INTRODUCTION
Pain assessment in humans and animals is critical for the development and monitoring of new pain therapeutics. While pain is inherently subjective to the individual experiencing it, current assessment techniques rely on verbal responses and questionnaires. Pain assessment can be even more problematic for individuals that lack the ability to communicate effectively, such as children and patients with neurodevelopmental disorders [5; 12]. The difficulty in translating pain assessment tools to the bench and back to humans extends to limitations on the use of mice as an experimental platform for understanding pain. To overcome these shortcomings, researchers have sought continuous objective measures of pain through the analysis of tissue biomarkers, electrical potentials, imaging procedures and assessment of muscle tension [7; 24]. These approaches are rational, but often lack correlation with other signs of pain and recording of these components in a clinical setting is often difficult. An exciting advancement in rodent pain assessment is the recent development of automated high-speed photographic monitoring of paw withdrawal touch-based assays, which revealed time-resolved features previously missed by manual monitoring [11]. However, translation of a paw withdrawal assay to patients is not feasible. The measurement of dynamic changes of facial features activated during a grimace response has been recognized as an alternative, viable objective biomarker for pain that can be observed across species, including humans [17].
Manual scoring of facial features to quantify grimace in laboratory animals is effective but is time intensive, requires training of graders to reduce subjectivity, and is limited to a noncontinuous three-level scale of severity (0, 1, 2) [13; 22]. Despite the challenges of facial grimace scoring, the assay has proven quite useful and highly translatable to humans [24; 28]. To address these challenges, video-based capture of facial images have been manually scored [13] or scored by trained convoluted neural networks for the presence or absence of facial pain [25]. However, many of these platforms rely on methods that can only detect extremes in facial pain, making it difficult to identify small changes in pain severity.
In this study, we sought to validate an automated video-based squint assay to measure pain by leveraging our previous finding that orbital tightening, or squint, is the principal component of the mouse grimace score [22]. The automated squint assay was validated using a well-known pain trigger, formalin, which has been reported to induce grimace [18]. We then compared automated measurements with manual measurements of grimace and orbital tightening in response to the neuropeptide CGRP, which is known to cause squinting and other components of grimace in mice [22]. Finally, the automated assay was used to test the CGRP-related peptide amylin, which has recently been shown to induce migraine in patients and migraine-like symptoms in mice [6].
MATERIALS AND METHODS
Animals
Wildtype C57BL/6J black (https://www.jax.org/strain/000664) and CD1 white mice (http://www.criver.com/products-services/basic-research/find-a-model/cd-1-mouse) were used for development of the automated squint model. In the formalin squint assay, cohorts of 10 C57BL/6J mice (5 males, 5 females) were used for both formalin and vehicle groups. In the CGRP dose response squint assay, cohorts of 20 C57BL/6J mice were used for each CGRP dose (10 males, 10 females) and vehicle (11 males, 9 females). For the amylin squint assay, amylin treated C57BL/6J (11 males, 10 females) and vehicle (8 males, 8 females) cohorts were used. Mice were 10–14 weeks old, with an average weight of 26 g for C57BL6/J mice and 30 g for CD1 mice. All strains of mice were housed in a temperature-controlled vivarium on a 12-hour light cycle with food and water ad libitum. C57BL/6J mice were housed in groups of 5, CD1 mice were housed in groups of 4. All behavioral experiments were performed between 8:00 A.M. and 5:00 P.M. after a minimum of 1-week habituation in the animal facility. All procedures followed the ARRIVE guidelines and were approved by the University of Iowa Institutional Animal Care and Use Committee and implemented in accordance with the standards set by the National Institutes of Health.
Drug administration
Rat α-CGRP (Sigma-Aldrich, St. Louis, MO) was diluted in Dulbecco phosphate buffered saline (PBS) (HyClone) and administered via intraperitoneal (IP) injection with a 30 g × 0.5-inch needle in the following quantities: 0.01 mg/kg, 0.05 mg/kg, 0.1 mg/kg, or 10 mL/kg PBS alone as a vehicle. Formalin (Thermo Scientific, Fair Lawn, NJ) was diluted in PBS and administered via subcutaneous (SC) intraplantar injection into the right hind paw with a 31 g × 15/16-inch needle (4%, 10 μL total volume) or PBS alone as a vehicle. Rat amylin (Bachem, Torrance, CA) was diluted in PBS and administered via IP injection at 0.5 mg/kg or 10 mL/kg PBS alone as a vehicle. Mice were handled gently with use of light anesthesia (isoflurane, 5% induction) for SC intraplantar formalin and vehicle injections as previously described [14]. It is important to note that the mice were anesthetized just long enough to be immobilized. No anesthesia was employed during IP injections. Investigators were blinded to drug treatment with all injections performed by either B.J.R. or L.P.S. Mice injected with CGRP or vehicle were allowed to recover for 30 minutes in their home cage before testing. Mice injected with formalin or vehicle were allowed to recover for 20 minutes in their home cage before testing. Mice injected with amylin or vehicle were allowed to recover for 15 minutes in their home cage before testing.
Development of video image capture for mouse facial detection
Video recordings of C57BL/6J and CD1 mice sampled at 1 frame every 0.1 second (10 frames per second) were used for training the automated facial detection software. Mice were acclimated to a customized gentle collar restraint prior to experimentation as previously described [22] to fix camera distance in order to constrain video magnification changes as well as to decrease struggle and head movement. Acclimation sessions for C57BL/6J and CD1 mice were 20 minutes each for 3–4 or 4–5 sessions, respectively. Acclimation was individually determined per mouse by willingness to remain still for at least 10 minutes. Mice that did not acclimate to the custom restraint were excluded. Eight synchronized cameras were employed from varying vantage points for each recording session (IDS Imaging UI-3240ML-NIR, Obersulm, Germany) with infrared light to visualize the face and ensure landmark detection with cameras having a Kowa LM35JC 2/3” 35 mm F1.6 manual iris c-mount lens (Kowa American Corp., Torrance, CA) with a focal distance of 254 mm and aperture adjusted accordingly. A custom graphical software tool was used to manually locate and place a predetermined set of 20 facial anatomical landmarks on video frames displaying varying degrees of mouse expression. Following facial landmarking, a proprietary algorithmic approach (FaceX, LLC, Iowa City, IA) was used to derive these landmarks in newly recorded video frames and identify the face and eye. The software incorporated a combination of deep learning and shape regression models to detect, align, and label facial and eye landmarks. The software-reported accuracy of the landmark placement was made by a secondary shape regression model that estimated a tracking error rate as the root-mean squared error of the predicted landmarks. After direct comparison of software-reported tracking error and manual assessment of eye landmark alignment by inspecting each frame of forty 5-minute video recordings, individual frames containing a tracking error rate of >15% were excluded. This error rate was where the software tracking error aligned with all manually assessed video recordings. For a completely closed eye, the software reports an area of ~670 pixels, which accounts for the eyelids. This area was not subtracted from the data.
Automated measurement and analysis of squinting behavior with formalin, CGRP, and amylin
For automated squint analysis, the custom gentle collar restraint and acclimation protocol previously described was employed [22]. Following a 5-minute baseline video recording and squint measurement in room light, mice were given their respective injections previously described (formalin, CGRP, amylin, vehicle) and returned to their home cage for recovery. After recovery, mice were restrained once and recorded at 10 Hz for squint assessment over 5 minutes in room light. Pixel area measurement for the right eye palpebral fissure was derived every 0.1 seconds (10 frames per second) using camera 2 in the recordings. Camera 2 was used since it provided the most perpendicular angle to the right eye for optimal measurement. Utilizing the trained facial detection software, the resulting values were compiled with custom MATLAB script. Individual frames containing a tracking error rate of >15% were excluded. In comparing manual grimace scoring versus automated squint assessment, the frame with largest pixel area delta between baseline and treatment in pixel area for each CGRP cohort was selected and scored using the Mouse Grimace Scale [13] by three blinded individuals. All data is freely available upon request.
Statistical analysis
All statistical analyses were performed using GraphPad Prism 9.2.0 based on data expressed as mean ± SEM. Statistical details are presented in Table 1. Differences in change from baseline to treatment with CGRP and vehicle were determined by 2-way repeated-measure ANOVA: treatment (3 different CGRP concentration groups, 1 vehicle group, 4 groups total) and condition (baseline and treatment) followed by Šídák’s multiple comparisons test to compare treatment effect with respective baselines. Differences in change from baseline to treatment with formalin and vehicle were determined by 2-way repeated-measure ANOVA: treatment (formalin, vehicle) and condition (baseline and treatment) followed by Šídák’s multiple comparisons test to compare treatment effect with respective baselines. Differences in change from baseline to treatment with amylin and vehicle were determined by 2-way repeated-measure ANOVA: treatment (amylin, vehicle, 2 groups total) and condition (baseline and treatment) followed by Šídák’s multiple comparisons test to compare treatment effect with respective baselines. All experiments were repeated with different cohorts in at least two independent sessions.
Table 1:
Statistical Analysis
| Figure # | Analysis | Statistics |
|---|---|---|
| Figure 2B | Two-way repeated measure ANOVA | |
| Interaction factor | F(1,18)=8.166, p=0.0105 | |
| Treatment factor (PBS, Formalin) | F(1,18)=3.530, p=0.0766 | |
| Condition factor (Baseline, Treatment) | F(1,18)=9.731, p=0.0059 | |
| Automated Squint | Šídák’s multiple comparison | |
| - Baseline vs. Treatment for PBS group (n = 10) | p=0.9790 | |
| - Baseline vs. Treatment for formalin group (n = 10) | p=0.0010 | |
| Figure 3B top row left panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(3,76)=9.743, p<0.0001 | |
| Treatment factor (4 treatment groups) | F(3,76)=4.756, p=0.0043 | |
| Condition factor (Baseline, Treatment) | F(1,76)=61.49, p<0.0001 | |
| Automated Squint | Šídák’s multiple comparison | |
| - Baseline vs. Treatment for PBS group (n = 20) | p<0.9999 | |
| - Baseline vs. Treatment for CGRP 0.01 mg/kg group (n = 20) | p=0.0387 | |
| - Baseline vs. Treatment for CGRP 0.05 mg/kg group (n = 20) | p<0.0001 | |
| - Baseline vs. Treatment for CGRP 0.1 mg/kg group (n = 20) | p<0.0001 | |
| Figure 3B top row middle panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(3,76)=12.33, p<0.0001 | |
| Treatment factor (4 treatment groups) | F(3,76)=1.225, p=0.3066 | |
| Condition factor (Baseline, Treatment) | F(1,76)=33.57, p<0.0001 | |
| Mouse Grimace Scale | Šídák’s multiple comparison | |
| - Baseline vs. Treatment for PBS group (n = 20) | p=0.4390 | |
| - Baseline vs. Treatment for CGRP 0.01 mg/kg group (n = 20) | p=0.3196 | |
| - Baseline vs. Treatment for CGRP 0.05 mg/kg group (n = 20) | p<0.0001 | |
| - Baseline vs. Treatment for CGRP 0.1 mg/kg group (n = 20) | p<0.0001 | |
| Figure 3B top row right panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(3,76)=12.70, p<0.0001 | |
| Treatment factor (4 treatment groups) | F(3,76)=1.789, p=0.1564 | |
| Condition factor (Baseline, Treatment) | F(1,76)=36.34, p<0.0001 | |
| Manual Orbital Tightening | Šídák’s multiple comparison | |
| - Baseline vs. Treatment for PBS group (n = 20) | p=0.4030 | |
| - Baseline vs. Treatment for CGRP 0.01 mg/kg group (n = 20) | p=0.1979 | |
| - Baseline vs. Treatment for CGRP 0.05 mg/kg group (n = 20) | p<0.0001 | |
| - Baseline vs. Treatment for CGRP 0.1 mg/kg group (n = 20) | p<0.0001 | |
| Figure 3B middle row left panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(3,37)=12.70, p=0.0028 | |
| Treatment factor (4 treatment groups) | F(3,37)=3.954, p=0.0153 | |
| Condition factor (Baseline, Treatment) | F(1,37)=17.55, p=0.0002 | |
| Male Automated Squint | Šídák’s multiple comparison | |
| - Baseline vs. Treatment for PBS group (n = 11) | p=0.9834 | |
| - Baseline vs. Treatment for CGRP 0.01 mg/kg group (n = 10) | p=0.9338 | |
| - Baseline vs. Treatment for CGRP 0.05 mg/kg group (n = 10) | p=0.0043 | |
| - Baseline vs. Treatment for CGRP 0.1 mg/kg group (n = 10) | p=0.0003 | |
| Figure 3B middle row middle panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(3,37)=10.06, p<0.0001 | |
| Treatment factor (4 treatment groups) | F(3,37)=0.8479, p=0.4766 | |
| Condition factor (Baseline, Treatment) | F(1,37)=12.84, p=0.0010 | |
| Male Mouse Grimace Scale | Šídák’s multiple comparison | |
| - Baseline vs. Treatment for PBS group (n = 11) | p=0.1775 | |
| - Baseline vs. Treatment for CGRP 0.01 mg/kg group (n = 10) | p=0.9810 | |
| - Baseline vs. Treatment for CGRP 0.05 mg/kg group (n = 10) | p=0.0032 | |
| - Baseline vs. Treatment for CGRP 0.1 mg/kg group (n = 10) | p<0.0001 | |
| Figure 3B middle row right panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(3,37)=11.12, p<0.0001 | |
| Treatment factor (4 treatment groups) | F(3,37)=1.990, p=0.1323 | |
| Condition factor (Baseline, Treatment) | F(1,37)=13.73, p=0.0007 | |
| Male Manual Orbital Tightening | Šídák’s multiple comparison | |
| - Baseline vs. Treatment for PBS group (n = 11) | p=0.2086 | |
| - Baseline vs. Treatment for CGRP 0.01 mg/kg group (n = 10) | p=0.9998 | |
| - Baseline vs. Treatment for CGRP 0.05 mg/kg group (n = 10) | p=0.0020 | |
| - Baseline vs. Treatment for CGRP 0.1 mg/kg group (n = 10) | p<0.0001 | |
| Figure 3B bottom row left panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(3,35)=4.325, p=0.0107 | |
| Treatment factor (4 treatment groups) | F(3,35)=3.512, p=0.0251 | |
| Condition factor (Baseline, Treatment) | F(1,35)=50.14, p<0.0001 | |
| Female Automated Squint | Šídák’s multiple comparison | |
| - Baseline vs. Treatment for PBS group (n = 9) | p=0.9293 | |
| - Baseline vs. Treatment for CGRP 0.01 mg/kg group (n = 10) | p=0.0131 | |
| - Baseline vs. Treatment for CGRP 0.05 mg/kg group (n = 10) | p<0.0001 | |
| - Baseline vs. Treatment for CGRP 0.1 mg/kg group (n = 10) | p<0.0001 | |
| Figure 3B bottom row middle panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(3,35)=4.225, p=0.0119 | |
| Treatment factor (4 treatment groups) | F(3,35)=1.321, p=0.2831 | |
| Condition factor (Baseline, Treatment) | F(1,35)=20.64, p<0.0001 | |
| Female Mouse Grimace Scale | Šídák’s multiple comparison | |
| - Baseline vs. Treatment for PBS group (n = 9) | p=0.9899 | |
| - Baseline vs. Treatment for CGRP 0.01 mg/kg group (n = 10) | p=0.3331 | |
| - Baseline vs. Treatment for CGRP 0.05 mg/kg group (n = 10) | p=0.0022 | |
| - Baseline vs. Treatment for CGRP 0.1 mg/kg group (n = 10) | p=0.0009 | |
| Figure 3B bottom row right panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(3,35)=3.861, p=0.0174 | |
| Treatment factor (4 treatment groups) | F(3,35)=1.273, p=0.2986 | |
| Condition factor (Baseline, Treatment) | F(1,35)=22.94, p<0.0001 | |
| Female Manual Orbital Tightening | Šídák’s multiple comparison | |
| - Baseline vs. Treatment for PBS group (n = 9) | p=0.9913 | |
| - Baseline vs. Treatment for CGRP 0.01 mg/kg group (n = 10) | p=0.0894 | |
| - Baseline vs. Treatment for CGRP 0.05 mg/kg group (n = 10) | p=0.0009 | |
| - Baseline vs. Treatment for CGRP 0.1 mg/kg group (n = 10) | p=0.0036 | |
| Figure 4B left panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(1,35)=2.133, p=0.1531 | |
| Treatment factor (PBS, Amylin) | F(1,35)=3.380, p=0.0745 | |
| Condition factor (Baseline, Treatment) | F(1,35)=7.149, p=0.0113 | |
| Automated Squint | Šídák’s multiple comparison | |
| - Baseline vs. Treatment for PBS group (n = 16) | p=0.6706 | |
| - Baseline vs. Treatment for Amylin 0.5 mg/kg group (n = 21) | p=0.0068 | |
| Figure 4B middle panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(1,17)=0.1064, p=0.7483 | |
| Treatment factor (PBS, Amylin) | F(1,17)=2.188, p=0.1574 | |
| Condition factor (Baseline, Treatment) | F(1,17)=2.208, p=0.1556 | |
| Male Automated Squint | Šídák’s multiple comparison | |
| - Male Baseline vs. Treatment for PBS group (n = 8) | p=0.7045 | |
| - Male Baseline vs. Treatment for Amylin 0.5 mg/kg group (n = 11) | p=0.3286 | |
| Figure 4B right panel | Two-way repeated measure ANOVA | |
| Interaction factor | F(1,16)=3.448, p=0.0818 | |
| Treatment factor (PBS, Amylin) | F(1,16)=4.253, p=0.0558 | |
| Condition factor (Baseline, Treatment) | F(1,16)=5.626, p=0.0306 | |
| Female Automated Squint | Šídák’s multiple comparison | |
| - Female Baseline vs. Treatment for PBS group (n = 8) | p=0.9294 | |
| - Female Baseline vs. Treatment for Amylin 0.5 mg/kg group (n = 10 | p=0.0118 |
RESULTS
Automated squint tracking and analysis
The trained FaceX software algorithm was able to automatically detect the C57BL/6J face using 20 facial landmarks (Fig. 1A, red dots, not all visible) from the 8 synchronized cameras (Fig. 1B). After detection of the face, the software automatically aligned and labelled the facial and eye landmarks and applied the final tracing to the eyelid borders (Fig. 1A, B, green line). By comparing pixel area values with time-stamped videos, we could identify any changes in pixel area resulting from struggle, movement, blinks, and eye opening (Fig. 1C, black arrows). The same analyses were performed with white CD1 mice (Fig. 1D–F). Over 1,000 images were trained for the CD1 (white) mouse model and over 2,200 images for the C57BL/6J (black) mouse model. The C57BL/6J mice required more training images due to the difficulty of discerning a black eyelid border and palpebral fissure against black fur. The 8 camera angles pictured in Fig. 1B, E illustrate the angles chosen for facial landmarking. In this restraint model, camera 2 provided the best angle and position for detection (Fig. 1B, E, red box) and analysis (Fig. 1C, F) of squint and was used for all experiments. These data demonstrated we could accurately track facial features of both black and white mice.
Figure 1: Framework and collection of automated squint data in C57BL/6J (black) and CD1 (white) mice.
(A) C57BL/6J mice were held with a gentle collar restraint to allow denotation of facial landmarks (red dots, not all visible) at varying levels of facial pain presentation. The right frame shows automatically derived fissure narrowing (outlined by green border) in response to CGRP (B) The face was located using object detection software based on the facial landmarks from eight synchronized cameras to ensure the optimum pose was captured for automated squint analysis. A decision forest of binary features along with additional trained regressors located the eye and applied the final optimal tracing of palpebral fissure shown as green borders. When utilizing the gentle collar restraint, camera 2 (red box) provided the best angle to accurately track the right eye and measure pixel area of the palpebral fissure. (C) Pixel area of the right eye palpebral fissure was sampled every 0.1 seconds (10 frames per second) with notable events such as struggle, blink, head turn, and sustained eye opening denoted by black arrows. (D-F) Same as panels A-C except with CD1 mice.
Validation of pain-induced squinting behavior through intraplantar formalin injection in C57BL/6J mice
To strengthen the connection between squinting behavior and pain, we utilized formalin, a commonly used nociceptive stimulus [10]. Formalin induces a multi-phase nociceptive response in mice with an early phase lasting 5 minutes post-injection and a second phase lasting 20 to 30 minutes [10]. Additionally, subcutaneous intraplantar injection of 4–5% formalin has been shown to induce grimace in C57BL/6J mice during the second phase [13; 18]. Using the automated assay, the mean pixel areas over time for baselines and treatment with vehicle or 4% formalin were determined during the second phase (Fig. 2A). We observed a significant squint response to formalin, but not vehicle (Fig. 2B). Male vs female comparison was made with no differences in squinting response (p = 0.9). However, we were not powered enough to make a definitive conclusion.
Figure 2: Intraplantar formalin injection induces significant squinting behavior in C57BL/6J mice.
(A) Mean pixel area over time for each baseline (5-minute recording, no injection, n = 10 for each) and respective treatment (5-minute recording, 20 minutes post-injection) for all mice injected with either vehicle (PBS, 10 μL volume, n = 10) or formalin (4%, 10 μL volume, n = 10). (B) Mean overall pixel area for each baseline (5-minute recording, no injection, n = 10 for each) and respective treatment (5-minute recording, 20 minutes post-injection) for all mice injected with either vehicle (PBS, 10 μL volume, n = 10) or formalin (4%, 10 μL volume, n = 10). Error bars indicate ± SEM. Two way repeated-measures ANOVA followed by Šídák’s multiple-comparison test to compare baseline and treatment conditions, **p = 0.0010.
Automated squint versus manual grimace scoring and orbital tightening of CGRP-treated C57BL/6J mice
Using the automated software, we compared CGRP-induced squint to manual grimace measurements. We have previously reported that CGRP at 0.1 mg/kg causes a grimace and squint response in mice [22] but a dose-effect of grimace responses had not been performed. Analyzed with the automated software, CGRP-induced squint could be visualized over time (Fig. 3A). CGRP-administered mice displayed a significant squint with even the lowest dose of 0.01 mg/kg CGRP (Fig. 3B, top row, left panel). When the manual Mouse Grimace Scale was used to score mice at the point with the largest delta in pixel area between baseline and treatment in the automated assay (Fig. 3A black arrows), we confirmed a grimace at the higher CGRP doses, but not at the lowest dose (0.01 mg/kg CGRP) (Fig. 3B, top row, middle panel). When the grimace orbital tightening unit was analyzed separately from grimace, it did not show a significant difference at 0.01 mg/kg CGRP (Fig. 3B, top row, right panel).
Figure 3: Automated squint analysis detects pain behavior with subthreshold levels of CGRP that Mouse Grimace Scale does not in C57BL/6J mice.
(A) Mean pixel area over time for each baseline (5-minute recording, no injection, n = 20 for each) and respective treatment (5-minute recording, 30 minutes post-injection) for all mice injected with either Veh (PBS, n = 20) or CGRP (0.01 mg/kg, 0.05 mg/kg, or 0.1 mg/kg, n = 20 for each). Black arrows indicate time synchronized frames with the greatest squint area delta between baseline and treatment independently assessed by three blinded individuals to maximize the likelihood of detecting a difference in Mouse Grimace Scale analysis. (B) Mean overall pixel area, mean grimace scores, and mean orbital tightening action unit scores from the grimace analysis from panel A for all mice during baseline (B) and treatment (Tx) conditions. The data are shown for all mice (top row), male only (middle row, n = 11, PBS; n = 10 for each CGRP concentration) and female only (bottom row, n = 9, PBS; n = 10 for each CGRP concentration). Error bars indicate ± SEM. Two way repeated-measures ANOVA followed by Šídák’s multiple-comparison test to compare baseline and treatment conditions, *p < 0.05, **p < 0.005, ***p < 0.001, ****p < 0.0001.
Interestingly, when analyzed by sex, only females, not males, showed a significant automated squint response to 0.01 mg/kg CGRP (Fig. 3B, middle and bottom rows, left panels). Both sexes responded to the 0.05 mg/kg and 0.1 mg/kg CGRP doses when analyzed with automated squint, Mouse Grimace Scale, and orbital tightening (Fig. 3B, middle and bottom rows, middle and right panels).
Automated squint detection of spontaneous pain induced by the peptide hormone amylin
With the validation of the automated squint assay, we measured the response of C57BL/6J mice to amylin, a peptide hormone closely related to CGRP. Human amylin has ~50% sequence identity with CGRP and the two peptides bind a shared receptor (AMY1) with equal affinity [4]. Mean pixel area over time for baselines and respective treatments are shown in Fig. 4A. C57BL/6J mice displayed significant differences in squint with 0.5 mg/kg amylin compared to baseline, while vehicle-treated mice showed no difference in squint response compared to respective baselines (Fig. 4B, left panel). When viewed separately, a sex difference was detected with amylin-induced squint response. The female C57BL/6J mice displayed a significant squint response when compared to respective baselines while the males did not have a detectable response (Fig. 4B, middle and right panels).
Figure 4: Amylin induces facial grimace in female but not male C57BL/6J mice.
(A) Mean pixel area over time for either vehicle (n = 16, PBS, left panel) or amylin (n = 21, 0.5 mg/kg, right panel). (B) Mean overall pixel area scores from panel A for all mice during baseline (B) and treatment (Tx) conditions. The data are shown for all mice (left panel), male only (middle panel, n = 8, PBS; n = 11, amylin), and female only (right panel, n = 8, PBS; n = 10 amylin). Two way repeated-measures ANOVA followed by Šídák’s multiple-comparison test to compare baseline and treatment conditions, *p < 0.05, **p < 0.01.
DISCUSSION
In this report, we describe an automated system for measuring squint as a faster, more efficient readout of facial pain in mice. The automation software was able detect and measure squint activity in both black C57BL/6J mice and white CD1 mice, demonstrating a range of applications with different mouse strains. We first validated the automated assay by detecting a squint response to the well-established nociceptive stimulus formalin, which has previously been reported to cause facial grimace [13; 18]. The squint response to 0.1 mg/kg CGRP (27% decrease from baseline) was comparable to the formalin response (26% decrease from baseline). We then compared automated squint versus manual scoring of CGRP-induced grimace in C57BL/6J mice. There was greater sensitivity with the automated squint assay, which detected responses to a low dose CGRP not seen with manual observations. This was observed only in female mice, not males. We also used the automated assay to document that amylin induces a small but significant squint response in female, but not male, C57BL/6J mice. These results demonstrate the applicability of a new automated method for detection of facial changes and specifically squinting associated with pain.
One limitation is that we chose a formalin dose (4%) based on a previous facial grimace study [18], but did not test lower doses of formalin (0.5%−2%) that have been shown to induce pain phenotypes [3; 10]. Another limitation may be the use of isoflurane to perform intraplantar injections of formalin since prolonged isoflurane (10–45 minutes) was reported to cause grimace in some mice strains not used in this study [9; 16]. However, in our formalin study the vehicle mice given a brief isoflurane exposure (about 30 seconds) did not demonstrate increased squint compared to baseline (Fig. 2B).
The ability of amylin to induce squint in mice builds on the recent report that the amylin analog pramlintide induces migraine-like headache in patients and that amylin induced two migraine-like symptoms of light aversion and cutaneous hypersensitivity in mice [6]. Interestingly, the amylin-induced squint was detected only in female mice. A similar sex bias was seen with light aversion and cutaneous hypersensitivity responses in mice [6], suggesting that amylin receptors might contribute to the female prevalence of migraine. While our experiments were not designed for direct comparison between CGRP and amylin, the relatively small amylin-induced squint response compared to the CGRP-induced response is consistent with observations that pramlintide and amylin also appeared to be less potent than CGRP in inducing migraine-like headaches and symptoms in patients and mice, respectively [6]. In mice, amylin must be given at a 5-fold higher dose (0.5 mg/kg) relative to CGRP (0.1 mg/kg) to induce observable phenotypes in the von Frey and light aversion assays [6; 26]. The same pattern holds true for this study where the squint response to CGRP (0.1 mg/kg, 27% decrease from baseline, 30% for female only) was greater than amylin (0.5 mg/kg, 13% decrease from baseline, 19% for female only). Thus, at present there appears to be a relationship between the magnitude of the CGRP and amylin-induced squint responses, hind paw hypersensitivity, and light-aversive behaviors. Future studies are needed to ascertain whether the difference in efficacy of CGRP and amylin might reflect differences in receptors i.e., CGRP binds the canonical CGRP receptor to a greater degree than amylin [8].
Our results highlight the use of the eye as an accurate measurement of spontaneous pain in a CGRP or amylin-evoked model. When compared directly to the facial grimace score, the automated squint detection was able to identify smaller changes than the human scored Mouse Grimace Assay. Of importance, the automated analysis enabled the detection of a sex difference, with females showing a response to low dose CGRP in our mouse model that was not detected manually with the Mouse Grimace Scale. This is consistent with our previous study that found orbital tightening to be the principal component in mouse facial grimace [22].
The development of squinting behavior for the assessment of pain may represent a step forward in pre-clinical pain assessment. However, like the facial grimace scale, squinting behavior may not always represent pain [22]. In fact, the squinting behavior used as the major tool in this study is only one component of the facial grimace assay. On the other hand, squinting and facial grimace may not occur with all pain states, e.g. grimace was not observed with chronic pain [13]. Nonetheless, our previous study demonstrated that the principal component of the grimace score was orbital tightening (squinting), at 77.1% of the total variation in the mouse grimace score. That correlates well with human studies that have demonstrated squinting behavior is one of the most consistent facial responses during painful states [2; 20; 21]. Finally, stress-induced squinting behavior cannot be ruled out. In fact, our previous publication demonstrated that restraint alone could induce a small increase in baseline squinting behavior [22]. Future studies should look to induce stress states in mice and measure facial grimace features to begin to separate stress responses from pain responses.
The development of automated pain analysis software is quickly evolving. Tuttle et al trained a convolutional neural network with mouse images previously evaluated using the Mouse Grimace Scale to distinguish the presence of pain/no-pain in a binary assessment [25]. However, small changes in pain state on a continuous scale were not possible. When using convolutional neural networks, the differential characteristics detected by the software are unbeknownst to the user. The automated system reported here utilized a definable unit (palpebral fissure area) on a continuous scale that was statistically supported as a read out for CGRP or amylin-induced pain. As more video frames of faces are landmarked and incorporated into the feature detection model, the automated system described here will be further optimized in its ability to detect pain states in mice utilizing more facial features in addition to squint. Additionally, this method and software can be applied to other species, including humans, in different environments and with different video recording systems. Detection of squint that can be objectively quantified in real time with low computing power as a read-out of pain is a key difference from neural network-based systems used to detect changes in facial features.
Other machine learning paradigms are being used to detect sub-millisecond changes in behavior, which will likely lead to a wealth of information. Indeed, automated tracking of evoked pain responses such as with foot touch stimulation will nicely pair with the automated squint assay. The tracked evoked foot response images published by Jones et al also showed varying degrees of squinting over the range of stimuli [11]. Thanks to high frame rate recording, these machine learning paradigms can detect time-resolved information undetectable by human observer-based assays. Automated methods to encompass pain detection by facial feature analysis will provide more sensitive and translatable readouts of pain analysis that could greatly increase the probability of drug discovery [19] applicable to human and veterinary medicine.
While advances have been made in this automated model, hurdles to overcome exist within the system. A major caveat is that this version of the automated squint software requires the mouse to be in a fixed position from the array of video cameras to keep the magnification constant between mice and compared to baseline. Restraint is a known inducer of stress in animals and may cover up smaller changes in pain behavior, as noted in our previous study [22]. Future efforts will focus on quantifying facial features in unrestrained mice. Direct measurement of the millimeter distance between the inner and outer canthus of the palpebral fissure of each animal will allow proper magnification scaling in free roaming mice so that restraint will not be a requirement.
Development of the automated squint analysis did allow us to make an unexpected finding. Prior studies have hinted at a greater response to CGRP in female mice, but generally the differences did not usually reach statistical significance [15; 22] With the automated assay, we were able to detect a significant sex difference at a low dose of CGRP that was not previously observed with the manual grimace assay [22], or with cutaneous sensitivity or light aversion assays [15; 23; 27]. Likewise, amylin was effective at inducing squint in only female mice, even at the relatively high single dose that was tested, and as noted above, greater responses in female mice to amylin was also seen with light aversion and cutaneous sensitivity tests [6]. While the reason for the greater sensitivity to IP CGRP and amylin in female mice is not known, it is interesting that Dussor and colleagues reported that only female mice responded to direct application of CGRP onto the dura [1]. This raises the possibility that perhaps at low CGRP doses, the site of action is preferentially at the dura, while other sites of action are recruited at higher doses. By extension, this suggests that amylin may also preferentially work in the dura. Future studies are needed to test these hypotheses.
In summary, this study is a step forward in the development of a quantifiable translational pain assay. To that end, facial analysis of pain in humans is evolving quickly and future studies should compare the use of animal grimace to human grimace. One of the key responses in human facial grimace is squinting behavior [2; 20; 21] and human squinting behavior may provide details about quantifiable pain responses through automated machine learning paradigms. The high throughput nature of the automated squint assay greatly reduces turnaround time for analysis after data collection and provides a continuous scale readout of a measurable component of the grimace response associated with pain.
Acknowledgments
Funding:
This work was supported by the National Institutes of Health (NS075599; NS113839), Department of Veterans Affairs Merit Award (1I0RX002101; I01 RX003523-0), Career Development Award (IK2 RX002010), and Center for Prevention and Treatment of Visual Loss (VA C6810-C), and Department of Defense (W81XWH-16-1-0071). The contents do not represent the views of VA or the United States Government.
Footnotes
Conflict of Interest: AFR is a consultant for Lundbeck, Amgen, Novartis, Eli Lilly, Allergan, AbbVie, and Schedule 1 Therapeutics. ASW was a consultant for Schedule 1 Therapeutics. RHK and P. Poolman are co-founders of FaceX, LLC. All other authors have no conflict of interest.
References
- [1].Avona A, Burgos-Vega C, Burton MD, Akopian AN, Price TJ, Dussor G. Dural Calcitonin Gene-Related Peptide Produces Female-Specific Responses in Rodent Migraine Models. J Neurosci 2019;39(22):4323–4331. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [2].Blais C, Fiset D, Furumoto-Deshaies H, Kunz M, Seuss D, Cormier S. Facial Features Underlying the Decoding of Pain Expressions. J Pain 2019;20(6):728–738. [DOI] [PubMed] [Google Scholar]
- [3].Bohic M, Marics I, Santos C, Malapert P, Ben-Arie N, Salio C, Reynders A, Le Feuvre Y, Saurin AJ, Moqrich A. Loss of bhlha9 Impairs Thermotaxis and Formalin-Evoked Pain in a Sexually Dimorphic Manner. Cell Rep 2020;30(3):602–610 e606. [DOI] [PubMed] [Google Scholar]
- [4].Bower RL, Hay DL. Amylin structure-function relationships and receptor pharmacology: implications for amylin mimetic drug development. Br J Pharmacol 2016;173(12):1883–1898. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [5].Crosta QR, Ward TM, Walker AJ, Peters LM. A review of pain measures for hospitalized children with cognitive impairment. J Spec Pediatr Nurs 2014;19(2):109–118. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [6].Ghanizada H, Al-Karagholi MA, Walker CS, Arngrim N, Rees T, Petersen J, Siow A, Morch-Rasmussen M, Tan S, O’Carroll SJ, Harris P, Skovgaard LT, Jorgensen NR, Brimble M, Waite JS, Rea BJ, Sowers LP, Russo AF, Hay DL, Ashina M. Amylin analog pramlintide induces migraine-like attacks in patients. Ann Neurol 2021. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [7].Glaros AG, Marszalek JM, Williams KB. Longitudinal Multilevel Modeling of Facial Pain, Muscle Tension, and Stress. J Dent Res 2016;95(4):416–422. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [8].Hay DL, Garelja ML, Poyner DR, Walker CS. Update on the pharmacology of calcitonin/CGRP family of peptides: IUPHAR Review 25. Br J Pharmacol 2018;175(1):3–17. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [9].Hohlbaum K, Bert B, Dietze S, Palme R, Fink H, Thone-Reineke C. Severity classification of repeated isoflurane anesthesia in C57BL/6JRj mice-Assessing the degree of distress. PLoS One 2017;12(6):e0179588. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [10].Hunskaar S, Hole K. The formalin test in mice: dissociation between inflammatory and non-inflammatory pain. Pain 1987;30(1):103–114. [DOI] [PubMed] [Google Scholar]
- [11].Jones JM, Foster W, Twomey CR, Burdge J, Ahmed OM, Pereira TD, Wojick JA, Corder G, Plotkin JB, Abdus-Saboor I. A machine-vision approach for automated pain measurement at millisecond timescales. Elife 2020;9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [12].Lalloo C, Stinson JN. Assessment and treatment of pain in children and adolescents. Best Pract Res Clin Rheumatol 2014;28(2):315–330. [DOI] [PubMed] [Google Scholar]
- [13].Langford DJ, Bailey AL, Chanda ML, Clarke SE, Drummond TE, Echols S, Glick S, Ingrao J, Klassen-Ross T, Lacroix-Fralish ML, Matsumiya L, Sorge RE, Sotocinal SG, Tabaka JM, Wong D, van den Maagdenberg AM, Ferrari MD, Craig KD, Mogil JS. Coding of facial expressions of pain in the laboratory mouse. Nat Methods 2010;7(6):447–449. [DOI] [PubMed] [Google Scholar]
- [14].Marquez de Prado B, Hammond DL, Russo AF. Genetic enhancement of calcitonin gene-related Peptide-induced central sensitization to mechanical stimuli in mice. J Pain 2009;10(9):992–1000. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [15].Mason BN, Kaiser EA, Kuburas A, Loomis MM, Latham JA, Garcia-Martinez LF, Russo AF. Induction of Migraine-Like Photophobic Behavior in Mice by Both Peripheral and Central CGRP Mechanisms. J Neurosci 2017;37(1):204–216. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [16].Miller A, Kitson G, Skalkoyannis B, Leach M. The effect of isoflurane anaesthesia and buprenorphine on the mouse grimace scale and behaviour in CBA and DBA/2 mice. Appl Anim Behav Sci 2015;172:58–62. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [17].Mogil JS, Pang DSJ, Silva Dutra GG, Chambers CT. The development and use of facial grimace scales for pain measurement in animals. Neurosci Biobehav Rev 2020;116:480–493. [DOI] [PubMed] [Google Scholar]
- [18].Neumann E, Kupfer L, Zeilhofer HU. The alpha2/alpha3GABAA receptor modulator TPA023B alleviates not only the sensory but also the tonic affective component of chronic pain in mice. Pain 2021;162(2):421–431. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [19].Pereira TD, Aldarondo DE, Willmore L, Kislin M, Wang SS, Murthy M, Shaevitz JW. Fast animal pose estimation using deep neural networks. Nat Methods 2019;16(1):117–125. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [20].Prkachin KM. The consistency of facial expressions of pain: a comparison across modalities. Pain 1992;51(3):297–306. [DOI] [PubMed] [Google Scholar]
- [21].Prkachin KM, Solomon PE. The structure, reliability and validity of pain expression: evidence from patients with shoulder pain. Pain 2008;139(2):267–274. [DOI] [PubMed] [Google Scholar]
- [22].Rea BJ, Wattiez AS, Waite JS, Castonguay WC, Schmidt CM, Fairbanks AM, Robertson BR, Brown CJ, Mason BN, Moldovan-Loomis MC, Garcia-Martinez LF, Poolman P, Ledolter J, Kardon RH, Sowers LP, Russo AF. Peripherally administered calcitonin gene-related peptide induces spontaneous pain in mice: implications for migraine. Pain 2018;159(11):2306–2317. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [23].Recober A, Kaiser EA, Kuburas A, Russo AF. Induction of multiple photophobic behaviors in a transgenic mouse sensitized to CGRP. Neuropharmacology 2010;58(1):156–165. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [24].Tracey I, Woolf CJ, Andrews NA. Composite Pain Biomarker Signatures for Objective Assessment and Effective Treatment. Neuron 2019;101(5):783–800. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [25].Tuttle AH, Molinaro MJ, Jethwa JF, Sotocinal SG, Prieto JC, Styner MA, Mogil JS, Zylka MJ. A deep neural network to assess spontaneous pain from mouse facial expressions. Mol Pain 2018;14:1744806918763658. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [26].Wattiez AS, Castonguay WC, Gaul OJ, Waite JS, Schmidt CM, Reis AS, Rea BJ, Sowers LP, Cintron-Perez CJ, Vazquez-Rosa E, Pieper AA, Russo AF. Different forms of traumatic brain injuries cause different tactile hypersensitivity profiles. Pain 2021;162(4):1163–1175. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [27].Wattiez AS, Sowers LP, Russo AF. Calcitonin gene-related peptide (CGRP): role in migraine pathophysiology and therapeutic targeting. Expert Opin Ther Targets 2020;24(2):91–100. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [28].Williams AC. Facial expression of pain: an evolutionary account. Behav Brain Sci 2002;25(4):439–455; discussion 455–488. [DOI] [PubMed] [Google Scholar]