Unraveling the nuanced trajectory of recovery following spinal cord injury (SCI) has remained a pivotal quest, fostering the need of innovative approaches to offer insight into the underlying dynamics of this complex phenomenon. Pioneering efforts in the rodent model, such as the Basso Mouse Scale (BMS)1 and kinematic analyses2, have yielded invaluable insights into the intricate neural mechanisms and observable behavioral patterns indicative of recovery. However, these methods predominantly focus on isolated aspects of behavior, such as stepping patterns and joint actions, thereby limiting the comprehensive understanding of each recovery epoch in the rodent. Consequently, there remains a conspicuous gap in fully elucidating complex motor behavior, leading to an incomplete grasp of the constituent behaviors characterizing individual epochs and the transitional behaviors signaling shifts between these distinct phases.

Behavior serves as a tangible reflection of underlying neural mechanisms3–7. Therefore, leveraging behavioral patterns not only aids in predicting underlying neural changes but also serves as a valuable therapeutic strategy to enhance and optimize the recovery process. Nevertheless enhancing our comprehension of the recovery process is crucial, as it can help alleviate misinterpretations and temper overinflated expectations during the translation of research findings8. Previous literature has underscored the importance of accurately delineating recovery epochs, highlighting their pivotal role in tailoring patient rehabilitation programs and therapeutic interventions. Leveraging behavioral indicators indicative of the transitional phases between these epochs has the potential to refine the targeting of therapeutic strategies, thereby potentially optimizing the overall recovery trajectory. Even incremental advancements resulting from such interventions can yield substantial improvements in a patient's overall quality of life9, illustrating the significance of this research direction in the field of SCI rehabilitation. This study utilizes behavioral biomarkers as a tool to redefine the delineation of recovery epochs and reveal shared behavioral patterns across varying degrees of injury severity. We employed a machine learning-driven, artificial intelligence, three-dimensional video analysis technique known as Motion Sequencing10–15. MoSeq integrates depth imaging with unsupervised computational algorithms, enabling the segmentation of mouse behavior into succinct (∼350 ms) behavioral "modules," including actions like “locomotion,” “rearing,” and “grooming” that occur in the preinjury/normal condition and adaptive counterparts that arise following SCI. Modules are characterized by their stereotyped nature, reusability, and concatenation into more extensive complex behaviors. This method facilitates the phenotyping of mice, allowing for the identification of deleterious or novel behaviors that may emerge due to dysfunction within spinal cord circuits11,14, particularly in the aftermath of SCI. Importantly, MoSeq captures an array of behavioral features, including those that might evade detection when focusing solely on traditional outcome measures such as individual joint movement. Our data-driven analysis reveals that behavioral biomarkers effectively partition the recovery process into discrete epochs, indicating common movement patterns across various injury severities. Additionally, we find strong correlations between these behavioral biomarkers and established metrics of recovery assessment, including individual components of the Basso Mouse Scale (BMS), as well as histological markers quantifying injury severity. Notably, the insights provided by behavioral biomarkers offer a more nuanced understanding of movement dynamics within each specific recovery epoch.

To gain insights into the dynamics of recovery, we deconstructed complex motor behavior preand post-SCI using unsupervised machine learning approaches10–15. We employed four distinct models of SCI — Transection ("Tx") injuries, Moderate ("Mod," 50 kDyn) contusion injuries, Mild (30 kDyn) contusion injuries, and Sham injuries (laminectomy only) — as well as included a fifth group of Naive animals for comparative analysis (Supplemental Figure 1A). Naive animals did not have surgery, but they were exposed to operative anesthesia for the approximate duration of a SCI surgery as well as postoperative analgesics. Our experimental timeline included recording naturalistic behavior in 20 min bouts before surgery, 2 days post surgery, 7 days post surgery, and every week thereafter or 10 weeks total (Supplemental Figure 1B). Figure 1A illustrates our motion sequence (MoSeq) modeling process, leveraging a downwardfacing depth camera to capture spontaneous behavior in a freely behaving mouse. Each "mouse" pixel was characterized by three-dimensional coordinates (x, y, z — where z represents the mouse's height above the ground of the arena). Subsequently, a hyperparameter enabled the identification of prevalent behavioral shifts at the sub-second to second timescale, demarcating distinctive behavioral modules such as "rear," "locomotion," and "groom," among others, observed across both normal and injured mice. We confirmed the precision of our models in distinguishing between various groups and time points (Supplemental Figure 1C-E). In Figure 1B, the heatmap rows depict the distribution of module similarities with module numbers clustered into a dendrogram using a hierarchical clustering approach with Euclidean distances. The columns are ordered by module usages exhibited by each group and time point using the same clustering approach. Employing a data-driven strategy, modules naturally coalesce into distinct epochs regardless of the severity of the injury. Prominent clusters or “usage hotspots” emerge, particularly: in the heatmap's top left corner, housing modules signifying paralysis, encompassing the Tx group and early time points of the Mod and Mild groups; the middle of the heatmap's lower section, embodying modules reflective of recovery or adaptive behaviors, representing mid to late time points of the "Mod" and "Mild" groups; and lastly, the middle of the heatmap's right side accommodates clusters of animals exhibiting normal behavior, including the Sham, Naive, and the preinjury time point.

The characterization of recovery, represented at the bottom of the heatmap's central region, is more finely delineated through an analysis of the F univariate statistical test10 (Figure 1C). This analytical examination more clearly ascribes the most relevant modules for discerning differences between recovery epochs. The rationale underscores the significance of modules whose frequency of use is statistically interdependent on a particular condition, thus serving as valuable markers for linear classification and indicative of that condition. These findings reveal various facets of recovery, each distinguished by its distinctive module usage patterns and marked transitions towards modules considered "normal." Furthermore, the results illustrate that behavioral patterns observed two days after Sham injuries align with later stages of recovery from SCI, implying that laminectomies may not serve as true "sham" controls. Upon reducing the data's dimensionality to a 2-dimensional linear discriminant analysis (LDA) space, a structured representation of recovery emerges (Figure 1D and Supplemental Figure 1F). Within this spatial arrangement, modules linked with paralysis are positioned in the upper left quadrant of the LDA space, while modules associated with normal behavior occupy the upper right quadrant. Serving as the intermediary, a recovery trajectory emerges between these two quadrants, demonstrating an overlap between the moderate and mild injury groups. The LDA space mirrors the “usage hotspots” in Figure 1B’s heatmap. This insight underscores the intrinsic nature of recovery patterns, which, when projected into the LDA space, reveals distinct clusters of behaviors that persist regardless of the severity of the initial injury.

The discovery of these enduring behavioral clusters illuminates a fresh perspective on the intricacies of the recovery process. Behavioral biomarkers can thus unveil distinct recovery epochs, forming cohesive modules that transcend traditional injury severity categorizations.

The Basso Mouse Scale (BMS) stands as the established metric for quantifying and comparing recovery among groups of SCI mice1. This 9-point scale provides a coarse evaluation of locomotor capability, utilizing indicators like Plantar Stepping, Coordination, and Trunk Stability, alongside other locomotion features. Notably, our experimental groups adhered to conventional recovery patterns (Supplemental Figure 2A). To confirm the applicability of behavioral biomarkers in gauging recovery, we conducted Pearson’s correlation analysis linking BMS criteria with modules (Figure 2A). Our findings revealed a robust negative (blue) correlation between modules denoting paralysis (top left usage hotspot on heatmap) and BMS criteria indicative of normal locomotion or advanced recovery stages, including Plantar Stepping and Coordination. Similarly, we observed a significant positive (red) correlation between modules reflecting uninjured movements (bottom right usage hotspot on heatmap) and BMS criteria signifying regular locomotion or advanced recovery phases, such as "up" Tail Position and Trunk Stability. Specifically focusing on the usage hotspot in the bottom center of the heatmap, in columns representing the Mod (orange) or Mild (green) groups, we excitingly found a substantial positive correlation between modules illustrating adaptive behavior during recovery and Plantar Stepping (Right, Left) and BMS scores (Left, Right, Avg). However, this correlation was comparatively less positive, slightly negative, or absent in the case of BMS criteria Coordination, Tail Position, Paw Position, and Trunk Stability, aligning with expectations.

Indeed, during the intermediate phase of recovery, animals exhibit plantar stepping but do not typically manifest coordinated stepping, elevated tail positions, parallel paw positions, or stable trunks.

With Pearson’s correlation analysis, we explored the association of behavioral biomarkers with prevalent histological indicators linked to injury (Figure 2AB). We included Myelin Basic Protein (MBP, green)16, a marker of damaged central nervous system (CNS) myelinated axons resulting from injury, and Glial Fibrillary Acidic Protein (GFAP, red)17, a CNS astrocytic marker and a constituent of astrocytes that aggregate around the lesion post-injury, in the analysis. Additionally, we examined the inflammation-to-lesion volume ratio (MBP/GFAP ratio). We observed a notably strong positive correlation between larger MBP and GFAP volumes and behavioral biomarkers suggestive of paralysis, matching anticipated outcomes. Moreover, irrespective of the lesion volume (GFAP), increased inflammation (MBP volume) demonstrated a positive association with modules reflecting both paralysis and recovery. However, larger lesion volumes showed notably strong positive correlations primarily with paralysis modules. We conducted an analysis of entropy, representing the capacity for diverse behavioral manifestations, in relation to BMS criteria (Figure 1C) and lesion severity (Figure 1DE). In the context of SCI, greater entropy reflects a broader range of behavioral patterns, while lower entropy signifies limited behavioral variations. Post-paralysis, the reduction in behavioral diversity is observed due to the limitation in available behavioral biomarkers. Notably, we observed a significant positive correlation between higher entropy and BMS criteria linked to normal locomotion or recovery (Figure 1C). Additionally, we identified a strong association between higher entropy and less severe injuries (Figure 1D, GFAP: slope = -0.08719611; yintercept = 0.62797916), as well as reduced inflammation (Figure 1E, MBP: slope = 0.09328262; y-intercept = 0.84207945).

These findings suggest a robust correlation between behavioral biomarkers and the established gold standard metrics used to assess recovery in SCI cases. Moreover, the utilization of behavioral biomarkers offers an enriched dataset, capturing nuanced movement patterns that are often not accounted for in conventional metrics.

Our investigation sought to redefine the course of spontaneous recovery post-spinal cord injury (SCI) through the utilization of computationally identified behavioral biomarkers. Our findings unveiled persistent behavioral clusters, independent of the severity of the injury. These behavioral signatures not only exhibited correlations with conventional measures assessing functional impairments and injury severity but also provided a nuanced portrayal of the recovery process, emphasizing the considerable potential inherent in behavioral biomarkers. Delving deeper into the granularity of behavioral modules can offer insights into specific signatures crucial for recovery, even when adaptive behavior appears to stabilize. This nuanced approach challenges the conventional notion of a recovery plateau and underscores the potential for continued progress as identified through behavioral biomarkers. These modules may be positioned at the intersection of epochs, and targeting them with therapies may facilitate the transition from one phase to the next. Furthermore, establishing correlations between behavioral signatures and underlying activities, such as brain and muscle activity, alongside histological markers, presents a promising avenue for comprehending the intricate interplay between behavior and underlying neural circuit changes.

Male and female 8- to 9-week-old C57BL/6 (Strain Code 027) mice were procured from Charles River Laboratories (Wilmington, MA, USA). The mice were provided ad libitum access to food and water and were maintained in a 12-hour light and 12-hour dark cycle within the animal facility at the Division of Life Sciences, Nelson Biology Laboratories, Rutgers University. Spinal cord injury (SCI) surgeries were conducted on these mice at an age of 9-10 weeks. All housing, surgical procedures, behavioral experiments, and euthanasia were carried out in strict adherence to Rutgers University's Institutional Animal Care and Use Committee (IACUC) guidelines, with the experiments following the ARRIVE guidelines.

The mouse thoracic SCI contusion model was conducted following the protocol described previously1. Briefly, 9- to 10-week-old mice were anesthetized with isoflurane (NDC 66794-01725, Piramal Critical Care, Inc., Bethlehem, PA, USA), administered at 5% initially and reduced to 2% during surgery. A thermo-regulated heating pad was utilized to monitor and sustain the body temperature throughout the surgery. Prior to surgical incision, the upper-back skin was shaved and sanitized using alternating preparations of Betadine scrub (NDC 67618-151-32, Purdue Products L.P., Stamford, CT, USA) and 70% ethanol. Preemptive and post-operative analgesia were provided by injecting 0.1 ml of 0.125% Bupivacaine (NDC 0409-1163-18, Hospira, Lake Forest, IL, USA) subcutaneously around the incision site. The slow-release Buprenorphine formulation, Ethiqa XR (Fidelis Pharmaceuticals, LLC, North Brunswick, NJ, USA), was subcutaneously administered at 3.25 mg/kg. To prevent corneal drying, lubricant ophthalmic ointment (Artificial Tears, NDC 59399-162-35, Akorn Animal Health, Lake Forest, IL, USA) was applied to the eyes post-anesthesia induction. A 3 cm skin incision was made along the median line on the animal's back, followed by laminectomy at the vertebrae level halfway between T9 and T10. The contusion SCI was induced using the IH Infinite Horizon Impactor (Precision Systems and Instrumentation, LLC; model IH-0400). The mouse was secured on the experimental table using spinal cord stabilizing forceps, and the exposed spinal cord was precisely positioned. The impactor tip was placed over the center of the exposed spinal cord. A force of 30 kDyne was used for mild injuries, and a force of 50 kDyne was employed for moderate injuries. The muscles were then sutured in two separate anatomical layers (Suture 40 Polyglactin 910 FS-2, Cat. No. 6549052, Henry Schein, Melville, NY, USA) and the incision closed with wound clips (Reflex-9, 9 mm, Cat. No. 201-1000, Cell Point Scientific, Inc.,Gaithersburg, MD, USA). Sham-injured animals underwent laminectomy but did not receive a contusion. Animals classified as "naive" received drugs during the approximate duration of the surgery but were not incised. Animals designated for complete transections underwent transections following laminectomy using aspiration and sectioning with microscissors as previously described1,18 using great care to ensure complete severance of all axon fibers.

Following surgery, mice were individually housed in cages with white Alpha Dri bedding (W.F.

Fisher & Son, Branchburg Township, NJ, USA) to monitor for potential bladder infections and bleeding. Subsequently, mice received subcutaneous injections of cefazolin (25 mg/kg) and 0.5 ml saline once daily for three days after injury. Additionally, cages were positioned partially on 37°C thermo-regulated heating pads for the duration of the study. Monitoring of food intake was performed using Labdiet Rodent Diet 5001 pellets (Cat. No. LD5001, W.F. Fisher & Son) placed within the cages, with DietGel (Clear H2O Dietgel 76A, Cat. No. CH2ODG762OZ, Fisher & Son) provided if necessary. Daily assessments for dehydration were conducted, and HydroGel (Clear H2O Hydrogel, Cat. No. CH2OHG2OZ) was administered as needed along with supplemental saline injections. Bladder care was administered twice daily from the day after injury until euthanization to prevent urine retention. Cases of hematuria were treated with Baytril (2.7 mg/kg) over a 10-day period. Mice were withdrawn from the study if they experienced weight loss exceeding 80% of their initial body weight or exhibited severe or persistent autophagia.

At 2 days and 1-10 weeks post-spinal cord injury, locomotor recovery in mice was assessed using the Basso Mouse Scale (BMS)1. Mice were placed in the center of a small wading pool (open field) with a 44-inch diameter for 4 minutes, and hind limb movements were evaluated by two trained scorers on a scale ranging from 0 to 9. A score of 0 indicated no hind limb movement, while higher scores denoted improved hind-limb joint coordination, with 9 representing "normal" walking. In cases of discrepancy between scorers, the lower score was selected. If necessary, the sub-score table, considering parameters such as plantar stepping, coordination, paw position, trunk instability, and tail position, was utilized.

Following the final behavioral tests at 10 weeks post-SCI, mice were deeply anesthetized with 5% isoflurane (NDC 66794-017-25, Piramal Critical Care, Inc., Bethlehem, PA, USA), followed by transcardial perfusion with saline containing 10 U/mL heparin (Cat No. H3393-100KU, Sigma-Aldrich, St. Louis, MO, USA), and 4% paraformaldehyde (Cat No. 15714-S, Electron Microscopy Sciences, Hatfield, PA, USA). The spinal cord columns were dissected, and the ventral side of the spinal cord was exposed. The tissue was postfixed in 4% paraformaldehyde overnight at 4°C, subsequently washed three times with phosphate-buffered saline (PBS), and stored in PBS containing 0.01% sodium azide for further processing.

For immunohistochemistry of the lesion site, 1 cm long spinal cords containing the lesion site at the center were cryoprotected by overnight immersion in 4°C PBS with increasing sucrose concentrations (10%, 15%, 20%), beginning with the lowest sucrose concentration overnight in each solution. Finally, spinal cords were incubated overnight with 50% Optimal Freezing Temperature (OCT) (Tissue-Tek, Cat No. 25608-930, VWR, Radnor, PA, USA) medium and 50% 20%-sucrose dissolved in PBS. Following these saturation steps, the tissue was frozen in OCT medium, and 20 µm thick serial sagittal cryosections were prepared, mounted on Diamond® White Glass Charged Microscope Slides (Globe Scientific Inc, Cat. No. 1358W), and stored at -20°C for immunostaining. The slides were thawed and dried for 1 hour at 20°C, washed three times with PBS, and then blocked with 3% bovine serum albumin (BSA) (Cat No. RLBSA50, IPO Supply Center, Piscataway, NJ, USA), 10% Normal Goat Serum (NGS) (Cat No. 01-6201, Fisher Scientific, Waltham, MA, USA), and 0.3% Triton-X 100 (X100-100ml, SigmaAldrich, St. Louis, MO, USA) for 1 hour at 20°C. Subsequently, the tissue sections were costained overnight at 4°C with a primary chicken antibody against the astrocyte marker Glial Fibrillary Acidic Protein (GFAP) (1:1000 in blocking solution; Cat No. 200-901-D60, VWR, Radnor, PA, USA) and a primary rat antibody against myelin basic protein (MBP) (1:50 in blocking solution; Cat. No. MAB386, Sigma-Aldrich, St. Louis, MO, USA). Following five 5minute washes at 20°C with PBS, the tissue was incubated for 2 hours at 20°C with corresponding Alexa Fluor 488 (1:500, anti-Chicken coupled to Alexa Fluor 488, Cat No.

A11039, Fisher Scientific, Waltham, MA, USA) and Alexa Fluor 565 (1:500, anti-rat coupled to Alexa Fluor 546, Cat No. A11081, Fisher Scientific, Waltham, MA, USA) secondary antibodies diluted in blocking solution. The tissue was washed five times for 5 minutes at 20°C with PBS, incubated for 10 minutes with DAPI solution diluted 1:5000 in PBS, and the sections were finally mounted under glass coverslips with Fluoromount-G (Cat No. 100241-874, VWR, Radnor, PA, USA).

Quantitative analysis of the sagittal sections was performed using the automated microscope IN Cell Analyzer 6000 (GE Healthcare, Chicago, IL, USA). Four microscope slides were loaded into the system, and the image capturing protocol from the manufacturer was implemented. The tissue sections were precisely located via preview, and high-resolution images were acquired for subsequent analysis.

https://github.com/tischfieldlab/moseq2-detectron-extract https://github.com/tischfieldlab/moseq-reports structure of

Neurosci.

FIGURE CAPTIONS:

study: Transection (Tx), Moderate contusion (Mod), Mild contusion, Sham injury, and Naive animals. (B) Experimental Timeline. Recordings occurred on the following time points: preinjury, 2 days post-surgery, 7 days post-surgery, and every week thereafter for a total of 10 weeks. Asterisks indicate recording days for MoSeq and BMS. (C-E) Linear Discriminant Analysis (LDA) Classification Confusion Matrix (C), Validation Curve (D), and Permutation Score (E). The model was able to appropriately discriminate between groups and time points. (F) LDA dimensionality reduction. Subplots indicate different groups within the LDA space. modules illustrating adaptive behavior during recovery. This correlation was comparatively less positive, slightly negative, or absent in the case of BMS criteria Coordination, Tail Position, Paw Position, and Trunk Stability. Lesion: We observed a strong positive correlation between larger MBP (demyelination) and GFAP (astrocytic inflammation) volumes and paralysis modules. Increased inflammation (MBP volume) had a positive correlation to modules that are associated with paralysis and recovery. Larger lesions (GFAP volume) demonstrated a strong positive correlation to paralysis modules. Groups include Transection (Tx, pink), Moderate contusion (Mod, orange), Mild contusion (green), Sham injury (purple), and Naive animals (blue). time points are denoted by colors. (B) Injury severity in five groups with demyelination (MBP) volume in green and lesion volume (GFAP) in red. (C) Pearson’s R Correlations of BMS criteria vs. entropy. Entropy denotes the capacity to demonstrate diverse behaviors. We observed a positive correlation between BMS criteria linked to normal locomotion or recovery and higher entropy. (DE) Correlations of lesion severity vs. entropy. We observed a strong correlation between less severe injuries (D, GFAP: slope = -0.08719611; y-intercept = 0.62797916), as well as reduced inflammation (E, MBP: slope = -0.09328262; y-intercept = 0.84207945) and higher entropy.

Supplemental Figure 2. BMS Scores. (A) All animals follow a typical recovery curve: Transections (Tx, purple), Moderate contusions (Mod, red), Mild contusions (green), Sham injuries (orange), and Naïve animals (blue).