A Fine Motor Task to Study Joint Kinematics in a Preclinical Model of Neurodegenerative Disease

Summary

The current study describes a fine motor behavior test for examining motor deficits in rodent models, including the TgF344-AD rat, using machine learning.

Abstract

Motor dysfunction is a critical, yet often underappreciated, component of neurodegenerative diseases such as Alzheimer's disease (AD) and related dementias. New research has shown that tasks involving fine motor behavior, visuospatial ability, and executive function in humans may have potential in early disease detection. However, little is known about the exact mechanisms by which these tasks predict cognitive and functional decline associated with AD. In the current method, a fine motor task, Kinematic Motor Ability Task (KINEMAT), was created to examine changes in motor ability across disease progression in preclinical models of AD relative to other neurodegenerative diseases and normal aging using free, open-source machine learning software. Here, proof-of-concept was established with the design, testing, and implementation of the fine motor task in a pilot study of Fisher 344 CDF rats with and without Harmaline-induced motor impairments. Additional validation, with an Alzheimer's disease rat model (TgF344-AD), further demonstrated model feasibility and task sensitivity. Rats were trained to reach through a small slot at the front of the chamber to retrieve a sugar pellet from one of three bowls with varying difficulty. Pellet retrieval was scored using one of three classifications: success (successfully grabbing and eating the sugar pellet), drop (grabs but drops the treat upon grasp), or failure (reach in which no pellet is procured, or pellets are knocked out of the bowl when attempting to grasp). Significant differences in motor ability were observed between control and Harmaline-treated rats, demonstrated by reduced total reaches and increased paw pose variability. TgF344-AD rats also showed reduced total reaches, as well as reduced successes across all bowls. In validating this as a fine motor task in rodents, it has future potential to study normal aging and preclinical models of neurodegenerative diseases, including AD, as pathology and symptoms emerge over time.

Introduction

Motor function has been shown to be impaired in people with dementia due to Alzheimer's disease (AD) or other subtypes (e.g., dementia with Lewy bodies). Grip strength and gait speed can distinguish between cognitively impaired and unimpaired older adults^1,2, but cannot necessarily differentiate between different subtypes of dementias (i.e., low specificity)^1,2. Thus, there continues to be a need for new diagnostic tools for early dementia, possibly paired with machine learning, for evaluating complex patterns and relationships between behavior and disease state^3,4. Animal models are valuable for investigating the progression of neurodegenerative diseases to offer insights into early disease onset, behavioral correlates, and integrating machine learning into analysis pipelines to distinguish between severity and presentation^5,6,7.

New behavioral methods in humans have identified tasks that involve fine motor skill along with cognitive processing, like visuospatial ability and executive function^8,9,10. For example, a functional upper-extremity task can track patterns of neurodegeneration and can distinguish mild cognitive impairment (MCI) from AD. In this task, humans move raw kidney beans (two at a time) from a central home cup to a target cup using a spoon. The phase of the task in which beans are acquired with the spoon shows stronger associations with cognition (specifically, visuospatial memory) than the phase in which the beans are transported to the target cup^11,12. As the spoon enters the cup, it may perturb the beans in the cup, meaning the participant's movements may make the task more difficult. Further, as the trial progresses and the number of beans in the home cup decreases, the task becomes more difficult (i.e., time to acquire the beans with the spoon increases)¹³. This task has been shown to correlate with AD-specific pathology (amyloid)^9,14 and neurodegeneration (hippocampal atrophy)^10,15, suggesting its performance may depend on neural circuitry and behavioral processes that are affected early in the disease process. It is known that movement initiation, planning, execution, and correction of movement are controlled by a diverse network traversing the brain, including the motor cortex, with cell bodies of upper motor neurons located in the cerebral cortex synapsing on lower motor neurons in the spinal cord¹⁶. These lower motor neurons directly communicate with muscles to control movement at the neuromuscular junction. Ascending tracts carry sensory information from the periphery to the spinal cord, while descending tracts within the brain communicate directional information regarding how and where to move back to muscles^17,18. Understanding how these pathways are selectively affected by AD pathology is critical to understanding disease progression, particularly since genetic animal models, including the 3xTg-AD mouse and TgF344-AD rat, have been previously used to study how AD genes impact locomotion, coordination, and other motor impairments^19,20,21. Typically, these tasks focus on gross motor impairments relating to whole-body movement, but motor deficits in humans are much more variable, which can be difficult to identify in rodents²². New tasks are therefore needed to bridge the gap between human and animal models to capture this variability.

One brain region that is commonly implicated in motor impairments and tremors is the cerebellum²³. The cerebellum is a significant regulator of precise and coordinated movements and acts as a processor of sensory information to provide feedback through descending pathways to correct and adapt this voluntary movement^24,25. It is important to note that cerebellar atrophy has been recognized in individuals with AD, but, contrary to other regions like the hippocampus, increased amyloid deposition in the cerebellum does not correlate with increased atrophy²⁶. While the progression of AD pathology is different in the cerebellum, there are multi-synaptic bidirectional connections between the cerebellum and hippocampus that are likely disrupted in AD^27,28,29. Notably, there are functional connections between the cerebellum and hippocampus, as memory deficits have been documented to occur after cerebellar tumor resection^{30,31,32,33,34,35,36,37}. In rodents, hippocampal place cells are disrupted with a cerebellar perturbation³⁸. These overlapping deficits in memory, sociability, and motor control highlight the importance of uncovering potential early changes in motor pathways that occur before cognitive decline^30,31,38,39. With the implementation of fine motor tasks in early AD research, we may begin to reshape the traditional understanding of AD as simply a cognitive disorder and redirect diagnostic efforts to early motor changes that may predict disease progression and may be more greatly distinguishable between cognitively impaired and unimpaired aging individuals.

To understand how fine motor behavior can be indicative of changes in brain structure, a rodent version of the human task described above was created and tested. This rodent task has been named KINEMAT (KINEmatic Motor Ability Task), and it deviates from previous reaching tasks that require the retrieval of a single pellet from a pedestal or staircase and the more simplistic scoring as a result of consistent pellet placement at all stages of the task^{40,41,42,43,44,45,46,47,48}. This fine motor reaching task includes new scoring methods, bowls with various difficulties, and self-perturbing features to mimic the human counterpart, with pellet movement and subsequent participant strategy adjustments central to the design. This task employs the machine learning tools, Social LEAP Estimates Animal Poses (SLEAP) and keypoint-MoSeq^49,50, allowing an unbiased and quantitative measurement of fine motor ability and reaching behaviors. While other notable computational models have been developed for skilled reaching tasks in rodents^44,45,51, keypoint-MoSeq eliminates noise to allow for precise quantification of movement and grasping poses⁵⁰. Task performance and motor learning strategy were reconstructed using a high-performance machine vision camera in typical and motor-perturbed rats using Harmaline⁵². Harmaline hyperactivates the inferior olivary nucleus and disrupts cerebellar function through glutamatergic climbing fibers^53,54. In addition, fine motor ability was examined in a subset of TgF344-AD rats. Thus, it was expected that the fine motor task would find measurable impairments in Harmaline-treated and TgF344-AD rats relative to their typical counterparts as a proof-of-concept. This protocol describes constructing the fine motor task, training and testing animals, and quantifying reaching behavior using SLEAP and keypoint-MoSeq.

Protocol

All methods described here have been approved by the Institutional Animal Care and Use Committee (IACUC) of Arizona State University.

1. Design and construction of the chamber

1. Prepare the clear, acrylic panels. The schematics of the box are located at https://github.com/verpeutlab/ADFineMotorTask. Cut clear acrylic panels to the following dimensions (all panels are 1 cm thick):
   1. Cut out two side panels: 26 cm × 22 cm.
   2. Cut out a top panel: 27 cm × 12 cm.
   3. Cut out the first bottom panel: 26 cm x 10 cm
   4. Cut out the second bottom panel: 26 cm x 11 cm
   5. Cut of a front panel: 20 cm × 11 cm.
   6. Cut out a back panel: 20 cm x 12 cm
   7. Cut out a front panel slot (140 mm x 12 mm). Position the slot centrally, with its lower edge aligned with the bottom of the panel.
   8. Cut out an inner groove: 1 cm. On the inner edges of the side panels, create an indent or groove to allow the front panel to slide in and out securely.
      NOTE: If the sliding feature is not required, the front panel may be permanently glued to simplify construction.
   9. Drill evenly spaced holes (5 mm in diameter) into the second bottom panel to allow for pellets to fall through if dropped. Space the holes 1 cm apart in a grid pattern.
2. Assemble the chamber with adhesive (Figure 1A).
   NOTE: Use a strong, transparent adhesive designed explicitly for acrylic materials.
   1. Side panels: Apply a continuous line of adhesive along the edges of the bottom panel. Position the side panels vertically and press firmly. Allow the adhesive to cure per the manufacturing instructions.
   2. Back panel: Apply adhesive to the rear edges of the side and bottom panels. Position the back panel and press it into place.
   3. Top panel: Apply adhesive to the top edges of the side and back panels. Position the top panel and press firmly.
   4. Front panel: Carefully slide the front panel into the grooves on the side panels, ensuring a snug but smooth fit, or apply adhesive along the edges of the front panel and secure it to the side, top, and bottom panels.
      1. Test the panel's movement and confirm that it aligns precisely with the rest of the chamber.
3. Set-up lighting
   1. Position the infrared (IR) light on the opposite end of the testing chamber, pointing towards the chamber. Place a sheet of white construction paper on top of the chamber to reflect the light down.
      NOTE: Lighting is critical for proper tracking, and all recordings should be made under IR light in a dark room.
4. Prepare the training bowl.
   1. Fabricate the training bowl with a 31 mm radius and a 32 mm extension using a 3D printer. STL file available on https://github.com/verpeutlab/ADFineMotorTask (Figure 1B: top)
   2. Attach the training bowl to the bottom of the chamber directly in front of the slot. Position the bowl so that its top edge is roughly above the eye level of the rat within the chamber, using a supportive stand, such as a block of wood.
5. Prepare the testing bowls (Figure 1B: middle and bottom).
   1. Print the bowls using the provided STL files. Two distinct bowl designs are provided: "Plain-Less" and "Plinko." (https://github.com/verpeutlab/ADFineMotorTask) 
      1. Inspect for dimensional accuracy and durability (Dimensions: 50 mm deep x 40 mm wide).

2. Computer and camera hardware

1. Use a tripod to mount the camera securely.
2. Position the camera approximately 15-20 cm from the front panel of the chamber.
   1. Adjust the tripod height and angle so the camera captures an isometric, 45° view of both the slot and the bowl.
      NOTE: Ensure the camera's field of view encompasses the rat paw and complete reach, the slot, and the training or testing bowl. Test different angles and distances if necessary to achieve optimal visibility.
3. Download software to pair with the camera for behavioral recordings (https://www.teledynevisionsolutions.com/products/spinnaker-sdk) and open the software to use the camera to record behavior.
   1. Use the following camera settings: acquisition mode: continuous, frame rate: 216 Hz, exposure mode: timed, auto exposure: off, exposure time: 4430 µs, gain auto: off, gain: 16.85 dB, gamma: 0.8.
   2. Recording settings are as follows: H264 file type, 5 million (5,000,000) bitrate.
   3. To record for 10 min, enter 130,000 frames and use streaming mode to save.
      ​NOTE: The use of a dedicated machine with a minimum of 16 GB RAM and a 1 TB internal solid-state hard drive for storing data is recommended.

Figure 1: Schematic of the fine motor task. (A) Cartoon representation of the fine motor task whereby a rat reaches through a slot to obtain a sugar pellet. Reaches are recorded for machine learning analysis. There are three separate bowls used: a (bottom left) training bowl, (bottom middle) Plain/Less bowl, and the (bottom right) Plinko bowl with obstructions circled in red. (B) Schematics of each bowl design, including (top) training bowl, (middle) Plain/Less bowl, and (bottom) Plinko bowl. Please click here to view a larger version of this figure.

3. Animal preparation

1. Pair house (at a minimum) Fisher 344 CDF rats (male and female, >3 months of age) on a reverse light/dark cycle to allow for testing when animals are most active.
2. Weigh all rodents 1 week prior to the beginning of habituation and provide ~30 g of food for a rodent weighing over 300 g and ~15 g for a rodent 100-300 g to maintain animals at 90% body weight.
   1. Daily weigh the rats and adjust the food accordingly, reducing the amount of food provided by 1-2 g per day until the rodent reaches 90% body weight. Daily handling will also help to habituate the animals to the experimenter.
   2. Throughout habituation and testing, continue this procedure, adjusting food by +/- 0.5 g when weight drops below or exceeds the threshold.
   3. During testing, weigh animals prior to task initiation, then provide allotted food immediately upon return to the cage following task completion.
      NOTE: Any sudden loss of weight (>15%) over 1-2 days could signal an underlying issue, and a veterinarian should be consulted.

4. Habituation (2 days)

1. Turn on the IR light, camera, and computer.
2. Place the training bowl with pellets at the opening slot, ensuring that the tip of the bowl extends slightly into the apparatus (approximately 2 cm initially).
3. Transfer cages to the testing room and allow rats to acclimate for 5 min.
4. Day 1: Add pair-housed rats to the chamber and allow them to explore for 20 min. Place 3-5 sugar pellets in the home cage post-testing to allow rats to become familiar with the pellets.
   ​NOTE: Pellet size of 45 mg allows for a single pellet retrieval.
5. Clean the chamber with a disinfectant wipe between rats. At the end of the day, spray with ethanol to disinfect.
6. Day 2: Habituate rodents individually to the apparatus using the same procedure (20 min). If rodents appear fearful or are avoiding the front of the apparatus and the bowl, continue habituating until this behavior resolves. Continue to give pellets as treats in the home cage (3-5 pellets).
   1. To encourage familiarity and approach to the bowl, move the elongated tube farther into the apparatus shutter progressively throughout habituation, moving the tip 0.5 cm at a time closer to the rodent to facilitate reaching.
      NOTE: Habituation, testing, and training should all be performed in the same location under red lighting to eliminate the impact of novel surroundings.

5. Shaping (10 days)

1. Turn on the IR light, camera, and computer.
2. Transfer cages to the room and allow rats to acclimate for 5 min.
3. Using the training bowl, provide 1-5 pellets along the extension, with the extension placed fully inside the apparatus on the first day. Train animals in 10-min sessions. Recording during shaping is optional.
   NOTE: To familiarize the rodents with the pellets, it is important to provide a few sugar pellets within the apparatus itself to reduce aversion to the pellets in the bowl. Rodents will likely eat the pellets within the apparatus because they are easily procurable and will recognize these pellets in the bowl. It is important to get them used to facing the correct direction of the apparatus and to get used to touching the bowl.
4. On subsequent days of training, begin retracting the bowl extension from the apparatus. This will require rodents to approach the apparatus opening at the front and encourage interaction with the extension just outside of the apparatus.
   1. The design allows for specific placement of pellets. For example, place the pellets at the end of the bowl closest to the rodent and gradually move farther away to encourage the rodent to reach out of the apparatus.
5. As rodents become more comfortable taking pellets from the edge of the extensions, even if only by mouth, begin to introduce pellets via the tweezers to encourage reaching behavior.
   1. Hold one pellet in the tweezers just outside the apparatus, allowing rodents to simply touch the tweezers. Immediately reward this behavior to encourage further reaching.
   2. As rodents become more comfortable touching the tweezers, allow them to reach a bit further out of the apparatus to touch or grasp the pellet by holding the tweezers just a bit farther from the apparatus opening.
      NOTE: Keep notes on all animal behaviors, in particular if animals are fearful of the bowl or do not reach. These animals may need additional shaping days.
6. Once rodents reach continuously and independently for a minimum of 3 days (Figure 2A), transition to the Plain bowl until reaching behavior is established (continuous reaching for 3 days with at least 10 independent reaches per day).
   1. When introducing the rodents to the Plain bowl, ensure that the bowl is filled as much as possible with pellets so the rodents can visualize the pellets over the side of the bowl.
      NOTE: Encouraging reaches using tweezers might be required at this stage.
7. Clean the chamber with a disinfectant wipe between rats. At the end of the day, spray with ethanol to disinfect.

6. Prepare for testing:

1. Testing is completed across 9 days (Figure 2A), with each rodent performing 3 trials per bowl, with each 10 min trial occurring on consecutive days.
   NOTE: Testing is best performed in the morning.
2. On each day of testing, weigh the rats. Once complete, rats are left in home cages for 30-60 min prior to testing, as weighing can induce stress and inhibit reaching behavior.
3. Turn on the computer and set up the bowl. Plug in the camera and light. Open the recording software and adjust camera settings as outlined in section 2 of the protocol.
   NOTE: Recording at least 200 frames per second (fps) to capture all paw movements is recommended.
4. Align the bowl in the frame of the camera and visualize it in the recording software by pressing the play button. Place the bowl flush with the apparatus and the camera at a 45° angle to the side of the bowl.
   1. Move the camera to capture the reaching directly head-on to study individual digit and palm movement, as well as movement related to the height of the palm above the bowl during reaching.
   2. Place the camera directly perpendicular to the bowl for capturing arm and hand movement to and from the bowl, but not digit placement or details about grasping behavior.
      NOTE: Although a 45° angle is recommended to capture movement along the x, y, and z planes, other configurations may more readily capture information for the desired study.

7. Testing (9 days)

1. Place the rats in the room to habituate for 1 h.
2. Place the first bowl on the bowl platform, securing it with dental wax. Ensure that the front of the bowl is flush and centered with the apparatus opening.
3. Name the video file within Spinview with the rat identification number, date of recording, and time of recording.
4. Place the rat in the apparatus. Close the lid and place the white cardstock on top to deflect the light.
5. At this point, double-check the camera angle to ensure the appropriate scene is captured and adjust the bowl if it is not centered.
6. Pour the correct amount of pellets into the bowl and press Start Recording.
   Day 1-3 Plain bowl: ~100 pellets
   Day 4-6 Less bowl: ~50 pellets
   Day 7-9 Plinko bowl: ~50 pellets
7. Remain in the room throughout testing to continuously monitor the rat, take notes on reaching performance and behavior, and add pellets as necessary to maintain the correct quantity.
   1. For the Plain bowl, ensure rodents can see pellets over the edge of the bowl and that pellets can always be obtained (i.e., filling the bowl throughout the trial as pellets decrease, pellets are layered).
   2. For the Less bowl: check that pellets are not stacked on one another and there is some room between pellets, allowing them to scatter when the rat attempts a reach.
   3. For the Plinko bowl: make sure there are enough pellets to be accessible between obstructions.
8. Return the animal to its home cage and provide the allotted food. Clean the chamber with a disinfectant wipe between rats.
9. At the end of the day, spray with ethanol to disinfect the entire chamber to prepare for the following day.
   ​NOTE: If testing both male and female rodents, make sure to counterbalance the sexes with some females and some males tested earlier in the morning, and a similar mix tested later. Maintain the order of testers throughout the entire experiment.

Figure 2: Training and testing of rats performing the fine motor task. (A) Timeline of experiment. Rats were habituated to the apparatus for two days then underwent shaping procedures for 10 days. Animals were tested for fine motor ability across all three bowls for three days each bowl. Then, animals were injected with Harmaline and retested on each bowl across three days. (B) Rats (n = 6) underwent two different bowls (training bowl and Plain) during shaping procedures, and the bowl type was switched on day 5 to improve reaches. There was a significant increase in shaping across 10 days to reach the criteria (F(9, 45) = 13.5, p < 0.001). The grey bar designates the switch between the original and the current elongated tube training bowl. (C) For testing all three bowls, there was a significant difference in total reaches across the 9 days of testing (F(8, 32) = 3.74, p = 0.003). Note: Rat 2 was removed after shaping. The Plinko bowl resulted in significant differences between the Less (p = 0.001) and Plain bowl (p = 0.027). Please click here to view a larger version of this figure.

8. Modeling fine motor disruption using Harmaline

NOTE: To identify the effects of motor disturbances on fine motor ability and performance in the fine motor task, rats were injected with Harmaline (10 mg/kg), a β-carboline alkaloid that is a reversible inhibitor of monoamine oxidase-A and targets the inferior olive^52,55,56,57. Harmaline in rodent models affects posture and balance, limits fine motor coordination, and induces motor disruption in the upper limbs.

1. Weigh the rats to calculate the proper Harmaline dose.
2. Allow rats to habituate to the testing room for 1 h.
3. Turn on the computer and set up the bowl. Plug in the camera and light. Open the recording software and adjust camera settings as outlined in section 2 of the protocol.
4. Inject rodents with Harmaline (10 mg/kg, i.p.).
5. After 15 min, pour the pellets into the proper bowl, and start recording. Test the rat on each bowl across 3 days.
   Day 1 Plain bowl: ~100 pellets
   Day 2 Less bowl: ~50 pellets
   Day 3 Plinko bowl: ~50 pellets
   NOTE: Monitor the injection site for abnormalities, including skin irritation, painfulbehaviors, ataxia, hypothermia, and postural changes for up to 3 h post-injection. Following the 3 h windows, continue to monitor for any changes in behavior, including latencies to eat or drink, or any residual tremor.

9. Video analysis using machine learning

NOTE: Following data collection and video recording, reaches were counted by hand and scored as follows: Success (retrieve and consume pellet), Drops (retrieves the pellet but drops the pellet prior to consumption), Failure (unable to retrieve pellet). This hand-scored data was used as ground-truth data for the machine learning analysis.

1. Using custom Python code at https://github.com/verpeutlab/ADFineMotorTask, comparisons were made between Baseline and Harmaline-treated animals to identify changes in reach performance (Figure 3, Figure 4, Figure 5, and Figure 6).
2. To analyze joint kinematics of reaching behavior, including reach trajectory, paw conformation and position, and changes in reach strategy across the varying bowl types, use the free and open-source machine learning software Social LEAP Estimates Animal Poses (SLEAP) (see https://sleap.ai/ for a more extensive explanation and a step-by-step guide to using the open-source program⁴⁹).
   NOTE: SLEAP⁴⁹ can be downloaded at https://github.com/talmolab/sleap.
3. Open a new project and add 3 videos of the reaching task.
4. Create a skeleton consisting of 15 nodes: 1 upper arm, 1 wrist, 1 palm center, 4 base, 4 knuckle, and 4 fingertip nodes (download skeleton at https://github.com/verpeutlab/ADFineMotorTask).
5. Add edges between the nodes such that the upper arm connects to the wrist, then the palm center. From the palm center, connect the base, knuckle, and fingertip nodes such that the palm center is always the starting node (Figure 7A).
6. Label a minimum of 50 frames prior to running training. Here one can train a new model or use the models for centered and centroid on the github: https://github.com/verpeutlab/ADFineMotorTask.
7. Once the model is trained, open a new video and run inference: trackin.match: greedy, n_instances:1, tracking.clean_instance_count: 1, tracking.target_instance_count: 1, tracking.pre_cull_to_target: 1.
   NOTE: Inference can be run as a bash script on a local cluster. Tracked videos may require proofreading (see https://sleap.ai/ for proofreading instructions).
8. To further analyze the nodes and understand differences between the groups in joint kinematics, use the free and open-source keypoint-MoSeq⁵⁰ (see https://github.com/dattalab/keypoint-moseq for details on download and instructions for use).
9. Create discrete syllables of behavior using keypoint-MoSeq⁵⁰ (Figure 7B). Perform additional analysis to map transition frequency, probability, and changes in the frequency of transitions between Harmaline and Baseline conditions per bowl (Figure 7C, D). See https://github.com/verpeutlab/ADFineMotorTask for analysis code.

Figure 3: Changes in reaching across bowls and treatment. (A) For baseline (n = 5) testing, across all 3 bowls, there was a significant difference in total reaches (H(df) = 8.48, p = 0.014, Kruskal-Wallis), with more reaches occurring with the Plinko bowl type (p = 0.011, Post-hoc Dunn's test with Bonferroni correction). (B) Harmaline (n = 5) perturbation did not result in differences between bowl types. Errors are reported as S.E.M. *p < 0.05. Please click here to view a larger version of this figure.

Figure 4: Comparison of performance between conditions and across bowl types. (A) There was a significant bowl (F(2) = 5.78, p = 0.007) and outcome difference (F(2) = 3.92, p = 0.029) during baseline testing. Reach attempts increased as bowl type became more difficult, although reaches were reduced from baseline across all bowls in the harmaline condition. (B) No differences were found in Harmaline-treated rats for reach counts across bowls. (C) There was a significant outcome difference (F(2) = 2.07, p < 0.001) during baseline testing when analyzed by performance. (D) No differences were found in Harmaline-treated rats for bowl or outcome performance. Errors are reported as S.E.M. Please click here to view a larger version of this figure.

Figure 5: Comparison of baseline and Harmaline conditions. (A) There were no differences in performance success between conditions. (B) For failure counts, there was a significant difference between Baseline and Harmaline conditions (H(df) = 4.85, p = 0.028, Kruskal-Wallis) with a significant effect found in the Harmaline condition (p = 0.028, Post-hoc Dunn's test with Bonferroni correction). Errors are reported as S.E.M. Please click here to view a larger version of this figure.

Figure 6: Learning and reaching rates. (A) Harmaline animals had a perturbed learning rate compared to the baseline condition (t-stat = 2.833, p = 0.022). (B) For total reaches per minute, there was a significant interaction between bowl (less) and condition (harmaline) (F(1,5) = 9.83, p = 0.002, mixed linear model) and bowl (plinko) and condition (harmaline) (F(1,5) = 22.67, p < 0.001, mixed linear model). Harmaline reduced the total reaches per minute across all bowl types (Less: p = 0.01, Plinko: p < 0.001) compared to baseline. Errors are reported as S.E.M. *p < 0.05, ***p < 0.001. Please click here to view a larger version of this figure.

Figure 7: Machine learning identifies unique poses and task strategy. (A) Node locations on the rat paw labeled using SLEAP. (B) Examples of syllables identified by keypoint-MoSeq during reaching. (C) Matrix of bigram transition frequencies between consecutive syllables. Transition probabilities are normalized by the total bigram count. (D) A transition graph comparing syllable transition differences between Harmaline and Baseline. Nodes of the graph represent syllables, and the edges indicate transitions between them. More prominent edges indicate syllable transitions with higher frequencies. Red connections indicate an up-regulated transition, while a blue connection indicates a down-regulated transition in Harmaline-treated animals. Red circles indicate up-regulated usage, while a blue circle indicates a down-regulated usage of the indicated syllable in Harmaline-treated animals. Pairs of syllables with frequencies less than 0.005 were excluded. Please click here to view a larger version of this figure.

Results

After rats reached a 90% body weight food restriction, they were habituated to the chamber for 2 days, then underwent shaping procedures for 1-2 weeks. The number of successful, dropped, and failed reaches were scored by a trained observer blind to each condition. Over the course of 9 days (Baseline), rats were tested for fine motor ability on 3 different bowl configurations (Plain, Less, and Plinko). Then, the same rats (n = 5 per group, 1 removed after training bowl shaping due to illness) were injected with Harmaline (10 mg/kg, i.p.) once a day across 3 days to determine changes in reach ability, strategy, and performance (Figure 2A). In the design and construction phases, the preliminary training bowl failed to increase the reaching ability (Days 1-4). This resulted in a redesign of the training bowl with an elongated feature, and animals began to reach consistently for a pellet (Days 5-10), reaching training criteria by Day 10 (Figure 2B). Across all bowls, individual total reaches varied, with increased reaches occurring as the task became more difficult, demonstrating persistence (Figure 2C and Figure 3A). Inducing motor impairment via Harmaline disrupted normal reaching and reduced total reaches (Figure 3B). The number of reaches increased with bowl difficulty during Baseline (Figure 4A), but did not change in Harmaline-treated rats (Figure 4B). Performance, measured by percent success, dropped, and failures, were mostly successes and drops during Baseline testing until rats reached the Plinko bowl, which increased failures (Figure 4C). Harmaline disrupted typical reach performance (Figure 4D), and while success performance did not differ between conditions, suggesting possible adaptation (Figure 5A), failures remained consistent across bowls in the Harmaline condition compared to Baseline (Figure 5B). Prior to Harmaline treatment, failures increased with bowl difficulty, as seen in the Baseline group (Figure 5B). Harmaline reduced the learning rate of the task (Figure 6A) and total reaches per minute across all bowl types (Figure 6B). Analysis of reaches using machine learning revealed unique poses and task strategy between Baseline and Harmaline conditions and within each bowl type (Figure 7A,B) using SLEAP and keypoint-MoSeq. Applying these software tools to this fine motor task, the software found unique transitions between syllables, which are defined as a stereotyped behavior or movement that the algorithm identifies and segments, and the use of specific syllables within each bowl type (Figure 7C) and between conditions (Figure 7D). Interestingly, syllables revealed less variation in grasping motions as bowls increased in difficulty, suggesting an improvement in task strategy and that unique poses are required for task success. Between conditions, Harmaline-treated rats were found to engage in different reaching strategies to match similar task performance to Baseline to possibly compensate for the induced acute neural impairment. Lastly, the fine motor task was implemented for a preclinical model of Alzheimer's disease, the TgF344-AD rat (n = 12 per genotype). Testing was done at 6 months of age to examine whether the fine motor task could detect motor deficits in Tg rats prior to the onset of neural loss and widespread accumulation of amyloid. Compared to wildtype littermates, transgenic (Tg) rats had reduced total reaches (Figure 8A). WT rats increased reaches as bowls became more challenging (Figure 8B), which was not seen in the Tg rats (Figure 8C). While performance type (success, dropped, failure) was variable within each group (Figure 8D,E), Tg rats had significantly reduced success performance compared to WT (Figure 8F). As bowls became more challenging, the number of failures increased in both WT and Tg rats (Figure 8G).

Figure 8: The fine motor task was used to examine a preclinical model of Alzheimer's disease at 6 months of age. (A) For total reaches there was a significant difference between the genotypes (H(df) = 17.6, p < 0.001, Kruskal-Wallis) with Tg rats (n = 12) having significantly less reaches than WT (n = 12) (p < 0.001, Post-hoc Dunn's test with Bonferroni correction). (B) WT rats demonstrated increased reach attempts across bowl types (F 2,33) = 4.422, p = 0.02, ANOVA). (C) No significant differences were found in reaching for Tg rats. No significant differences were found in (D) WT or (E) Tg performance across bowls. (F) There was a significant effect of genotype between WT and Tg rats for success performance (H(df) = 7.63, p = 0.006, Kruskal-Wallis). (G) For failure counts, there was a significant bowl difference, but no differences in genotype (H(df) = 25.41, p < 0.001, Kruskal-Wallis). Errors are reported as S.E.M. *p < 0.05, **p < 0.01, ***p < 0.001. Please click here to view a larger version of this figure.

Discussion

Reaching tasks for rodents are commonly restricted to single pellets and do not include obstructions or self-perturbing features, which are common struggles for patients with neurodegenerative pathology. The method presented here, using both a reimagined pellet-reaching task and machine learning, can detect both coordination and strategy through analysis of pose features and transitions, in addition to individual digit movements. The fine motor task created here detected coordination and strategy via analysis of pose and transitions between poses. A typical reaching task commonly involves a shelf upon which a single food reward is placed for the test subject to retrieve. With its reliance on a single, repetitive reaching strategy, the single-pellet iteration lacks metrics related to task difficulty, potential learning when used longitudinally, and challenges in strategy adjustment with pellet placement changes throughout testing, which are captured by this new method. This omission in the traditional single-pellet task introduces a ceiling effect in performance in which rodents plateau in both speed and accuracy due to mastery of a single, repetitive reach⁴². By varying bowl types that differ in pellet quantity and obstructions as shown here, rodents must continually update a strategy to navigate bowl variability and pellet movement as adjacent pellets are perturbed^41,58. The self-perturbing features of this task may also more readily recruit a larger network of brain regions. Variability in the placement of pellets within the bowl requires the involvement of both sensation and movement-related systems, exemplified by greater cross-communication between the cerebellum, basal ganglia, and hippocampus. This is not substantially captured by a single-pellet task, which may recruit regions involved simply in the initiation and precision of movement once the task is learned^46,59, forgoing a greater cognitive or learning component. This distinction is exemplified by the use of Harmaline. Harmaline had a notable impact on performance, specifically in learning and reaching rates from Baseline (Figure 6A). Despite this finding, successful performance remained constant, suggesting adaptation to the task. Furthermore, the performance of 6-month-old TgF344-AD rats was also impaired (Figure 8), showing additional proof-of-concept that the task may be sensitive to early motoric deficits in AD.

Preliminary iterations uncovered specific design elements that required further consideration and reconfiguration. In initial testing of bowl configurations, it was apparent that the bowl shape was crucial to the advancement of reaching behavior. Of note, reaching performance showed an upward trend following day 6 of shaping, which was the first day the rats were introduced to a revised bowl design (Figure 2B). As opposed to previous configurations that were simply circular dishes housing many pellets, this training bowl added an elongated extension, allowing for the retrieval of single pellets at a close distance during the shaping phase. The width of the extension was small enough to fit into the apparatus to allow for the movement of the bowl both farther into and farther out of the rodent's reach to promote reaching behavior and approach to the front of the apparatus. With the simple transition from a bowl with raised edges to a bowl with increased visibility and opportunity for early single-pellet reaching to learn the task, there were significant improvements in rodent behavior. Moreover, the shutter cutout at the front of the apparatus was also revised several times prior to testing. Rudimentary designs had a T-shaped cutout, allowing rodents to utilize both paws to procure pellets. Despite the added room to use both extremities, rodents were erratic in behavior, failing to grasp pellets in a manner that allowed for individual reach analysis and counting. Reaches appeared aimless until the introduction of the current shutter design, which limited the reaching space to a single, vertical opening. With the latter configuration, rodents are forced to reach one arm into the bowl, allowing for the quantification of individual reaches and the analysis of reaching pose using machine learning techniques. It is also important to take into consideration the size of the provided sugar pellets and their impact on proper analysis. The pellets utilized are manufactured in two available sizes, 25 mg and 45 mg. Both pellet sizes easily fit into the palm, but the larger size ensures that only singular pellets are retrieved. The 25 mg option allows rodents to grasp multiple pellets at once, often resulting in both a success and a drop within the same reaching bout. To eliminate this ambiguity in reach categorization, the larger 45 mg pellets are recommended to encourage rats to only grab one pellet at a time.

Additionally, it is essential that rodents are food-deprived throughout the habituation, shaping, and testing phases to maintain motivation to procure a food reward. Testing with the elimination of food deprivation yielded limited results, with rodents indifferent to approaching the front of the apparatus, unwilling to reach during the 10-min trials, and often preoccupied with grooming at the back of the apparatus, out of the camera's frame. As outlined in the protocol, rodents should be food-deprived a week prior to the start of habituation to encourage apparatus exploration and potentially limit the impact of the novel environment with the presentation of the food reward (i.e., sugar pellets) at the earliest stages. Rodents food-deprived prior to task initiation have been shown to express greater interaction with a food object along with greater exploratory behavior despite slight weight loss, without limiting locomotion⁶⁰. To further facilitate the association between the apparatus and the presentation of a food reward, pellets should be placed within the apparatus for easy access at the rodent's first exposure to the apparatus.

Several methodological limitations are pertinent to mention. With the implementation of machine learning, motor planning, learning, and pose throughout KINEMAT can be analyzed in an unbiased manner. However, the reach counts of each outcome (success, dropped, failure) were scored by several trained observers by hand, without the use of machine learning. Future experimenters can use this data as ground truth and create automated scoring methods. The reaching outcomes (i.e., success, drop, and failure) must be standardized across experimenters to ensure that reaches are properly classified. Reaching analysis could be further improved with additional cameras to capture top-down, opposite side angle, and body position throughout the task for 3D pose estimation. Further, the size of the rat, but not the strain, may pose a constraint. Animals larger than 800 g in body weight will not comfortably fit within the proposed apparatus, and those much smaller than 100 g cannot comfortably reach for the pellets with the elevated bowl without rearing. Thus, behavior may be restricted if these requirements cannot be met. Lastly, while animals were handled daily, rats did not receive injections of saline to control for injection-related impacts on behavior prior to Harmaline. Future experiments should include a saline or vehicle injection condition.

The strength of this fine motor task lies in its translational value, bridging preclinical rodent work, allowing for more invasive investigations of the neural networks imperative in disease progression and decline, with important preliminary clinical findings. Here, it is shown that the fine motor task is sensitive to changes in reaching and performance for both an acute cerebellar perturbation model and a preclinical Alzheimer's disease model, the TgF344-AD rat, specifically prior to the onset of cognitive impairments. The dynamic, self-perturbing design and innovative analytics have significant clinical value. The use of machine learning accommodates discrete types of analysis, allowing for the identification of minute fine motor differences between animals of the same strain, different conditions, and potentially different genotypes. Importantly, this task leverages prior work on rodent limb, head, and tail kinematics to study gait^32,61,62,63 and other tasks analyzing endpoint precision kinematics^44,64, particularly focusing on cerebellar deficits. With a myriad of bowl designs and the use of machine learning to examine additional features of animal pose and strategy, this task expands upon former work. With the use of 15+ nodes and principal component analysis, this task allows for both discrete and global metrics focused on single node movements as well as coordination between these nodes of interest. Future additions to this preliminary study include paw cohesion analysis and looking at the distance between digits (i.e., knuckle and fingertip separation), which has been found to be a significant indicator of localized brain atrophy in patients with AD⁶⁵. Unlike traditional tasks that assess repetitive movements, this task attempts to mirror real-world challenges encountered by patients with neurodegenerative conditions that require continuous strategy adjustment as opposed to singular, repetitive movements. By requiring animals to adapt to shifting pellet locations and bowl configurations, this paradigm models adaptability deficits often seen in human patients, providing a more comprehensive assessment of motor learning and strategy formation.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
3D printing material	Amazon (Overture)	B07PGY2JP1	PLA Filament 1.75 mm PLA 3D Printer Filament, Dimensional Accuracy +/- 0.03 mm
Camera	Edmond Optics	BFS-U3-13Y3M-C	1.3 MP, Mono, 170 FPS, ON Semi PYTHON 1300 x2
Camera cables	Edmond Optics	ACC-01-2300	USB 3.1, 3 m, Type-A to Micro-B (Locking) Cable
Camera lens	Thor labs	MVL5WA	4.5 mm EFL, F/1.4, 1/2" Format Machine Vision Lens
Camera to tripod connection	Edmond Optics	88-210	Connection for camera to tripod
Cleaning wipes	Amazon (Sani-Cloth)	B00KMZ7KMO	AF3 Surface Disinfectant Cleaner Wipe Canister Mild Scent 160 Ct P13872
Computer	Dell	Precision	Intel Core i7-13700 @ 2.1 GHz, NVIDIA T1000 8 GB, RAM 16 GB
IR light	Amazon (Serlium)	B0BJFDBT44	Camera IR Light 48 LED IR Illuminators Lights Waterproof Infrared Night Visionss Light for Security CCTV Camera
Panel glue	Amazon (Sdintar)	B0B1DLRPNZ	Glass glue
Plexiglass	Amazon (TOOLINHAND  US)	B0BRJ1L8TR	12 × 12" Clear Cast Acrylic/Plexiglass Sheets - Transparent 1/8” Thick (3 mm)
Python	Python Software Foundation	Python 3.7
Spinview	Teledyne Technologies	Version 3.1.0.79
Sugar pellets (45 mg)	Bioserv	F0023	45 mg, Unflavored 50,000/Box; Ingredients: Sucrose, Dextrose, Magnesium Stearate, Calcium Silicate, Mineral Oil
Tripod	Amazon (YOUYI)	B086PYJF9D	Amazon Flexible Desk Webcam Tripod
Tweezers	Fine Science Tools	11064-07	Iris Forceps
Wax	Amazon (Orthomechanics)	B00FKDC0UU	Genuine Orthowax