations were used to warp mouse atlas onto the prairie vole reference brain. In order to validate atlas registration, prairie vole Neurotrace stained brains were registered at high resolution onto the prairie vole reference brain.

Brains were registered to a standardized reference brain as previously described (Renier et al., 2014; Renier et al., 2016). Initial 3D affine transformation was calculated using six resolution levels followed by a 3D B-spline transformation with three resolution levels. Similarity was computed using Advanced Mattes mutual Information metric by Elastix registration toolbox. In order to enhance image registration, both brain images and reference brains were pre-processed to reduce the impact of imaging artifacts during the computation of mutual information (Video 1). First, brain image illumination was corrected to homogenize illumination across sections. Second, both brain images and reference brain intensities were smoothed to reduce imaging artifacts (Muñoz-Castañeda and Osten, In Preparation).

Whole-brain Neurotrace staining was performed with a modification of the iDISCO+ protocol (Muñoz-Castañeda et al., 2021; Muñoz-Castañeda and Osten, In Preparation). Samples were initially washed in phosphate-buffered saline (PBS) and incubated for in PBS + Triton X-100 + DMSO + glycine. Then samples were transferred and incubated in a solution with PBS and Neurotrace. Finally, samples were washed in PBS.

For area-based volume quantifications, the reference brain was registered onto each single imaged brain and all volume areas were automatically quantified with custom made scripts (Muñoz-Castañeda and Osten, In Preparation). These area volumes were used for comparisons within and between species (n = 96 female voles, 94 male voles, and 108 male mice).

Before imaging, both prairie vole and mouse brains were embedded and cross-linked with oxidized 4% agarose as previously described (Kim et al., 2017; Ragan et al., 2012). Whole-brain imaging was achieved using the automated whole-mount microscopy with serial two-photon tomography (STPT). The entire brain was coronally imaged at an X,Y resolution of 1 µm and Z-spacing of 50 µm (Kim et al., 2017; Ragan et al., 2012). After imaging, brains were registered to the reference brain for anatomical validation (see 3D registration of the vole brain).

c-Fos+ cells segmentation was performed using convolutional neural networks as previously described (Kim et al., 2016; Kim et al., 2017). After cell segmentation, all cell centroids were calculated for whole-brain distribution analysis. For the analysis of area-based c-Fos+ cell counts, the mouse reference brain was registered onto each prairie vole brain, using the output transformations to warp the atlas onto each individual brain for region of interest (ROI) cell counting and distribution. For voxel analysis, the inverse transformation output was used to move all c-Fos centroids onto the reference brain.

Study subjects were 8–12 week old sexually naive prairie voles. There were eight treatments that varied based on partner type and cohabitation time (Figure 2). Subjects were partnered with a familiar same-sex cage mate (siblings) or an opposite-sex individual (mates) for 0, 2.5, 6, or 22 hr. The 0 hr timepoint represents a baseline state before pairing takes place, the 2.5 hr timepoint is ~2 hr after the first mating bout, the 6 hr timepoint is when an unstable partner preference is established, and the 22 hr timepoint is when pair bonding becomes stable (i.e., after overnight mating) (DeVries and Carter, 1999; Williams et al., 1992). It is important to note that the opaque divider in the acclimation period prevented physical interactions, but it is possible that animal pairs may have detected each other through olfactory or auditory cues. The experiment was run in six testing blocks. Each block was composed of eight testing days spread across 2 weeks, with two adjacent testing days per cohabitation timepoint (0, 2.5, 6, and 22 hr). Two pair sets were tested in a day, either female siblings and male siblings or two mating pairs. The order of partner type (adjacent days per timepoint) and order of cohabitation timepoints (four per block) were counterbalanced across blocks. Individuals were randomly assigned to the partner type condition.

There were 12 sibling pairs and 12 mating pairs tested for each cohabitation timepoint, for a total of 96 female and 96 male study subjects. Of this total, 190 voles were used for behavioral analyses, and 189 voles were used for IEG analyses. A mating pair from the 0 hr timepoint group lacked behavioral data due to a camera malfunction (audio was unaffected). Three brain samples were not used in IEG analyses: one male sibling from the 6 hr timepoint group, one female mate from the 0 hr timepoint, and one female mate from the 6 hr timepoint. The first sample was not stained for c-Fos due to major issues with the perfusion. The latter two samples were identified as outliers by a Rosner’s test (EnvStats R package; Millard, 2013); their whole-brain c-Fos+ cell counts were higher than the rest of the samples (R = 4.61 and 5.504, p < 0.05).

Voles were housed in new home cages, isolated from their sibling cage mates, 4–5 days prior to testing. During this isolation period, all females were induced into estrus with daily 0.1 ml subcutaneous injections of 2 μg estradiol benzoate dissolved in sesame oil (Amadei et al., 2017; Carter et al., 1988). Voles were screened for mating capacity on the third day of isolation. This screening involved a brief exposure (<10 min) to an opposite-sex vole until the first mount. Voles that did not show a mating attempt in the first exposure were retested with a different animal. We used this mating assay to restrict study subjects to voles that showed lordosis (females) or mounting behavior (males). By selecting voles who showed sexual behavior, we could control the estrus state and timing of mating across the 0, 2.5, 6, and 22 hr study groups. This selection process also ensured that animals assigned to the same-sex sibling pair and opposite-sex mating pair groups had similar sexual motivation and experience.

Behavioral testing began in the morning after lights-on between 8:00 and 10:15 (mean = 8:44 am). Subjects were fitted with colored collars (pipe cleaner attached to a miniature cable tie) for identification and automated video tracking. Then, subjects were placed on opposite sides of a custom made acrylic testing arena (12″ × 24″ × 12″), separated by an opaque divider, for a 2 hr acclimation period. The testing arenas were outfitted with fresh bedding and ad libitum access to food and water. The arenas were housed within an enclosed experiment box (42″ × 34″ × 24″) constructed from expanded PVC. These experiment boxes had controlled white:red lighting (Phillips Hue light strips) on the same photoperiod cycle as the colony room. For the 0 hr timepoint, subjects only underwent a 2 hr acclimation period. Otherwise, the divider was removed at the end of acclimation so that subjects could interact freely for 2.5, 6, or 22 hr. After this acclimation (0 hr group) or cohabitation (2.5, 6, and 22 hr groups), subjects were promptly removed from the arena and perfused. Arenas were cleaned with 70% ethanol between tests.

Video and audio data were recorded from Basler Ace-IMX174 cameras (1920 × 1200, 2 MP resolution, Basler AG, PA) and Ultramic 364K BLE microphones (Dodotronic, Italy) suspended above each testing arena. Cameras were outfitted with a 16-mm lens and 46-mm linear polarizer. Video data were recorded with Pylon viewer (v 5.1.0.12681, Basler AG, PA) at 25 fps with white balance set to ~1.93 and exposure levels between 15 and 25 ms. Audio was recorded with SeaPro2 software (v 2.0j, CIBRA) at a 192-kHz sampling rate and saved in 30 min WAV files (starting on the hour and every half hour). With this setup, ultrasonic vocalizations (USVs) could be detected per pair, but not localized to specific individuals. Recording equipment was connected directly to a Windows 10 PC. The start of both video and audio recordings were timestamped with PC system time, which enabled the synchronization of audio and video data on a second-by-second timescale.

Behaviors were measured during observation windows that corresponded to peaks in c-Fos induction (Kim et al., 2015; Renier et al., 2016), specifically 60–120 min before each pair’s perfusion (i.e., midpoint between exsanguination times of each vole pair). Both automated and manual methods were used to characterize vole behavior.

Automated scoring of movement and proximity behaviors was done with Ethovision (v 10.1). Individual collars were tracked during white light periods, with detection settings set for each video and optimized for collar color (blue/green, mean ± standard deviation [SD]: huemin = 97.0 ± 1.4/63.1 ± 2.5, huemax = 113.0 ± 1.4/79.2 ± 2.6, saturationmin = 113.3 ± 11.8/85.4 ± 10.5, saturationmax = 255.0 ± 0.2/254.9 ± 0.7, brightnessmin = 84.1 ± 10.5/70.2 ± 8.9, brightnessmax = 254.0 ± 4.6/254.7 ± 1.6, marker size = 50). Body area was recorded across both white light and red light periods with the grayscaling method (white/red light: detectionmin = 0, detectionmax = 108.7 ± 3.3/5.96 ± 0.8, pixel sizemin = 3000, pixel sizemax = 125,000, contour erosion = 1, contour dilation = 1). The largest body area was recorded when animals were in physical contact. Overall activity (% pixel change between frames) also was recorded for white light and red light periods separately (white/red light: threshold = 1, background = 10, compression = ‘on’ for all 24 overnight videos/‘on’ for 18 overnight videos). Individual tracking, body area, and activity were recorded at 12.5 samples/s, down sampled to 1/s, and used to compute an array of behavioral measures across entire recording sessions and the 1 hr observation windows (Figure 2—figure supplement 1 and Supplementary file 3).

Manual scoring of mating, investigative, grooming, and contact behaviors was done with BORIS (v 7.9.6) (Friard et al., 2016). Two trained observers, who were blind to animal sex and partner condition, independently labeled social and non-social behaviors in cohabitation videos (2.5, 6, and 22 hr groups) during the 1 hr observation windows (Supplementary file 3). Self-grooming during acclimation videos (0 hr group) was labeled by one of the trained observers. Inter-observer reliability was assessed with Pearson correlations. For dyadic behaviors (e.g., mounting and huddling), reliability was assessed across vole pairs. For individual behaviors (e.g., anogenital sniffing and self-grooming), reliability was assessed separately for individuals with blue collars and green collars. All social behaviors had high agreement between the independent observers (Pearson correlations, r > 0.90; Supplementary file 3).

Ultrasonic vocalizations (USVs) were detected with DeepSqueak (v 2.6.1 in MATLAB 2019a) (Coffey et al., 2019). To initially label USVs, the ‘AllShortCalls’ network was used along with the following detection settings: frequency range of 20–90 kHz, overlap of 0.1 s, 5 s chunk length, and a high precision recall. Then, a trained observer revisited all USVs labels during the 1 hr behavior observation windows. Using the DeepSqueak interface, the observer removed labels of background noise (e.g., water bottle sounds) and adjusted USV label boundaries to exclude noise. This trained observer was blind to the partner condition, caller sex, and behavioral context. The start and stop times of all USV labels were exported and the times were adjusted to align with the video frame timestamps.

Absolute and relative volumes of brain regions were compared across species (mouse vs. vole) and sexes (male voles vs. female voles) with negative binomial regression models. Relative volumes were computed as the ratio of region volumes to whole-brain volumes. To control for multiple comparisons, p-values were corrected with the False Discovery Rate (FDR) method (Benjamini and Hochberg, 1995).

All statistical analyses for behavioral experiments were carried out in R (v 3.5.3) (R Development Core Team, 2016). Welch two-sample t-tests were used to compare behavioral measures between mate pairs and sibling controls. Paired t-tests were used to compare behavioral measures between female mates and male mates. T-tests were run for all timepoints combined and for specific timepoints. Significance values were adjusted for multiple comparisons using the FDR method, with an alpha threshold of q = 0.05. Pearson correlations were used to make associations between behavioral measures and ROI c-Fos+ cell counts.

To identify brain regions that are sensitive to pairing status, we used a model comparison approach. For each voxel or ROI, two general linear regression models (‘GLMs’; a quasi-Poisson link function was used to model over-dispersed count data) were used to predict IEG c-Fos expression across all individuals (n = 189) (Figure 3, Video 2). These models differed in whether the partner condition (sibling vs. mate) was a part of the formula. The ‘null’ model (Equation 1) included main effects of sex [S], timepoint [T], and experiment block [B]. The hypothesized ‘bonding’ model (Equation 2) included the same predictor terms in addition to a main effect of partner type [P] and interactions between partner with sex or timepoint. In these models, sex was a categorical factor (males vs. females). Timepoint was an ordinal variable from 1 to 4 (0, 2.5, 6, and 22 hr) and was included as a polynomial term to account for both linear and quadratic (i.e., 0/22 vs. 2.5/6 hr) effects. Block was an ordinal variable from 1 to 6. Partner was a categorical variable (mates vs. siblings). This model comparison allowed for the identification of voxels/ROIs where c-Fos induction variation was explained better when accounting for the partner condition.(1)Yi=β1Si+β2Ti+β3Ti2+β4Bi+ei(2)Yi=β1Si+β2Ti+β3Ti2+β4Bi+β5Pi+β6Pi∗Si+β7Pi∗Ti+β8Pi∗Yi2+ei

The null and bonding GLMs were compared with an analysis of variance (ANOVA) test. For the ROI-level analysis, a Monte Carlo permutation approach (10,000 random shuffles of the data) was used to determine a null distribution for the ANOVA F-statistics and compute p-values. For the voxel- and ROI-level analyses, the FDR method was used to correct for multiple tests, with an alpha threshold of q = 0.1.

We ran an additional model comparison to specifically test for sex differences in c-Fos induction (Video 2). In this analysis, we compared the full version of the model formula (Equation 2) to a reduced version of the model (Equation 3). The reduced model excludes a sex by partner interaction term. The FDR method was used again to correct for multiple tests across all voxels and all ROIs, with an alpha threshold of q = 0.1.(3)Yi=β1Si+β2Ti+β3Ti2+β4Bi+β5Pi+β6Pi∗Ti+β7PiTi2+ei

ROIs were chosen for subsequent analyses if they surpased the FDR alpha threshold (q = 0.1) in Equation 1 vs. Equation 2 model comparison and were mutually exclusive with one another. To select mutually exclusive ROIs, we constructed an structural hierarchy from the ARA. For any significant ROIs that overlapped anatomically, we iteratively chose the ROI with the higher ANOVA F-statistic. This method excluded larger brain regions in lieu of smaller and more localized ROIs.

Significant ‘chosen’ ROIs were assigned to groups by using hierarchical clustering with the ward D2 method, with a Euclidean distance matrix extracted from c-Fos+ cell counts (Murtagh and Legendre, 2014; R Development Core Team, 2016). The resulting tree was cut so that ROIs were grouped into anatomically similar clusters. We used multi-dimensional scaling (MDS) to further interpret the degree of similarity between chosen ROIs and their clustering groups based on the Euclidean distance matrix. The MDS method is a form of non-linear dimensionality reduction to visualize similarity in Cartesian space (Cox and Cox, 2008).

To confirm whether chosen ROI clusters represented structural circuits, we compared cluster assignments to published data on structural connectivity in the mouse brain (Knox et al., 2019). First, we refined a matrix of ROI–ROI ipsilateral normalized connection densities to align with our list of chosen ROIs. Some of our chosen ROIs did not align with this matrix because they represented subregions of the ROIs in the matrix. In those cases, we used data from the next inclusive ROI that was available (e.g., BST data used for BSTpr). Then, for each cluster, we found the mean normalized connection density between the regions. We excluded data from the matrix diagonals (e.g., BST to BST) to emphasize connections between, rather than within, regions. We took an average of these cluster densities values to capture the overall connection density based on our cluster assignments. A permutation test was used to assess whether this connection density was higher than expected by chance. The rows of the connectivity matrix (rows = origin ROIs, columns = target ROIs) were randomly shuffled prior to computing the mean cluster density, and this was done 10,000 times to construct a null distribution. This null distribution was then compared to the observed density to estimate its probability.

We used canonical correlation analysis (CCA) to investigate the relationships between behavioral variables and IEG induction patterns in chosen ROIs. CCA is an unsupervised method that finds linear combinations of two variable sets with the strongest correlation (González et al., 2008; R Development Core Team, 2016; Wang et al., 2020). This approach enabled us to isolate discrete canonical correlates (CC), where each individual animal is assigned scores for each variable set per CC factor. We used correlations between individual behavior/ROI measures and their CC scores to interpret CC factors and identify specific behaviors and ROIs with the strongest associations. Wilk’s lambda test statistic was used to confirm which CC factors represented a significant association between the two variable sets.