Early-life sleep disruption (ELSD) has been shown to have long-lasting effects on social behaviour in adult prairie voles (Microtus ochrogaster), including impaired expression of pair bonding during partner preference testing. However, due to the limitations of manual behaviour tracking, the effects of ELSD across the time course of pair bonding have not yet been described, hindering our ability to trace mechanisms. Here, we used pose estimation to track prairie voles during opposite-sex cohabitation, the process leading to pair bonding. Male-female pairs were allowed to interact through a mesh divider in the home cage for 72-h, providing variables of body direction, distance-to-divider and locomotion speed. We found that control males displayed periodic patterns of body orientation towards females during cohabitation. In contrast, ELSD males showed reduced duration and ultradian periodicity of these body orientation behaviours towards females. Furthermore, in both sexes, ELSD altered spatial and temporal patterns of locomotion across the light/dark cycles of the 72-h recordings. This study allows a comprehensive behavioural assessment of the effects of ELSD on later life sociality and highlights subtle prairie vole behaviours. Our findings may shed light on neurodevelopmental disorders featuring sleep disruption and social deficits, such as autism spectrum disorders.

Subjects

Prairie voles were bred and reared by both parents at the Veterans Affairs Portland Health Care System. Litters with both males and females were submitted to early-life sleep disruption (ELSD) or Control conditions when pups were at postnatal day 14 - 21 (P14-P21; see below for ELSD procedure). Subjects were weaned at P21 into groups of 2-4 same-sex siblings per cage [Male-ELSD (n = 13), Male-Ctrl (n = 14), Female-ELSD (n = 13), Female-Ctrl (n = 15)] and co-housed at the same breeding site until reaching adulthood. The groups of siblings were transferred between P50-P90 to the University of Michigan Medical School for the main recordings. Prairie voles were allowed to acclimate to the facility transfer for two weeks before experimentation. Housing conditions were the same throughout experiments, including controlled temperature, ventilation and humidity, bedding, ad libitum food (mixed diet of rabbit chow, corn, and cracked oats) and water (bottles/hydrogel), environmental enrichment (cotton nestlets and wooden blocks/sticks), and 14:10 h light/dark cycle (lights on at 5:00 am). Cages and nestlets were changed weekly. The prairie voles we used derived from a colony at Emory University (Dr. Larry Young), which in turn originated from field-caught animals in Illinois. Genetic diversity has been maintained through bi-annual donations among researchers across the USA (North Carolina State, University of California Davis, University of Colorado Boulder, and Florida State University).

Early-life sleep disruption

Litter-containing home cages with both parents were placed on an orbital shaker (turned on every 110 seconds for 10 seconds, 110 rotations per minute) when pups were at P14-P21 of age, thus generating ELSD phenotypes. Control animals were moved into the room with the shakers, but cages were not agitated. Hydrogel was provided instead of water bottles during orbital shaking to prevent spillage in ELSD cages. Hydrogel was equivalently provided in Control cages. As described by our previous study, ELSD is a gentle sleep disturbance method that predominantly affects infant REM sleep while preserving parental care and hormonal markers of stress.

 Cohabitation recording

Individuals emerging from the housing and sleep manipulations described above were submitted to cohabitation recordings during adulthood (26.5 +/- 6.9 weeks of age, mean +/- standard error). Each adult was assigned to an opposite-sex mate. Pairs were formed randomly with all possible sex vs. sleep manipulation combinations: Male-ELSD/Female-ELSD, Male-ELSD/Female-Ctrl, Male-Ctrl/Female-ELSD, Male-Ctrl/Female-Ctrl. A sexually naïve male and female were then placed in bedded home cages (48.3 cm length, 25.4 cm width, 20.3 cm height), but separated from each other by a lab-made mesh divider, meaning that individuals could only roam within the confinements of their quadrants. Cage dividers were made with metal wire mesh (square mesh, 6.5 mm aperture, 22 cm wide, 28 cm high) covered on both sides with a plastic sheet (clear polycarbonate, 0.5 mm thick) to prevent animals from climbing. The bottom rectangular portion of the mesh barrier was left exposed without the plastic sheet (6.5 cm high), allowing animals to exchange bedding and sniff each other. Crocs with chow/gel were placed uniformly across quadrants, with chow and gel being always positioned away from the divider, on the left and right sides of the animal, respectively. No environmental enrichment objects were placed in the quadrants, thus forcing animals to interact with each other or the mesh divider. Finally, we placed cardboard barriers between neighbouring cages, preventing male-female pairs from distracting each other. 

Animals were allowed to behave freely in their quadrants while being filmed from a 90° overhead angle for 72 h. For video recording, we customized a vibration-free optomechanical assembly (ThorLabs) holding two infrared-sensitive cameras (specifications below), each camera surrounded by four infrared illuminators (made in the lab from inexpensive LED boards). Two mesh-divided home cages were placed below each camera, allowing the recording of four male-female pairs at once. This system was installed in a housing-approved room with circadian light/dark switching. Light/dark switching was innocuous to video brightness, as imaging was obtained with infrared reflectance.

We used two grayscale cameras (Basler, acA1300-60gm), each one attached to a fixed focal length lens (Edmund Optics, 6 mm UC Series). The cameras communicated via a network adapter (Intel Pro 1000/PT) with a host computer, and videos were written into an array of hard disks (RAID) for protected data storage. Cameras were configured in Pylon software (Basler) with 8-bit depth, 800-pixel frame width, 896-pixel frame height (no binning), 1328 kbps, and 20-Hz frame rate. Exposure and brightness were adjusted by manipulating the lenses and infrared illuminators, without further adjustments to the camera software. Videos were acquired using StreamPix software (NorPix) into 6-h mp4 segments (H.264 codec) from the two cameras synchronously. Each camera recorded four quadrants, i.e., two male-female pairs. Quadrants were re-framed using Adobe Premiere and re-exported using Adobe Media Encoder, resulting in smaller videos with one individual per video (384-pixel frame width, 416-pixel frame height, 1025 kbps, 25-Hz frame rate, mp4 format, H.264 codec).

Behavioural tracking

Re-framed videos were submitted to marker-less body tracking using DeepLabCut (DLC) v2.2 [26]. We trained the DLC network (resnet_50) to label seven body parts per individual: nose, left/right ears, shoulder, two locations along the back and tail base. For network training, we used manually selected video frames (n = 157) representing a variety of scenarios: from clear imaging of the animal (no motion blur or obstruction of body parts) to challenging situations (e.g., with motion blur, curled posture when sleeping or eating, tail base hidden under bedding, nose hidden by the mesh divider when sniffing the cage mate) across recordings from representative animals, 3 males and 3 females. We then trained the network overnight using a lab server (operating system: CentOS, a Linux distribution. CPU: Intel Xeon E5-2640 v3 @ 2.60 GHz. RAM: 512 GB. GPU: NVIDIA GeForce GTX 1080 Ti). See also Movie S1.

Data analysis

Body part coordinates per video frame were saved as CSV files, imported to Matlab (MathWorks), and converted from pixels to centimetres. We then obtained three measures per video frame explained as follows. (1) Body direction relative to the divider: we fitted a line to the sequence of coordinates from tail to nose and determined the angle of the fitted line relative to the horizontal plane of the video frame. The angle was then rescaled from -1 (opposite from the divider) to +1 (toward the divider), with left and right directions treated equally. By doing this we eliminated the circular dimension, i.e., we ignored information on clockwise or counterclockwise body rotations, which were outside of our scope. This resulted in a straightforward measure of “interest” in the partner vole through the mesh divider. (2) Distance to divider: length between shoulder and divider on the horizontal axis of the video frame, regardless of the position on the vertical axis. In the absence of unrestrained touch between animals, this measure provided a measure of proximity. (3) Locomotion speed: difference (hypotenuse) between the shoulder coordinates of frame n and frame n+1. See also Movie S1.

All measures were timestamped per video frame according to clock time in number of seconds x number of frames per second. For example, the first frame after 8:00 am on Day 1 of recording was identified as 720001 (8 h * 60 min * 60 s * 25 frames + 1 frame). Timestamping was made without restarting the clock at midnight so that each frame could have a unique identifier across the 72-h recording. These timestamping procedures were used to align all recordings onto a common 72 h axis, given that all recordings were intentionally made with 15-30 min margins for later trimming. Clock time information was obtained from the file naming system of the video acquisition software (StreamPix, NorPix).

Body direction relative to the divider, distance to the divider, and locomotion speed data per individual were averaged into 1 h bins or 20 min bins separated into the three recording days. In either case, binned data were submitted to statistical comparisons between sexes or ELSD treatments per sex (two-way repeated measures ANOVA, followed by Tukey’s post-hoc comparisons at each time bin). The same data were averaged across 24-hour periods or light/dark periods of Day 1 and submitted to one-way ANOVA.

The three behavioural measures were also examined for ultradian periodicity (biological cycles in the scale of hours), an exploratory approach to evaluate if the above behaviours oscillate in an ultradian manner. Data were resampled from the original video frame rate to 1 s bins and analysed using Welch’s power spectral density (PSD) estimate (6-h Hamming windows, frequency range of 0-3 cycles per hour in steps of 0.03 cycle). By examining the PSD curves, we found that peaks were mostly prominent within the 0.1-0.6 cycle/h band, which we interpreted to represent the ultradian fluctuations we observed in the raw data. Thus, we summed PSD values within the 0.1-0.6 cycle/h band per individual and did the same across behavioural variables and time periods. Differences between groups and sexes were examined using two-way repeated measures ANOVA, followed by Tukey’s post-hoc comparisons at each period. We additionally subtracted ultradian power curves by their moving averages with a sliding window of 3 data points (i.e., 3 periods) on a per-animal basis. This resulted in curves with magnified light/dark alternation, representing circadian fluctuations. Data from light and dark periods were then separately averaged. Within-sex comparisons per light or dark period were made using one-way ANOVA.

Finally, using data in 1 s bins, we created spatial maps depicting both area occupancy and body direction. A 100 x 100 cell array was created in Matlab to represent a 2 mm grid of the home cage quadrant (quadrant dimensions: 20 x 20 cm). The cell array was cumulatively populated with body direction values across time bins, according to the animal’s position at each time bin. We then averaged the values per cell, which resulted in the maps. Six maps were created per individual, corresponding to the light/dark periods of Days 1-3. Such maps were arranged three-dimensionally, and Z scored across the 3rd dimension within males or females. We then averaged each map vertically to obtain body direction vs. distance to divider curves. These curves were submitted to within-sex statistical comparisons (two-way repeated measures ANOVA, followed by Tukey’s post-hoc comparisons per spatial bin).

Proprietary: Matlab (Mathworks). Open-source alternative: Python.