Measuring Mouse Somatosensory Reflexive Behaviors with High-speed Videography, Statistical Modeling, and Machine Learning

Abstract

Objectively measuring and interpreting an animal’s sensory experience remains a challenging task. This is particularly true when using preclinical rodent models to study pain mechanisms and screen for potential new pain treatment reagents. How to determine their pain states in a precise and unbiased manner is a hurdle that the field will need to overcome. Here, we describe our efforts to measure mouse somatosensory reflexive behaviors with greatly improved precision by high-speed video imaging. We describe how coupling sub-second ethograms of reflexive behaviors with a statistical reduction method and supervised machine learning can be used to create a more objective quantitative mouse “pain scale.” Our goal is to provide the readers with a protocol of how to integrate some of the new tools described here with currently used mechanical somatosensory assays, while discussing the advantages and limitations of this new approach.

Introduction

Pain is a complicated and subjective experience. In humans, pain is usually assessed from the patient’s own description of experience, which is roughly quantified by self-assignment along a single-score pain scale [1–5]. To investigate pain mechanisms and screen for new pain treatment reagents, animal models, such as rodents, are commonly used [6–9]. However, more and more researchers are realizing that objectively measuring and interpreting pain sensation in rodents is a big challenge [10–14].

Since rodents are non-verbal, researchers have relied on their behaviors to infer their pain state. The reflexive withdrawal assays are the most widely used measurements of pain sensation in rodents, in which a stimulus is applied to a region of the rodent, such as the paw or tail, and the withdrawal frequency is quantified as a readout for the animal’s pain state [15–19]. In this chapter, we will focus exclusively on the paw withdrawal assay and our recent efforts to improve it.

The reason why pain researchers like to use the paw withdrawal test to measure pain sensation in rodents is because it’s easy to perform this test as well as train novice experimenters to collect data [20–23]. Additionally, the results of the paw withdrawal test are instantaneous (i.e. presence or absence of paw lift). Moreover, the behavioral setup to conduct this test is inexpensive and relatively easy to setup, as opposed to more expensive platforms [24, 25]. Other pain behavioral tests, such as spontaneous pain behaviors or the operant assays, may require hours or even days for animals to perform or learn to perform tasks [26–33]. Lastly, the striking similarity between the paw withdrawal reflexes of rodents and human withdrawal responses evoked by noxious stimuli (i.e. humans also display withdrawal reflexes to calibrated von Frey hair filaments or noxious thermal stimuli) adds face validity for using this test as a readout of pain [34–36]. As a field, we have learned much mechanisms and physiology underlying pain processing by using this simple behavioral assay [37–39]. Notably, however, for bench-to-bedside translation of new pain treatment reagents, the successful rate from discoveries made in rodents has not been impressive [40–43]. Though many reasons could contribute to this low success rate, this has caused many pain researchers, clinicians, and pharmaceutical companies to question the specificity, robustness, and validity of the behavior assays used to assess pain sensation in rodents [44–46]. In this chapter, we describe innovation pieces added to the commonly used paw withdrawal assay that might help to mitigate some of the traditional limitations.

A major limitation from focusing solely on the incidence rate of paw withdrawal in response to application of sensory stimuli, as the field traditionally does, is that behaving rodents often display a similar responsive rate to both noxious and innocuous mechanical stimuli. For example, the dynamic brush test uses a soft innocuous brush applied to the mouse paw, and many researchers report a paw withdrawal frequency in approximately 80% of trials [47–50]. This incidence rate is similar to what is commonly reported for potentially noxious stimuli like pinpricking needles or high force von Frey hair filaments [47, 48, 51]. Since the withdrawal frequency itself is not sufficient to distinguish a touch or mechanical pain response, the experimenters have to assign what a given stimulus might mean to the animal, i.e. “noxious” or “innocuous”. The problem with this “assignment” process is that rodents and humans, or even different human individuals, may perceive some stimuli quality differently. There are massive debates in the field on what some of the most commonly used sensory stimuli are actually measuring in rodents. For example, some reports suggest that 1.4grams of force with von Frey hair filaments is a noxious stimulus to mice, while other reports suggest that this stimulus is perceived as gentle touch [52, 53]. This kind of discrepancy makes it difficult to translate findings across labs. How can we avoid this subjective “stimulus quality assignment” process?

Mechanical somatosensory stimuli turn on sensorimotor neuronal circuits of an animal, which oftentimes trigger very rapid movements, within 100 milliseconds from the stimulus onset [54–57]. Therefore, when researchers tried to measure rodent reflexes to mechanical sensory stimuli with the unaided eyes, extracting movement features besides incidence of paw lifting has been challenging if not impossible. To gather more detailed information related to the animal’s behavioral response, others and we have turned towards high-speed videography to capture sub-second paw withdrawal features. For example, the Woolf and Ginty labs recorded state dependent sub-second pain behaviors at 1,000 frames per second (fps) with an optogenetic approach to activate nociceptive neurons [58]. The Lechner group also recorded sub-second pain behaviors with optogenetic activate of nociceptors at 240 fps using an iphone6 smartphone [59]. The Iadarola group likewise recorded unrestrained rats with a high-speed camera at 500 fps, using noxious mechanical and thermal stimuli and observed multi-segmental body movements in response to the stimuli [60]. In our methodology, as detailed below in the Methods section, we also recorded sub-second behaviors at 500 fps or 1,000 fps in freely behaving mice [61]. A major distinction in our work compared to some of the other studies is the application of both innocuous and noxious stimuli and the quantification and a composite statistical analysis of behaviors, transforming the multifactorial datasets into a single dimension with principal component analyses. This analysis of behaviors to diverse stimuli allowed us to generate a threshold that distinguishes mouse paw withdrawal reflexes indicating touch from those indicating pain. Furthermore, with a supervised machine learning method, we could determine the “pain-like” probability from the paw withdrawal reflex on a trial-by-trial basis. Taken together, our results reveal that capturing rapid behaviors using high speed imaging and merging multiple behavior features into one numeric parameter can greatly improve precision and rigor of rodent somatosensory behavioral assays.

Materials

Animals

We performed our high-speed assessment analysis with wild type mice purchased from the Jackson Laboratories and Charles River first. Specifically we used the C57BL/6J inbred strain (stock no. 000664) and the CD-1 (stock no. 022) outbred strain. For optogenetic activation of nociceptors we used the following three mouse lines, which are available from the Jackson Laboratories: ChR2 f/f Ai32 (stock no. 012569) [62], Mrgprd^CreERT2 (stock no. 031286) [63], TrpV1^Cre (stock no. 017769) [64].We conducted behavior assays with both male and female mice to control for the potential sex differences.

Somatosensory stimuli

1. Natural mechanical stimuli with pre-defined quality: We first applied four mechanical stimuli with pre-defined quality for our studies: cotton swab, dynamic brush (make-up brush from CVS, e.l.f.), and light or heavy application of a pinprick needle (Austerlitz). Details on how these stimuli are applied to the mouse hind paw are described in the Methods. Cotton swab and dynamic brush are commonly used by the field as “innocuous touch” stimuli whereas pinprick is commonly used as a “noxious” stimulus. In collaboration with Dr. Xinzhong Dong’s group at the Johns Hopkins University using whole animal dorsal root ganglion calcium imaging [65–67], we demonstrated that the cotton swab and dynamic brush mainly activate large diameter touch neurons, while the pinprick stimuli mainly activate small diameter nociceptors [61], further validating the stimulus quality. These stimuli were applied on different days in limited number of trials (usually no more than 3 times/day) to avoid any sensitization to stimuli. Further testing should be conducted if labs want to give the animals multiple stimuli in a single session or day.

2. Von Frey hair filaments (VFHs): These are the most commonly used mechanical stimuli [20, 68–70], which deliver a known quantity of mechanical force to a surface. We used three different VFHs (Stoelting Compnay, 58011) that are frequently used in the field: 0.6 grams, 1.4 grams, and 4 grams.

3. Peripheral optical stimuli: To activate the blue light sensitive ion channel channelrhodopsin (ChR2) transdermally through the paw skin, we delivered 473nm blue laser light (Shanghai Laser and Optics Century, BL473T8–150FC/ADR-800A) pulsed at 10hz sine frequency controlled by a pattern generator (Agilent 10MHZ Function Waveform Generator (33210A). The laser light source was connected to an FC/PC optogenetic patch cable with a 200 mm core opening (ThorLabs, M72L01), in order to get the light from the source to the animal’s hind paw. Light power intensity was held between 5–20mW for each experiment and was measured with a digital power meter with a 9.5mmaperture (ThorLabs, PM100A). The technique of using optogenetics to activate somatosensory afferents through the skin of mice has also been performed by many other research groups [71–73].

Testing platform and holding chambers

For a platform to place mice on top of during testing (part A in Fig. 1), we used a custom designed mesh floor with perforated openings from below that allows experimenters to reach the mouse hind paw to directly deliver a mechanical stimulus. For holding the mice in chambers that allow free yet restricted movement, small plexiglass rectangular holding chambers with dimensions (11.5 × 11.5 × 16 cm) were used that did not allow mouse standing up or raring (Part B in Fig. 1). We have found that this design helps with behavioral characterizations, as mice do not stand up with front paws in the air. Although we built these items in house, the similar testing platform and mouse holding chambers can be purchased commercially from many vendors.

Figure 1: Set up for the High-speed imaging method.

This picture shows all the equipment and parts necessary for our high-speed image recordings. (A) Raised mesh platform where mice stand while behavioral studies are performed, (B) A mouse holding chamber that allows free but restricted movement, which also facilitates side imaging of behavior. (C) Computer attached to the camera for real-time viewing of the behavioral data that is being captured and stored. (D) Lens attached to the high-speed camera for focusing on the mouse. (E) Infrared lighting attached to the camera, which illuminates the mouse to the camera. (F) High-speed camera. (G) Tripod outfitted to hold to the high-speed camera.

High-speed camera

To capture quick paw withdrawal and nocidefensive behaviors at sub-second speeds we used a high-speed camera, Photron FastCAM UX100 (800K-M-4GB - Monochrome 800K with 4GB memory) (Part F in Fig. 1). This particular camera allows capturing of behaviors at much fast speeds around 1,000 fps or higher. Therefore, it is possible that some less expensive cameras, whose maximum frame rate is between 500 – 1,000 fps, may also be sufficient to use our method. In addition, a Zeiss lens (Milvus 2/100M ZF.2) was attached to the camera for zoom and focus capabilities (Part D in Fig. 1). The camera and lens were held with a fitted tripod (Part G in Fig. 1) approximately 1–2 feet away from the mice and directed at a 45° angle from the testing setup. To maximally activate the camera, we included a far-red shifted LED light (Tech Imaging) (Part E in Fig. 1), which mice can’t see. This is an important consideration for having better imaging effects without affecting the mouse behaviors.

Methods

Performing somatosensory behaviors

We ordered testing mice directly from commercial vendors and allowed the animals to habituate to our animal facility for at least two weeks prior to testing. Next, we habituated animals for one hour per day for approximately one week in our behavior room with animals in the holding chambers on top of the testing platform. Additionally, the experimenter performing the behavior remained in the room with the animals during the entire habituation period and the experimenter waved a gloved hand under the animal platform to mimic stimulus delivery. These habituating measures were taken to avoid startle to the animals when tests were commenced.

On the actual testing day, mice were allowed to habituate for 15–30 minutes before we applied somatosensory stimuli one hind paw on either side of the animal, depending upon which paw was in view of the imaging setup. To obtain high quality data that is easy to interpret, it is imperative that the animals are calm and still when a sensory stimulus is delivered. Our testing platform allowed 5 animals to be mounted on the platform at once, and tested in an assembly line fashion one-by-one. We put dividers between the animals so that they could not see one another and thus affect each other’s behavioral responses. We only delivered one kind of stimuli per day to each mouse tested.

For application of the cotton swab stimulus, we used a q-tip for gentle application to the mouse hind paw. For dynamic brush, we used a soft make-up brush to stroke the hind paw gently from back to front. For light pinprick, we used an insect-pinning needle to apply a weak application of the stimulus to the hind paw and the stimulus was withdrawn when the mouse began paw lift. With heavy pinprick, we used the same needle as above, but applied more intense force upwards and withdrawing the needle when approximately 1/3 of the pin’s length had passed through the mesh platform. Lastly, for application of von Frey hair filaments, the stimulus was delivered upward to the mouse hind paw until the filament bent. As the majority of positive responses occurred within 100 milliseconds of stimulus delivery, if an animal failed to respond within 2 seconds, we considered it as a negative response. We chose this criterion to limit the false positive rate of scoring a response to normal animal movement that was unrelated to the actual stimulus.

For all of these stimuli, if mice did not respond in the first trial, we would repeat the stimulus until we could record a response, with trials separated by at least 5 minutes. We curated and saved each video prior to moving to the next animal to ensure a high quality video and correct delivery of stimuli. On average, it took 2–3 minutes for one video to save to the internal memory within the high-speed camera.

Scoring sub-second withdrawal and escape behaviors

After completing the behavioral testing, the next step is to work at a computer to quantify the behaviors at sub-second resolution. Since we used the Photron high-speed camera, we did behavioral analysis in the companion Photron Fastcam Viewer software. This software is freely downloadable and should be compatible with videos taken from other cameras. Although we initially scored many movement features, we settled upon scoring six that contributed to the majority of variance in the data (Fig. 2). We completed this task using an exploratory factor analysis with orthogonal Varimax rotation conducted with SPSS software. Basically, we chose features that explained the majority of the observed results and excluded other features that were likely redundant or not significant for distinguishing touch from pain. The six features can be divided into sensory-reflexive behaviors (maximum paw height and velocity) and defensive/coping/affective behaviors (paw guarding, paw shaking, jumping, orbital tightening) (Fig. 2 A-D). For measuring paw height, we simply drew a vertical line in the software to measure the distance between the paw on the floor to its maximal height (Fig. 2 E). For measuring paw velocity we tracked the paw during its ascent away from the stimulus and captured a snapshot of velocity during its maximal peak, recording the distance between two points and the time it took the paw to get from one point to the next (Fig. 2 E). For the “affective” features [74, 75], we used a yes/no (ie. 0 or 1) binary system for presence or absence of the corresponding behaviors (Fig. 2A-D). We combined these 4 measures into one number that we called a “pain score.” For example, if an animal displayed orbital tightening and paw shaking it received a score of 2. With these measures, we observed clear statistical separation between the touch stimuli (cotton swab and dynamic brush) and the pain stimuli (light and heavy pinprick) [61]. Lastly, we tested both male and female mice, and did not observe statistically significant sex differences in the behaviors [61].

Figure 2: Sub-second behavioral movements that distinguish touch from pain.

Pain sensation associated defensive and coping behaviors: paw shaking (A), jumping (B), paw guarding (C), and orbital tightening (D). (E, E’) Reflexive behavioral measurements were made by determining the maximal height of the stimulated paw and the speed taken to reach that height. Panel E shows the animal close to the time of stimulation just before the paw ascends in the air. The red asterisk labels the “starting rest position” of the stimulated paw. Panel E’ shows where the paw at rest (red asterisk) to its maximum withdrawal height (blue asterisk). Panels E and E’ are offset by approximately 15 milliseconds for this example. (F) Graphical abstract summarizing how we move from behavioral assessment to creation of a “pain scale” by collapsing the data into a single dimension via PCA analysis or predicting pain-like probabilities via machine learning on a trial-by-trial basis. Images in panels A-D, F were adapted from [61], with permission from the publisher Elseiver.

Transforming data into a single dimension with principal component analysis

To produce a simple one-dimensional score that encompassed the six sub-second movement features we scored, we used principal component analysis. We first converted the raw data measured in different units (ie. mm, mm/sec, 0,1) into z-scores to normalize numbers and units and to visualize how far individual datapoints diverged from the mean of the entire group. Next we used principal component analysis [76] for dimension reduction using SAS software, which generated contributing weights, or eigenvalues, for each measurement. After plotting the principal component analysis scores (or PC scores) the data aligned along a continuous scale from −3 to 3. We noticed that the PC scores to innocuous stimuli were typically negative, while the scores to painful stimuli were typically positive. Therefore, ‘0’ appeared to serve as a nice cutoff with this tool to statistically separate touch from pain (Fig. 2F). Therefore, this linear distribution of PC scores seems to make up a mouse “acute pain scale”.

Supervised machine learning to make predictions about pain-like probabilities

To provide a more “user friendly” output for interpreting mouse paw withdrawal reflex, we could use part of the collected data (cotton swab and heavy pinprick) to train a SVM algorithm [77], which was then able to predict “pain-like” probability of any given mouse paw withdrawal trial. For details of our SVM machine-learning pipeline, we direct you to our prior published study [61]. This “pain-like” probability basically mirrored our findings using the PC scores (Fig. 2F).

Application of our method to study VFH and peripheral optical stimuli triggered paw withdrawal reflexes

After creating this new mouse “pain scale” method to measure and interpret mouse paw withdrawal reflexes, we asked whether we could use this new method and collected database to map the mouse sensory experiences in response to VFHs and optogenetic activation of different types of nociceptors.

For VFHs, we used three stimuli that are frequently used to measure pain in mice (0.6 grams, 1.4 grams, and 4 grams, Fig. 3A-C) and tested them on CD1 male mice. Consistent with other publications, we found ~50% withdrawal rate with 0.6 gram VFH and close to 100% withdrawal rate for both 1.4 gram and 4 gram VFHs. [61, 78]. As described above, we quantified the paw height and velocity, orbital tightening, jumping, and paw shaking and guarding. From the raw data, we calculated Z score and principal component (PC) scores using the CD-1 baseline data with cotton swab, dynamic brush, and pinpricks as the reference. With this approach, we observed that responses to 0.6grams and 1.4grams had on average negative PC scores, while the responses to 4 grams had positive PC scores (Figure 3A). In other words, 0.6grams and 1.4grams map in the non-pain domain with behavioral responses reminiscent of cotton swab and dynamic brush – despite 1.4 grams having close to 100% paw withdrawal rates. With a SWM trained by CD1 male cotton swab and heavy pinprick data, it predicted that 0.6grams had a low probability of being pain-like (~30%), 1.4grams was near the threshold that separates touch from pain (~50%), and 4grams had a high probability of being pain-like (~80%, Figure 3B). It is worth noting that the field usually used 50% paw withdrawal, 0.6 gram mechanical force here, as the mechanical pain threshold [49, 78]. However, our analysis argued that this 50% withdrawal threshold indicates mechanical touch rather than mechanical pain.

Figure 3: Measuring mouse sensation in response to VFHs and optical stimuli.

(A-C) PC scores and SVM machine learning reveal different sensations and the pain-like probabilities triggered by three different forces of von Frey hair filaments. (D-F) PC scores and SVM machine learning reveal different sensations and the pain-like probabilities triggered by optogenetic activation of two distinct nociceptor populations. These images were adapted from [61], with permission from the publisher Elseiver.

To measure the sensory experience evoked by optogenetic activation, we used two transgenic mouse lines to optically activate two different populations of nociceptors. We used Mrgprd-ChR2 to activate C-fiber non-peptidergic nociceptors [63] and Trpv1-ChR2 which targets Trpv1-lineage neurons including the majority of nociceptors [58, 74, 79] (Fig. 3D-F). We activated the sensory afferent terminals in the hind paw skin using 473nm blue laser light. For negative control experiments, we applied the same blue light to littermate animals that did not express the light-sensitive ChR2 and we did not observe non-specific responses. In the experimental animals, we observed high paw withdrawal rates in the entire group of mice. We next measured the behavioral parameters described in Figure 2, followed by calculating PC scores and pain-like probabilities with a SVM trained by all C57 and CD1 cotton swab and heavy pinprick data. Optically activating Trpv1-ChR2 line triggered pain sensation, indicated by both the PC scores and pain-like probability (Figure 3E-F), which is further validated by a reversion of responses after painkiller administration (Figure 3E-F). Despite high paw withdrawal rates (Figure 3E-F), optical stimulation of MrgD-ChR2 mice triggered non-painful paw withdrawals. Only after local inflammation by paw injection of the heat-killed bacteria complete Freund’s adjuvant (CFA), optical stimulation triggered pain-like paw withdrawal reflexes (Figure 3E-F), which was also validated by painkillers (Figure 3E-F). Our data strongly suggests that activation of Mrgprd+ nociceptors at baseline states doesn’t cause a pain response. Other research groups have arrived at similar conclusions about these neurons using chemogenetic approaches and/or operant assays [80–82].

Together, these data revealed the power of our new “pain scale” method to measure mouse pain sensation from paw withdrawal reflexes.

Notes

In summary, the relatively simple improvements we described here seem to facilitate greater accuracy in assessing the mouse somatosensory experience from their paw withdrawal reflexes. Since this method is congruent with the behavioral setups that most labs currently use, we do not anticipate major barriers for adopting this approach as long as a high-speed camera is available. Although we have delineated some of the advantages of using this technology, there are still some limitations to be considered. What we have presented is a platform to measure responses to mechanical stimuli under baseline and inflammatory states, but not during chronic neuropathic pain conditions. Neuropathic pain is of the most common, yet hard to treat, chronic pain in the clinic, and we need more robust ways to measure it in preclinical animals. Since many neuropathic pain models involve injury of a nerve, usually the sciatic nerve [6, 83–91], two movement features we emphasized here, paw withdrawal height and velocity, will be affected. Whether our approach could be used for studying neuropathic pain or not will need to be tested in different neuropathic pain models. In addition, this method measured evoked but not spontaneous responses. Spontaneous pain is an important dimension of the pain experience in both the clinic and in animal models [92–101]. Therefore, it would be the best to combine this platform with technologies measuring mouse spontaneous pain behaviors, such as the grimace scale for example or other spontaneous behavior detection platforms [27, 102–111], for achieving a more complete picture of the rodent pain state. Moreover, we manually delivered the stimuli and quantified the behaviors for the previously published study [61]. Automation for some or all of these steps would help to reduce human errors as well as subjectivity. Another issue that will be mitigated with time and when more researchers adopt this technology is that it is currently unclear whether the sub-second behavioral features we defined are similar for different mouse strains or other rodent species such as rats. Nonetheless, these limitations do not preclude adopting this approach now to move beyond the simple yes-no incidence rate quantification of paw withdrawal reflexes. The methodology described here, alongside others with the capacity to detect spontaneous bouts of pain, should greatly increase our confidence in measuring the mouse sensory experience and potentially speed up the rate of translating basic science findings into the clinic.