Wake-like skin patterning and neural activity during octopus sleep – Nature

Experimental animals

All research and animal care procedures were carried out in accordance with institutional guidelines, approved by the OIST Animal Care and Use Committee under approval numbers 2019-244-6 and 2022-364. Adult octopuses (O. laqueus, mantle length roughly 3 cm) of both sexes were collected in Okinawan tidal pools and housed in 12 l tanks connected to a seawater system with open circulation to the ocean. Animals were provided with an enriched environment including sand, plants, rocks and coral rubble, as well as a shelter (terracotta pot).

O. laqueus were carefully selected for this study after assessing several other options due to (1) their compact brain and body size made them suitable for Neuropixels recording and light-sheet imaging, (2) their white resting skin pattern aided detection of AS bouts, (3) their nocturnal behaviour meant we could film sleep behaviour under white lighting and (4) they were locally available, a regulatory requirement for keeping in the OIST marine station. The brain of O. laqueus resembles coastal diurnal octopuses in possessing a seven-gyrus VL (Extended Data Fig. 3d). The VL occupies 9.05% of central brain volume, slightly higher than other coastal nocturnal species, such as the commonly studied O. vulgaris and O. bimaculoides⁵¹.

Behavioural filming

Experiments were conducted in closed seawater systems, circulating filtered natural seawater. Water was filtered mechanically and biologically, UV sterilized, oxygenated and exchanged with fresh seawater twice a week. Unless otherwise reported, temperature was cooled to 22 °C, with lighting alternating on a 12/12-h light/dark cycle with a 30-min taper in light intensity. Animals were fed live crabs three times a week during subjective night, while awake. Animals were given a 2-week acclimation period of living in experimental tanks. Experiments started after animals were seen to be resting normally during the daytime.

Low-resolution recordings (Fig. 1b,g–j and Extended Data Figs. 2 and 5b) were filmed using three custom filming chambers. Each chamber placed a single 4K camera (Basler ace acA4024-29uc, 4,024 × 3,036 pixels, 24 fps) viewing four transparent acrylic 100 × 150 × 100 mm tanks from the top, with 13.4 pixel per mm resolution using 12-mm lenses. Lighting was mounted on two sides of the group of four tanks, with white LED day lighting (Koval Smart Aquarium Light, 300 mm, 654 lx) and red LED night lighting (Leimac IDBA-HM300R, 300 × 40 mm Barlight, 129 lx). Other than during feeding, a 5-mm-thick glass cover was placed over the tanks to prevent animal escape. For recording animal movements (Fig. 1c–f and Extended Data Fig. 1), we placed animals in transparent acrylic 300 × 200 × 200 mm tanks fitted with shelters (three-dimensionally printed and terracotta pots), with shelter entrance facing the tank wall. Cameras (as above, 24 or 30 fps) were positioned facing the shelters of single animals. Lights (Leimac IDBA-HM300W, 300 × 40 mm, 3 klx) were placed on tank top and/or sides.

Recordings were made with PylonRecorder2 (v.0.6), using online hardware compression to h264 format and writing to solid-state drives. A single computer running Windows 10 was equipped with two graphics cards (Nvidia Quadro P1000 and Quadro P5000), which run up to seven cameras simultaneously.

High-resolution recordings (Figs. 1a and 5) were made by placing an 8K camera (Canon EOS R5, 8,192 × 5,464 pixels, 30 fps) fitted with a Canon Macro Lens EF 180 mm lens on a gimbal, and filming top-down on a single animal using a 45° mirror (96 pixels per mm). Three white bar lights (Leimac IDBA-HM300W, 300 × 40 mm, 3,933 lx) were used to light the tank. All recordings were made using the following camera settings: 1/200, F32, ISO 3,200. Because AS events are periodic, AS recording was started a few minutes before the expected AS time, and ended after AS pattern dynamics ceased. Some pairs of waking and sleeping patterns were shot at 4K resolution (Basler ace acA4112-30uc, 4,096 × 3,000 pixels, 30 fps, 32.3 pixels per mm, 1,000 lx), looking down on animals using a 50-mm lens. High-resolution recordings were shot at 24 °C (room temperature water).

To put lighting levels in context, daylight ranges from 10 to 25 klx when sunny, 1,000 lx when overcast. Nighttime light levels range from roughly 0.3 lx during a full moon to 0.002 lx when the moon is not visible. Animal sleep time, AS duration and interval seemed normal under various experimental lighting conditions (Extended Data Fig. 5). Darker skin patterns did not appear under 3 klx lighting while awake, but were observed under 1 klx lighting.

Behavioural experiments

For measuring arousal threshold, mechanical stimulation was delivered using a solenoid, fixed in a constant position on the tank wall and controlled by an Arduino (Arduino Mega 2560). Synchronization of camera and solenoid hit time was done using a red LED placed in view of the camera and out of view of the animal. Stimulus strength was calibrated using a hydrophone (DolphinEar DE200), placed in the position of the octopus. Three strengths of stimulus 6, 40 and 86 dbV were delivered to the octopus during active bouts, QS and wake in the terracotta pot. QS was defined as the interval between two active bouts, in which the octopus had the characteristic lack of movement, flat posture, smooth texture, closed eyes and white colour. All trials were conducted between 12:00 and 17:00.

For sleep homeostasis, octopuses were recorded continuously for 48 h before the start of sleep deprivation. The following day, animals were kept awake from 07:00 to 17:00 by gently brushing their skin with a paintbrush every 2–3 min. Movement, elevated posture and eye opening were considered indicators of wakefulness. Animals were left to behave freely after 17:00. Sleep deprivation was then repeated, using the above procedure, for a second day. Post-deprivation behaviour was subsequently recorded for 48 h without any disturbance. For the active bout interruption experiment, a random subset of active bouts were interrupted using a paintbrush. The amount of QS preceding the following AS bout was then compared in interrupted and uninterrupted ‘trials’.

For measuring circadian rhythm, octopuses were acclimated to a 12-h light/dark cycle for 2 weeks. After 48 h of continuous filming, they were subject to either 72 h of continuous daylight or 72 h of continuous darkness. Feeding was halted during this period to prevent cues from live crabs. Animals were then switched back to a 12-h light/dark cycle.

For measuring temperature dependence, the temperature of the water circulating in the behavioural filming system was cooled to 19, 22 and 24 °C using a cooler (Poafamx AL-300) attached to the water circulation system. Each temperature was maintained for 2 days. Higher temperatures were achieved by turning off the cooler. Tank water temperature was measured using a temperature sensor (Tinytag Aquatic 2 TG-4100).

Surgery for electrophysiology

Animals were anaesthetized (2% ethanol in filtered sea water) more than 3 days before the recordings, and the distal 2–4 cm of all arms were surgically shortened to prevent them from removing future head fixation. Wounds were sealed with tissue glue (Histoacryl, B.Braun). Following this procedure, animals recovered in an experimental tank equipped with a closed seawater circulation system (above), in which they could move and eat immediately on waking from anaesthesia. Octopus did not demonstrate signs of pain (that is, arm grooming behaviour⁵²), therefore local analgesia was not applied, preventing administration related stress and sleep disruption. One day before recording, animals were anaesthetized as above and the top of the head was placed out of the water. The skin, soft tissue and muscle over the cartilaginous head capsule were removed with microscissors. A three-dimensionally printed plastic ring with a pole was glued to the head capsule over the central brain with tissue glue and dental light cure adhesive (3M, Transbond XT Light Cure Paste Adhesive). The inside of the ring was filled with silicone sealant (Kwik-Cast, WPI) to prevent the wound from touching sea water. Animals were left to recover overnight in the recording tank. On the day of recording, animals were anaesthetized as above. A small hole was cut into the head capsule to expose the central brain, and the sheath surrounding the brain was removed using fine forceps. The ring was covered again using silicone sealant. Fresh seawater was washed over the animal, before moving it to the experimental tank. It was then head-fixed using a metal rod. Animals recovered from anaesthesia within minutes of exposure to fresh seawater. The water level in the recording tank was reduced to have the top of the animal’s head above the water line during probe insertion. The animal’s body was supported with an aquarium wool filter mat. A Neuropixels v.1.0 probe shank was coated with CM-DiI (100 μg to 50 μl ethanol, Thermo Fisher Scientific, CellTracker CM-DiI Dye) for post hoc probe position localization. The probe was mounted on a motorized micromanipulator (New Scale Technologies, M3-LS). After removing the cured silicone sealant, the Neuropixels probe was lowered into the brain at the speed of 200 μm min⁻¹. The depth of the recordings varied across experiments. We explored a range of depths, insertion sites and angles. After lowering, the inside of the head-fixation plastic ring was again filled with silicone sealant and water level was raised so that the animal’s head was submerged. The recording tank circulated aerated seawater at a rate of roughly 0.2 l min⁻¹ at room temperature (24 °C).

Electrophysiological recording

Neuropixels recordings were performed using SpikeGLX software (v.3.0). with sampling rates at 2.5 kHz (for LFP signals) and 30 kHz (for extracellular spike signals). The mantle of the animal was simultaneously filmed using a 4K camera as in behavioural recordings (Behavioural filming above, shot at 1,034 lx). Video was synchronized to electrophysiological recording by sending a 25-Hz transistor-transistor-logic (TTL) signal from an Arduino to trigger camera frame exposure and to log TTL time using spikeGLX. Animals showed AS 12.7 ± 6.6 (s.d.) h after recordings began, demonstrating AS bouts similar in duration and interval to that of freely behaving animals (Extended Data Fig. 5).

Tissue clearing and light-sheet imaging

Following recordings, the Neuropixels probe was removed from the brain and the animal was euthanized by gradually increasing ethanol concentration from 2 to 5% in sea water. The head of the recorded animal was dissected, with tissues surrounding the head capsule removed as much as possible. The brain in the head capsule was fixed in 4% paraformaldehyde at 4 °C for 24–48 h. Dissected brain tissue was cleared using a second-generation CUBIC method⁵³. First, the tissue was incubated in 50% CUBIC-L solution at 25 °C overnight, followed by 100% CUBIC-L incubation at 25 °C for 24 h. After PBS wash, the tissue was immersed in BOBO-1 nuclear staining dye (ThemoFisher B3582; 1/800 dilution) for 3 days. The tissue was washed with PBS, then placed in 50% CUBIC-R overnight followed by 100% CUBIC-R for 24 h (Extended Data Fig. 3a). The cleared sample was embedded in a transparent agarose gel for mounting on a microscope. To scan the brain, we custom-built a light-sheet microscope using the technique of axially swept light-sheet microscopy⁵⁴. The microscope was equipped with a 10X detection objective lens (Olympus XLPLN10XSVMP) and a 10X illumination objective lens (KYOCERA SOC Corporation, CS03-10-30-152). Images were acquired with (x, y, z) = (0.65, 0.65, 2.5) μm resolution. BOBO-1 was imaged with a 488 nm excitation laser and 536/40 nm bandpass filter, whereas CM-DiI was imaged with a 532-nm excitation laser and 593/40-nm bandpass filter (Extended Data Fig. 3b,c).

Brain registration and atlas construction

To construct a reference brain atlas of O. laqueus, we cleared and stained an adult octopus brain following the procedure described above and scanned the brain with (x, y, z) = (0.65, 0.65, 2.5) μm resolution. We cropped the supra-oesophageal mass from the whole brain image and downsampled it to (x, y, z) = (10, 10, 10) μm. We used 3D Slicer⁵⁵ to manually annotate this 3D image, referencing existing anatomical atlases^25,26 (Extended Data Fig. 3d,e). 3D images of individual octopus brains were mapped to this reference atlas using the symmetric image normalization method (SyN) method implemented in the Advanced Normalization Tools (ANTs) library⁵⁶. Before the registration, the BOBO-1 channel was downsampled to (x, y, z) = (10, 10, 10) μm. A binary mask was then created manually using 3D Slicer, masking tissues other than the supra-oesophageal mass. The brain was then mapped to the reference brain by a two-step transformation. First, an affine transformation was computed to roughly align the two brains, using mutual information as a metric function. Second, a non-linear warping was computed using the SyN algorithm with cross-correlation as a metric function. In ANTs command line, the following parameters were used: –transform Affine[0.1] –metric MI[${fix_img},${mov_img},1,128,Regular,0.5] –convergence [1000x1000x1000,1e-5,15] –shrink-factors 8x4x2 –smoothing-sigmas 3x2x1vox –transform SyN[0.1,4.0,0.2] –metric CC[${fix_img},${mov_img},1,6] –convergence [500x500x100x30,1e-6,10] –shrink-factors 8x4x2x1 –smoothing-sigmas 3x2x1x0vox. This generated an affine transformation matrix and a warp field given as a four-dimensional matrix (Extended Data Fig. 3g). We controlled the parameters in ANTs to prevent excessive warping, which was quantified by the values of Jacobian determinants (Extended Data Fig. 3g). The atlas and the registered brain were overlaid, showing visually precise alignments (Extended Data Fig. 3f). Alignment quality was quantified by computing the voxel-wise normalized cross-correlation⁵⁶ value with window radius of 4 voxels (Extended Data Fig. 3h), which showed positive values in most of the areas and especially high positive values at the boundary between brain lobes. We also generated an average nuclear stained image from n = 9 independently aligned brains (Extended Data Fig. 3h). The lobe structure was maintained, further supporting the quality of our registration.

To analyse the location of a Neuropixels probe, the CM-DiI channel from a 3D brain image was first downsampled to (x, y, z) = (5, 5, 5) μm. The CM-DiI probe track was then manually labelled using 3D Slicer. This labelled track image was smoothed by first skeletonizing the binary object using the morphology.skeletonize_3d function implemented in the scikit-image library, and then fitting the resulting skeleton with a B-spline. The xyz coordinates of Neuropixels probe channels were then mapped to the reference space using the transformation computed above. Finally, each recording channel was assigned a unique region ID on the basis of the atlas region. Determining probe depth from CM-DiI images is sometimes a non-trivial problem due to the spread of the dye by diffusion. Following previous heuristic treatments⁵⁷, after mapping we inspected the characteristic LFP patterns at the boundary of the anatomical regions. If necessary, we shifted the probe location along the depth axis to increase LFP-anatomy correspondence. The automatically determined locations were usually very accurate, and the maximum correction was ten channels (roughly 100 μm).

To divide brain regions into anterior and posterior halves (Fig. 4e inset), we took the most anterior point and posterior point of each brain region as its A–P minimum and maximum values and computed the midpoint between them as (min + max)/2. Channels were divided into anterior or posterior on the basis of whether they were anterior or posterior to the brain region midpoint.

Behavioural analysis

To measure octopus skin brightness in behavioural recordings (Fig. 1), we segmented octopuses from background with the FAIR Detectron2 platform⁵⁸ (v.0.1.3), using a pretrained base model (COCO Instance Segmentation with Mask R-CNN, R50-FPN, 3× schedule), fine-tuned with octopus training datasets. Training set labelling was done using the Labelbox platform. To accelerate data processing, octopuses were segmented every 100 frames (4.16 s), with mean intensity calculated on every frame using the nearest preceding segmentation. Multi-day videos were processed in parallel by using FFMPEG to cut videos into 1-h clips, running each clip independently, then recombining.

Colour flashes during QS were detected using the mean skin brightness trace during QS. The trace was filtered 0.005–2 Hz using a three-pole Butterworth filter, and peaks were detected on the negative of the z-scored signal, with a minimum peak height of 2, minimum prominence of 0.5 and minimum separation in time of 10 s. Colour flash duration was calculated by taking a window of the mean skin brightness trace from 500 frames (21 s) before a colour flash peak to 1,000 frames (42 s) after a peak. This time series was z-scored, and threshold crossings (2z) on either side of the peak time were taken as the start and end time.

For sleep time, duration and interval, AS and QS times were identified manually using the mean skin brightness recording coupled with video confirmation. Wake times were similarly identified manually. AS inter-event intervals were calculated between bouts in which the animal did not wake up. AS duration was determined by considering a window 10 s before and 100 s after AS start times. These time series were z-scored, and low pass filtered at 0.1 Hz using a two-pole Butterworth filter. The length of the largest continuous stretch of data falling below a threshold was taken as the AS duration. We explored a range of thresholds (Extended Data Fig. 5d), deciding on 0.2 as a good subjective match to video data.

Histograms in Fig. 1i,j used 2-h binning, and a continuous rate estimate was calculated by smoothing a 0.1-h binning using a Gaussian kernel with s.d. 40 bins (4 h). Figure 2b was calculated similarly, using 1-h histogram binning. Figure 2d used 5-min binning, a probability density was estimated using a kernel density estimate (MATLAB ‘ksdensity’).

For movement analysis, in the arousal threshold experiments (Fig. 2a and Extended Data Fig. 1g–i), we extracted clips 1 s before stimulation time to 1 s after. Animals were segmented from background (as above) on the first frame of a clip. Within the segmentation mask, prominent features were detected using SIFT⁵⁹ keypoint detection (contrast threshold of 0.05). Lukas–Kanade optical flow⁶⁰ (window size of 512 pixels) was then used to track these points over frames. Movement magnitude was calculated as the mean optic flow magnitude between neighbouring frames. For calculating reactionary movements, a baseline mean magnitude for 25 frames (1 s) before stimulation time was subtracted from the mean magnitude for 25 frames (1 s) following stimulation.

For analysis of animal movements during QS and AS (Fig. 1c–f and Extended Data Fig. 1a–c), we extracted 2-min clips centred on AS start times. The eye, body and the anterior mantle (for measuring breathing) were manually segmented from the first frame of this clip. Movement magnitude was calculated as above, separately for each segmentation mask. To isolate eye and breathing movements from overall body movements, the average movement within the eye/anterior mantle mask was removed from each frame and we have reported residual movements. Breathing rate was extracted from anterior mantle residual movements, with inhalation detected through peak detection in the z-scored, smoothed (10 frames) trace, with a peak prominence of 0.05. Breathing rate was then linearly interpolated to video frame rate. Figure 1 reports average movement magnitude for the first 30 s of the clip (QS), and the third 30 s of the clip (AS). Calculation of breathing arythmicity for waking animals (Extended Data Fig. 1b) was calculated as QS/AS, on separate 30-s video clips.

Behavioural analysis was performed using OIST’s Saion HPC system, using up to 32 GPUs (Nvidia V100 and P100s). Core analysis was written in Python (v.3.6 and 3.7), with further analysis written using MATLAB 2019a.

Electrophysiological analysis

LFP data were preprocessed by resampling from 2.5 to 1 kHz, filtering 0.1–150 Hz and re-referencing by subtracting the median of ten channels located out of the brain from all channels. Spectrograms were calculated using a continuous wavelet transform with a Mortlet wavelet (MATLAB ‘cmor1.5-1’), scales logarithmically spaced between 1 and 100 Hz. Spectrograms were normalized in amplitude by dividing all values by the maximum value. Spectra were calculated on non-overlapping 1-s chunks of data using the Chronux toolbox (v.2.12, http://chronux.org/)⁶¹ with a time-bandwidth product of five and nine tapers. The results were then averaged over data chunks.

For calculating channel intensity during different behavioural states, a uniform procedure was conducted on different selections of data. For AS, QS and wakefulness, 60 s of LFP data were loaded, beginning at the transition of every detected AS or wake phase and taking the 60 s before AS times as QS. For QS colour flashes, 700-ms chunks of data were loaded, centred on colouration flash events (detected as above). After data loading, two filtered versions of the data were then generated, at 0.1–10 and 20–150 Hz for every recording channel. The envelope of these filtered signals was calculated using a 150-tap Hilbert filter. Signal strength for a channel was calculated as the mean of this envelope. We median filtered more than five channels to remove noisy channel readings. We then averaged this vector over all events.

For QS oscillation events, 1,200 s of LFP data were loaded preceding every AS bout and any wake events (manually detected, above) were removed. QS oscillation events were detected by filtering the data 4–40 Hz, then finding peaks in the z-scored signal with a minimum height of two, minimum prominence of two and a minimum separation of 1 s. Estimates of oscillation event rates were taken per data chunk and averaged. This was then smoothed with a five-channel median filter, as in activity strength measurements. To calculate LFP activity strength over brain regions, we averaged the channel intensity for all electrodes located within a brain region. We required data from a minimum of two probes to consider activity strength for a brain region. Correlations between different activity strength measurements were taken over all electrodes from these brain regions. Unless otherwise stated, filtering was performed using three-pole Butterworth filters.

Skin pattern analysis

To quantify AS skin patterns, we adapted techniques developed for describing cuttlefish camouflage²⁸. High-resolution octopus videos were processed by first detecting the octopus mantle, using the Detectron2 platform⁵⁸ as above. Mantles were aligned by choosing a single source image and mapping all images onto this source image by ellipse fitting and similarity transformation. Determination of anterior versus posterior direction was done manually for waking images and the first frame of AS video clips. Images were then cropped and downsampled to 1,004 × 675 pixels (20% image size after cropping to segmented mantle), with background pixels coloured uniform grey (as in Fig. 4c). Following standard preprocessing (zero-centring), 400 × 400 pixel crops of the dorsal mantle were evaluated by a VGG-19 (ref. ⁶²) network pretrained on the ImageNet⁶³ database, using the Keras⁶⁴ platform (included in TensorFlow v.2.0). We used the max-pooled fifth layer activations (‘block5_pool’) as our skin pattern metric⁶⁵. This resulted in 512-dimensional vectors describing skin patterns for every frame in a video clip.

The starting points of AS trajectory dynamics were aligned across video clips by calculating the first principal component of the 512-feature by frame matrix, and thresholding the absolute value of its approximate derivative (difference between neighbouring points in time, threshold 0.1). All further analysis was done on start-time aligned trajectories. To estimate the dimensionality of AS space, we ran Parallel Analysis⁶⁶ several times on 10,000 randomly selected images from the total dataset. To estimate the overlap of different animals’ patterns within AS space, we similarly calculated the Silhouette score⁶⁷ several times on 10,000 randomly selected images from the total dataset. Intra-trajectory distance was calculated as the mean Euclidean distance between neighbouring points in time along a trajectory. Inter-trajectory distance was calculated between two trajectories as the mean element-wise Euclidean distance, from trajectory start time to the end of the shorter trajectory. Distances between trajectories after dynamic time warping⁶⁸ were divided by trajectory distance to compare with non-time warped inter-trajectory distances. Nearest inter-trajectory distances (between AS bouts and between waking images and AS bouts) were calculated by taking, for all pairs of trajectories, the minimum Euclidean distance between points.

For aligning select waking and AS skin patterns, similar patterns were extracted from videos manually. Precise alignment was achieved by manually selecting corresponding points and interpolating between these points using a moving least-squares algorithm⁶⁹ to produce a mapping from the AS image to the waking image. Images were uniformly and linearly brightened for display. Green rectangles in AS images (Fig. 5d and Extended Data Fig. 10) are approximate crops (non-linear mapping). To show overlaid matches, images were grey scaled, inverted and thresholded (image specific threshold of 8-bit grayscale at 180–230) to show the dark pattern regions.

Statistics and reproducibility

Unless stated otherwise, data are mean ± s.e.m. For box plots, margins are 25th and 75th percentiles; middle line, median; whiskers, boundaries before outliers; outliers (+) are values beyond 1.5× interquartile range from the box margins. Experiments were repeated independently several times with similar results, with numbers of repetitions and sex (female or male) as follows: temperature modulation n = 9 animals, arousal threshold n = 5 animals, homeostasis n = 6 animals, active bout movement n = 3 animals, continuous light on/off n = 6 animals, electrophysiology n = 9 animals, skin pattern dynamics n = 3 animals and wake–sleep pattern matching n = 5 animals.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.