Animals

Female wild-type C57BL/6J mice (JAX: 000664; RRID: IMSR_JAX:000664), aged two months old at time zero of the experiment, were used in all experiments. We chose to use female mice to ensure consistent postoperative care and reduce variability associated with urinary obstruction and bladder management, which are more frequent in males following thoracic SCI. This design also minimized mortality during the acute post-injury phase and improved overall data consistency. Mice were housed in a temperature-controlled facility under a 12-hour light/dark cycle with unrestricted access to food and water. Sample sizes were determined empirically, and all animal procedures were approved by the Institutional Animal Care and Use Committee (IACUC) at UT Southwestern Medical Center. We acknowledge that the exclusive use of females represents a limitation of the present study.

Spinal cord injury model

Adult female C57BL/6J mice were anesthetized using a ketamine (100 mg/kg) and xylazine (10 mg/kg) cocktail, followed by a laminectomy at the thoracic level 9 (T9) spinal segments to expose the spinal cord. Mice were assigned to either the sham or SCI group, with each group consisting of eight animals. For the SCI group, a standardized contusion injury was induced using the IH impactor (Precision Systems and Instrumentation, Lexington, KY; Model IH-0400) with a 1-mm-diameter impactor tip, applying a force of 60 kdyn perpendicular to the central axis of the exposed dorsal spinal cord. The tip was immediately retracted after impact to prevent excessive tissue damage. Sham animals underwent identical surgical procedures, including laminectomy and impactor tip positioning, but no injury was delivered [13]. This contusion model produces a bilateral lesion with a central cavity surrounded by spared tissue, mimicking clinically relevant traumatic spinal cord injuries. Postoperatively, all mice were monitored in their home cages and provided supportive care, including manual bladder expression twice daily until reflexive bladder control was regained.

Tissue processing and immunohistochemistry

Mice were euthanized via CO₂ overdose and transcardially perfused with ice-cold phosphate-buffered saline (PBS), followed by 4% paraformaldehyde (PFA) in PBS. The entire spinal cord was carefully extracted, post-fixed in 4% PFA overnight at 4°C, and cryoprotected in 30% sucrose at 4°C. A 1.5-cm segment encompassing the injury site was sectioned at 20 µm using a Leica cryostat. Immunohistochemistry was performed following a previously established protocol [14]. Tissue sections were incubated overnight at 4°C with primary antibodies, including chicken anti-GFAP (1:1000, Abcam, ab4674) and rabbit anti-PDGFR beta (1:500, Abcam, ab32570), diluted in blocking buffer containing 3% goat serum in 0.3% PBST. Sections were then incubated with secondary antibodies for two hours at room temperature, including donkey anti-chicken Alexa Fluor 488 (1:1000, Thermo Fisher Scientific, A78948) and donkey anti-rabbit Alexa Fluor 555 (1:1000, Thermo Fisher Scientific, A31572). Confocal imaging was performed using a Nikon A1R confocal microscope with a 20 × objective, and image stitching was carried out using NIS-Elements software. ImageJ was used for fluorescence area analysis.

Behavioral assessments

All behavioral assessments were conducted by a single experimenter blinded to group assignments. To ensure consistency, animals were acclimated to the testing environment in two sessions: a one-hour habituation period on the day before testing and an additional 30-minute session immediately prior to testing. Sensory assays, including mechanical and thermal sensitivity tests, were performed at baseline (7 days prior to SCI) and weekly from 7 to 42 days post-injury (dpi). Motor assessments were conducted at baseline (7 days prior to SCI), 1 dpi, and then weekly until 42 dpi.

Von Frey mechanical sensitivity testing

Mechanical sensitivity was evaluated using the Simplified Up-Down (SUDO) method [15] to determine withdrawal thresholds to calibrated von Frey filaments. Mice were acclimated to an elevated plastic chamber with a mesh floor, allowing access to the plantar surface of the hindpaws. Filaments ranging from 0.02 g to 1.4 g were applied to the left and right hindpaws, and a positive response was recorded if the animal exhibited behaviors such as toe spreading, flinching, licking, or paw withdrawal. To minimize sensitization, a five-minute interval was maintained between applications.

Hargreaves thermal sensitivity

Thermal sensitivity was evaluated using the Hargreaves method, where animals were acclimated to a plastic chamber positioned on a heated (30°C) glass platform before testing [16,17]. A focused beam of light was directed at the plantar surface of the hindpaws to measure withdrawal latency, with beam intensity calibrated to elicit a response within approximately 10 seconds in uninjured wild-type animals. Each hindpaw was tested three times with a 5-minute interval between tests to ensure consistency. A cutoff time of 30 seconds was implemented to prevent potential tissue damage in cases of delayed withdrawal.

Motor recovery analysis

Motor behavior and recovery were assessed using the Blackbox One motion tracking system [12] in combination with Palmreader v0.7.0 and Analysis v0.3.0 software. Animals were placed in a 6 × 6-inch opaque enclosure and recorded from below for 5 minutes under low-light conditions. The system utilized frustrated total internal reflection (FTIR) technology and OptiTouch sensors to detect luminance in response to surface contact, which corresponds to the pressure applied by each limb. Concurrently, video-based pose estimation was used to track body position and movement. Palmreader software generated summary files containing averaged metrics such as paw luminance (arbitrary units (A.U.) normalized to paw area (pixels^2) as defined in the Blackbox Manual v1.5), limb separation, and limb angles, while raw tracking data was stored in HDF5 format for further analysis. To encourage continuous movement, animals were not acclimated to the testing chamber before recording.

Significant features for body tracking and luminance data were analyzed using Python code provided by Blackbox One [18]. The z-score for each feature was calculated using the standard deviation and mean values from both sham and SCI across all timepoints. Significance was determined using a two-way ANOVA and Bonferroni adjusted p-values are reported.

Skeleton mislabeling analysis

To evaluate the accuracy of skeletal pose estimation in the Blackbox system following SCI, a manual frame-by-frame analysis of skeleton data was conducted. Videos generated by Palmreader were converted into individual JPEG frames at a rate of one frame per second. Mislabeled skeleton components were identified and quantified based on predefined criteria, including (1) incorrect connections between body parts, such as a forepaw linked to a hindpaw; (2) lines appearing in empty space without corresponding body parts, such as a misidentified tail position; and (3) exaggerated skeletal extensions, such as toe lines extending beyond the anatomical boundaries of the limb.

Tracking features generation

To develop novel movement metrics from the Blackbox tracking dataset which contains coordinate values for multiple body parts recorded throughout each session. We implemented custom scripts in the R programming language. We generated a number of metrics not previously available through the Blackbox software including Occupied bins, Average speed, Average acceleration, Maximum acceleration, Acceleration standard deviation, and Fore-to-hindpaw speed ratio. The tailbase coordinate was used as the primary reference point for all tracking analyses due to its high reliability. The complete code for these analyses is available at GitHub (https://github.com/helenLaiLab/BlackBox_Analysis/tree/main/tracking_metric_generation).

1. (1) Speed Metrics: The Euclidean distance formula (Distance = √[(x₂ – x₁)² + (y₂ – y₁)²]) was used to calculate displacement between consecutive frames, which was divided by the 0.022-second frame interval (the reciprocal of the 45-frames-per-second rate) to obtain frame-by-frame speed. The average speed was computed by taking the mean of these frame-by-frame speed values derived from the tailbase coordinate. To compute the fore-to-hindpaw speed ratio, frame-by-frame speeds were similarly calculated for each paw using their coordinate values. The average speeds of the left and right forepaws were first computed separately, then averaged together to obtain the overall mean forepaw speed. The same process was applied to the left and right hindpaws to identify the mean hindpaw speed. We then divided the overall mean forepaw speed by the overall mean hindpaw speed to compute the final fore-to-hindpaw speed ratio.
2. (2) Acceleration Metrics: Acceleration was derived from the speed data of the tailbase coordinate. For each recording, frame-by-frame speed values were calculated as described above, and acceleration was computed as the rate of change in speed across consecutive frames, using the same frame interval of 0.022 seconds. This generated a continuous time series of acceleration values representing changes in locomotor velocity over time. From this data, we computed the mean acceleration, maximum acceleration, and acceleration standard deviation.
3. (3) Occupied Bins Count: The recording area of the Blackbox enclosure was divided into 2,500 equally sized bins, each with an area of 0.36 cm². For each frame, the bin corresponding to the animal’s body coordinate was identified, and the total number of unique bins entered during the session was calculated as the occupied bins count. A lower number of occupied bins reflects reduced spatial exploration, consistent with conventional open-field paradigms used to assess exploratory behavior. This approach follows recent studies that applied grid-based spatial occupancy metrics to quantify exploration in SCI mouse models [19–21].

Principal Component Analysis (PCA)

PCA was conducted using the R programming language to analyze selected significant features derived from Palmreader software and additional tracking-based metrics. The analyzed features included fore and hindpaw distances, hindpaw angles (left and right), femur width, and distance traveled. To generate the PCA plot, the selected features were averaged across all mice and time points, and the prcomp() function in R was used to compute the principal components. Principal components 1 and 2 were plotted on the x- and y-axes, respectively, with ellipses indicating time-point clusters. To assess statistical differences between clusters in PCA space, a multivariate analysis of variance (MANOVA) was performed using the manova() function in R, comparing sham and SCI groups at each time point. The resulting p-values were visualized in a bar plot. Effect size was calculated using Pillai’s Trace, a statistic included in the MANOVA analysis. The complete PCA generation script is available at GitHub (https://github.com/helenLaiLab/BlackBox_Analysis/tree/main/PCA).

Blackbox data analysis app

A user-friendly data analysis application was developed using the Shiny framework in R to facilitate processing and visualization of Blackbox motion tracking data (https://helenlailab.shinyapps.io/BlackboxDataAnalysisPlots/). The app is designed to handle various file types outputted by Blackbox and enables detailed behavioral and locomotor analyses. The raw scripts are available at GitHub (https://github.com/helenLaiLab/BlackBox_Analysis/tree/main/BlackBox_Plotting_Raw_Scripts).

The application converts the tracking HDF5 file into an R data frame using the rhdf5 package [22], allowing extraction of key locomotor features, including:

1. (1) Speed — Calculated as described in the speed metrics section.
2. (2) Likelihood — Extracted values representing the confidence of coordinate estimates from the tracking data.
3. (3) Trajectory — Visualization of movement paths, color-coded by time to reflect changes over the recording period within a scaled 15.24 x 15.24 cm arena.
4. (4) Cage Occupancy Heatmap — Generated by binning coordinate values into 2,500 regions covering the recording area, with color intensity indicating time spent in each region. Transparent areas denote regions where no movement was recorded.
5. (5) Batch Extraction of Speed Averages — Computes and exports speed metrics for selected body parts, summarizing results into an Excel file.

The app also processes Blackbox feature HDF5 files, extracting and renaming feature metrics for improved readability. It generates:

1. [1] Feature Plots — Time-series line or raster plots visualizing numeric and ethogram (Boolean) data.
2. [2] Batch Extraction of Feature Averages — Exports averaged feature values per file into an Excel summary.

Additionally, the app supports importing Blackbox summary CSV files for generating bar plots and allows PCA summary file imports for on-demand PCA plot generation. Data preprocessing is performed with the dplyr [23], readxl [24], jsonlite [25], writexl [26], and stringr [27] packages. Data visualization is performed using the colourpicker [28], ggplot2 [29], patchwork [30], randomcoloR [31], rgl [32], shinyjs [33], shinyWidgets [34] packages.

Motion Sequencing Analysis (MoSeq)

To identify nuanced behavioral motifs, the Keypoint Motion Sequencing (MoSeq) platform was utilized. Analysis was performed using the MoSeq team's Jupyter Notebook via Google Colab. The customized Colab notebooks and configuration files are available at GitHub (https://github.com/helenLaiLab/BlackBox_Analysis/tree/main/MoSeq).

Input data included tracking HDF5 files and tracking-AVI videos generated by Blackbox. The dataset was configured following the “Option 1: Setup from DeepLabCut” approach, and a subset of 19 tracking files was randomly selected for model training. PCA was performed, followed by hyperparameter tuning to determine an optimal kappa value of 10¹⁰. The dataset was then modeled using an Autoregressive Hidden Markov Model (AR-HMM), and behavioral syllables were sorted by frequency before exporting model outputs to CSV. Given the dataset size, the trained model was applied in a batch-wise manner, splitting the dataset into two halves for analysis before merging results. MoSeq-generated visualizations included trajectory plots, grid movies, and syllable similarity dendrograms. For statistical analysis, data were categorized into three groups: (1) control (sham and baseline SCI), (2) 1 dpi SCI, and (3)7–42 dpi SCI, to improve interpretability while maintaining behavioral relevance. The statistical analysis followed the MoSeq Jupyter Notebook pipeline, generating plots and summary data frames. For syllable frequency analysis, the MoSeq stats_df data frame was used to identify the ten syllables with the highest overall variability across all metrics. Variance was determined using weighted standard deviation calculations in R, with values summed across metrics for each syllable. MoSeq-generated trajectory plots were examined to assign behavioral classifications to syllables, informed by corresponding grid movies.

Statistical analysis

All statistical analyses were conducted using R and GraphPad Prism v10.4.1. Comparisons between sham and SCI groups were performed using unpaired, two-tailed Student’s t-tests for single-variable comparisons. For multi-variable and time-series data, two-way ANOVA with Tukey post hoc corrections was applied to control for multiple comparisons. Data are presented as mean ± standard error of the mean (SEM). Statistical significance was defined as p < 0.05, and figures were generated using GraphPad Prism to visualize recovery trajectories and group differences.

Characterization of the SCI model and injury lesion

In a T9 contusion SCI model, we assessed motor and sensory recovery in female C57BL/6J mice one day post-injury (1 dpi) and then every week up to 6 weeks post-SCI compared to sham controls (Fig 1A). As part of this recovery analysis we utilized the Blackbox system, which allows simultaneous bottom-up video recording of up to four animals and produces skeleton-based motion tracking data as well as luminance-based paw pressure data (Fig 1B). To validate our contusion SCI model, we conducted histological analyses at 42 dpi using immunohistochemistry (IHC) (Fig 1C-E). We determined the extent of tissue damage in transverse spinal cord sections stained with DAPI (nuclear marker), GFAP (astrocyte marker), and PDGFRβ (fibroblast marker). GFAP staining marks the glial scar, which plays roles in both neuroprotection and regeneration inhibition [35], and PDGFRβ staining marks the fibrotic scar, which forms a dense extracellular matrix barrier that inhibits axonal regrowth [36]. Representative images from sham and SCI groups highlight substantial structural differences (Fig 1C). The sham control group exhibited well-preserved spinal cord architecture with uniform GFAP expression, whereas the SCI group demonstrated a prominent fibrotic lesion core with a significant accumulation of PDGFRβ+ fibroblasts encased by a dense GFAP+ astroglial scar. Quantitative analysis of the injury area in four animals, revealed a significantly larger fibrotic lesion in the SCI group compared to sham controls (p < 0.0001) (Fig 1D), confirming extensive tissue damage post-injury.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Experimental design and verification of glial and fibrotic scar formation after spinal cord injury (SCI).

(A) Experimental schematic: Wild-type mice were subjected to SCI at day 0, and spinal cord tissue was analyzed at 42 dpi using IHC. Hargreaves and von Frey behavioral assays were assessed at baseline (BL) (7 days prior to SCI) and every 7 days for 6 weeks. Blackbox was assessed at BL, 1 dpi and 7 day intervals for 6 weeks. (B) Blackbox system: Allows simultaneous bottom-up video recording of up to four animals in individual chambers. The recordings are then analyzed by the Blackbox system to produce skeleton-based motion tracking data and luminance-based paw pressure data. (C) SCI induces astrocyte and fibroblast activation, leading to glial and fibrotic scar formation. Representative transverse spinal cord sections stained with GFAP (green, astrocytes), PDGFRβ (red, fibroblasts), and DAPI (blue, nuclei). Sham-operated mice exhibit preserved spinal cord structure, while SCI sections demonstrate a prominent lesion core marked by increased GFAP and PDGFRβ expression. Scale bar = 100 µm. (D) Quantitative analysis of the PDGFRβ+ injury area (µm²) indicates a significantly larger lesion size in the SCI group compared to sham controls (****p < 0.0001, two-tailed unpaired t-test, n = 4 per group). Data corresponds to the largest PDGFRb area such as the representative SCI image in (C). (E) Higher magnification views of the lesion microenvironment. In SCI conditions, GFAP-positive astrocytes (green) form a glial scar encasing the lesion site, and PDGFRβ-positive fibroblasts (red) are enriched within the lesion core, contributing to fibrotic scar formation. The dashed line outlines the lesion boundary. DAPI (blue) highlights the distribution of nuclei. Scale bar = 50 µm.

https://doi.org/10.1371/journal.pone.0343998.g001

Further investigation into the lesion microenvironment (Fig 1E) revealed PDGFRβ-positive fibroblastic cells were enriched within the lesion core, consistent with fibrotic scar formation, whereas GFAP upregulation was most prominent in reactive astrocytes in the peri-lesional tissue and at the lesion border, underscoring the dual nature of the injury response. This spatial distribution reflects the classic organization of the injury site, in which a PDGFRβ-positive fibrotic core is surrounded by a GFAP-positive astroglial scar that demarcates the lesion, limits the spread of inflammation, and stabilizes adjacent spared tissue [35–38]. These findings confirm the successful establishment of a contusion SCI model, characterized by hallmark histopathological features of fibrotic and glial scar formation.

Mice exhibit postural adjustments during motor recovery

Postural impairments following SCI were revealed using the automated machine learning-based whole body pose estimation tracking data within the Blackbox system (Fig 2A). Analysis of the skeleton-based tracking features revealed several parameters in posture and locomotion that were significantly different upon SCI (Fig 2B, standard summary metrics described on the right side of the heatmap are automatically generated by the Blackbox Palmreader software). These include distance between various body parts, paw and limb angles, and distance traveled. We followed up on several of these features to understand how posture and locomotion change post-SCI.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Changes in animal posture and locomotion following SCI.

(A) Representative skeletal visualizations of animal posture at baseline (BL), 1-day post-injury (dpi), and 28 dpi in sham and SCI-treated animals. Skeletal representations illustrate limb positioning and coordination changes over time. (B) Hierarchical heatmap of the mean z-score for all detected features automatically provided in the standard summary metrics output of the Blackbox Palmreader software utilizing pose estimation skeleton tracking. Significance is represented as *** = p < 0.001, ** = p < 0.01, and * = p < 0.05. A sample of the significantly different postural and motor parameters are: (C) Distance between forepaws; (D) Distance between hindpaws; (E, F) Right and Left Hindpaw angles, measured as the angle between the intersection of a vector from the tail base to the sacrum and a vector from the heel to the center of the paw; (G) Femur width, (H) Total distance traveled during the recording session, representing overall locomotor activity. Tracking data was also used to analyze additional parameters: (I) Occupied bins counts reflecting exploratory behavior and movement patterns; (J) Average speed in SCI animals over time. A two-way ANOVA is used in (B). A two-tailed unpaired Student’s t-test for comparisons between sham and SCI at specific timepoints is used in (C-J). Abbreviations: Left and Right Forepaw (LFP, RFP), Left and Right Hindpaw (LHP, RHP).

https://doi.org/10.1371/journal.pone.0343998.g002

At baseline, both sham and SCI animals displayed similar postural alignment. However, by 1 dpi, SCI-treated animals exhibited noticeable asymmetry. SCI-treated animals consistently positioned their forepaws farther apart compared to sham animals, with a significant increase observed at 7 dpi (p = 0.0266) that persisted through 42 dpi (p = 0.0202; Fig 2C), likely reflecting a compensatory mechanism to shift weight forward and stabilize the forelimbs due to hindlimb deficits. Conversely, the distance between hindpaws significantly decreased at 14 dpi (p = 0.0019) and remained diminished compared to sham controls throughout the 42-day study (p = 0.0004; Fig 2D). Similarly, the distance between the right and left femurs initially increased at 1 dpi (p = 0.0002) but subsequently decreased by 14 dpi (p = 0.0003), with reductions persisting through 42 dpi (p = 0.0110; Fig 2G), indicating restricted hindlimb range of motion. The increasing trend observed in femur width in shams, as well as in SCI animals from 14 to 42 dpi, can be ascribed to the increasing mass as the mice age. We found that the femur width is positively correlated with mass (Correlation Coefficient: 0.92, R² = 0.85), which increases as they age (data not shown). Consistent with this, the femur width in both sham and SCI animals steadily increases as time goes on after SCI.

Additionally, both left and right hindpaw angles (hindpaw-tail base-sacrum angle) were significantly reduced in SCI-treated animals relative to sham controls. The right hindpaw angle showed a significant reduction at 1 dpi (p = 0.0308) and exhibited partial recovery over time, though transient reductions remained significant throughout recovery (Fig 2E). Similarly, the left hindpaw angle was significantly reduced at 1 dpi (p = 0.0120) and, despite partial recovery, remained significantly lower at 42 dpi (p < 0.0001; Fig 2F). These findings suggest a compensatory medial rotation of the hindpaws post-SCI. We analyzed whether any of these postural adjustments (i.e., forepaw distance, hindpaw distance, right hindpaw angle, left hindpaw angle, and femur width) correlated with fibrotic scar area. We did not find any positive correlation with scar area (data not shown); however, we were likely underpowered for this analysis (n = 4 SCI animals) and the scar measurements were taken only at the chronic endpoint (42 dpi). In addition, fibrotic scar area represents only one component of lesion pathology. Prior studies have shown that recovery of coordinated locomotion depends more strongly on ventrolateral white matter sparing and propriospinal circuit preservation, which are not captured by scar size alone [39–41]. Thus, a single histological measure at the lesion site may not fully capture the distributed neural circuits underlying recovery.

We note that we performed this analysis with the original version of the Blackbox machine learning-based whole body post estimation software. With this version, we observed that the software made transient errors in pose estimation during the acute phase of SCI recovery (S1 Fig). Blackbox continues to improve the DeepLabCut model and future versions of the software will take SCI models into account.

Changes in locomotor activity and coordination during motor recovery

Body part tracking data also revealed changes in locomotor activity and coordination. The total distance traveled in 5 min. was substantially decreased at 7 dpi compared to sham controls (p = 0.0107) and remained until 42 dpi (p = 0.0007; Fig 2H). In addition to the standard summary metrics generated by the Blackbox platform, we reanalyzed the tracking coordinate data to calculate novel parameters, including exploratory behavior (occupied bins count), average speed, fore-to-hindpaw speed ratio, average acceleration, maximum acceleration, and acceleration standard deviation (Fig 2I–J, S2 Fig associated with Fig 2). We found significant differences in occupied bins counts, average speed, and fore-to-hindpaw speed ratio (Fig 2I–J, S2D Fig associated with Fig 2). Occupied bins count, a metric representing the level of exploratory behavior and arena coverage, was consistently lower in SCI animals compared to sham across most time points (Fig 2I) (1 dpi p = 0.0009; 7 dpi p = 0.0011; 14 dpi p = 0.0011; 21 dpi p = 0.0004; 28 dpi p = 0.0014; 42 dpi p = 0.0012), indicating less exploratory behavior in SCI mice. Average speed (Fig 2J) was decreased across most time points in the SCI group compared to sham (1 dpi p = 0.0419; 7 dpi p = 0.0416; 21 dpi p = 0.0010; 28 dpi p = 0.0002; 42 dpi p = 0.0071). Such speed loss is reflective of the decreased motor activity exhibited by SCI mice and provides an important indicator of the extent of locomotor impairment in SCI animals. The fore-to-hindpaw speed ratio underscored the coordination deficits in SCI animals (S2D Fig associated with Fig 2). At 1 dpi, the SCI group exhibited a significant decrease in this ratio compared to sham (p = 0.0071), indicating a loss of coordinated movement immediately following injury. The ratio remained relatively consistent in the SCI and sham groups after 1 dpi with a notable deviation at 35 dpi (p = 0.0328). No differences were found in average acceleration, maximum acceleration, or acceleration standard deviation (S2A-S2C Fig associated with Fig 2). This indicates that the ability of mice to accelerate and the variability in that acceleration is not significantly affected by SCI under the conditions tested. These additional metrics provide deeper insight into the locomotor and behavioral differences between sham and SCI groups.

Sustained weight distribution alterations following SCI

Utilizing luminance data, which reflects the pressure (or weight) placed onto each paw normalized to paw area, we observed a persistent shift in weight distribution to the forepaws in SCI-treated animals. At baseline, both sham and SCI-treated animals exhibited uniform weight distribution across all limbs (Fig 3A). At 1 dpi, SCI-treated animals exhibited reduced hindlimb luminance intensity, indicative of decreased weight bearing on the hindlimbs. This reduction correlated with hindpaw dragging (Fig 3A, white arrows) and compensatory forelimb support (Fig 3A, white arrowheads). By 28 dpi, hindlimb weight-bearing improved in most animals, with weight placed on all four paws (Fig 3A). Analysis of the detected luminance-based features revealed that this post-injury shift in weight-bearing is primarily observed in fore and hindpaw luminances (Fig 3B, standard summary metrics described on the right side of the heatmap are automatically generated by the Blackbox Palmreader software). We analyzed the luminances in more detail to understand precisely how the animals altered their weight distribution after injury. While sham animals maintained stable weight distribution throughout the study, weight distribution still favored the forepaws after SCI as reflected in the average fore-to-hindpaw luminance ratio (Fig 3C). At 1 dpi, a significant shift to the weight-bearing on the forepaws is seen as an increased fore-to-hindpaw ratio in SCI-treated animals (p = 0.0041). Despite partial recovery, this fore-to-hindpaw ratio remained significantly elevated compared to sham controls until 42 dpi (p = 0.0007) (Fig 3C). The increase in the fore-to-hindpaw luminance ratio indicates a redistribution of pressure toward the forepaws, suggesting that SCI-treated animals shift their body weight forward.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Changes in paw pressure following SCI.

(A) Representative paw pressure luminance in Sham and SCI-treated animals at baseline (BL), 1-day post-injury (dpi), and 28 dpi, illustrating the distribution of weight-bearing across all four paws, with changes in intensity reflecting shifts in pressure distribution. (B) Hierarchical heatmap of the mean z-score for all detected features automatically provided in the standard summary metrics output of the Blackbox Palmreader software using luminance detection. Significance is represented as *** = p < 0.001, ** = p < 0.01, and * = p < 0.05. Several luminance values were assessed during locomotion: (C) Average fore-to-hindpaw luminance ratio over time. This ratio represents the relative pressure applied to the forepaws compared to the hindpaws. (D-E) Luminance (arbitrary units (A.U.)/pixel^2) of the right and left forepaws over time, measured as an indicator of weight-bearing on each forelimb. (F-G) Luminance of the right and left hindpaws over time, reflecting changes in hindlimb weight support. Data are shown as mean ± SEM for n = 6 Sham animals and n = 8 SCI animals. Two outliers in the Sham group over three standard deviations from the mean were removed because of urine interfering with the luminance signal. Statistical significances using a two-way ANOVA are shown for (B) and using two-tailed unpaired Student’s t-test for (C-G) comparisons between Sham and SCI groups at specific time points.

https://doi.org/10.1371/journal.pone.0343998.g003

Analysis of individual forepaw and hindpaw luminance data indicates that the increased fore-to-hindpaw ratio is largely driven by increases in forepaw pressure. The right and left forepaw luminances increased upon SCI (Fig 3D–E). These findings indicate increased weight-bearing on the forepaws following SCI. Conversely, the right and left hindpaw luminances significantly decreased at 1 dpi (p = 0.0003 and 0.0006, respectively), but were relatively unchanged for the remaining days post injury with the exception of the right hindpaw at 28 dpi (p = 0.0300) (Fig 3F–G). The sustained increase in pressure on the forepaws with relatively minimal decreased pressure on the hindpaws, highlights a persistent shift in weight distribution toward the forepaws immediately after SCI, which is maintained even after the animals have visually regained normal mobility.

Principal component analysis of behavioral parameters reflects late-stage motor recovery differences between SCI and Sham controls

Single-metric plots provide insight into specific parameter differences but lack a holistic representation of behavior. This limitation prevents a comprehensive understanding of overall changes in motor function. To address this, we implemented principal component analysis (PCA) to condense six selected significant motor metrics (i.e. fore and hindpaw distances, hindpaw angles (left and right), femur width, and distance traveled) into principal component space (Fig 4A). PCA components 1 and 2 capture the majority of the variance (Fig 4B). Additionally, to assess statistical significance between groups at each time point, we conducted multivariate analysis of variance (MANOVA).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Principal Component Analysis (PCA) of locomotor and behavioral features following SCI.

(A) PCA plot of selected significant features from individual mice across all time points. Each point represents a single mouse, with numbers and colors indicating the experimental group and time post-injury (see legend). The x-axis represents Principal Component 1 (PC1) and the y-axis represents Principal Component 2 (PC2), capturing the primary dimensions of variance in the dataset. Ellipses illustrate grouping trends among sham and SCI animals across different time points. (B) Scree plot showing the percentage of variance explained by the first six principal components. The x-axis represents the principal component number, and the y-axis represents the percentage of variance explained, indicating the relative contribution of each component to the overall variance in the dataset. Data are presented for n = 8 animals per group. Statistical significances using MANOVA are indicated between sham and SCI groups across time points.

https://doi.org/10.1371/journal.pone.0343998.g004

Throughout the study period, sham mice consistently clustered together across all time points, with baseline SCI mice localizing within a similar principal component (PC) space, suggesting uniform motor activity across uninjured animals. This observation was confirmed through MANOVA analysis, which showed no significant difference between the baseline sham and baseline SCI groups. At 1 dpi, SCI mice exhibited strong deviations in PC space localization, with many data points diverging substantially from both the sham and 7–42 dpi SCI clusters. This widespread distribution indicates a drastic and variable motor response to spinal cord injury. MANOVA analysis confirmed a statistically significant difference between SCI and sham animals at 1 dpi (p = 0.0004), further supporting this finding. By 7 dpi, although variability decreased, injured mice still exhibited dispersed localization within the PCA plot, reflecting persistent fluctuations in motor function and recovery rates (p = 0.0062). Differences between SCI and sham persisted for up to 42 dpi, with 7–42 dpi SCI forming distinct, separate cluster, indicating persistent motor differences despite visual motor recovery. These differences were validated through MANOVA, which demonstrated statistically significant difference between SCI and sham mice at each post-injury time point (14 dpi: p < 0.0001; 21 dpi: p < 0.0001; 28 dpi: p < 0.0001; 35 dpi: p = 0.0002; 42 dpi: p < 0.0001) (Fig 4A). Finally, MANOVA analysis confirmed a significant difference between SCI animals at baseline versus 42 dpi (p < 0.0001), indicating that even at the latest time point, SCI mice exhibited persistent motor impairment (Fig 4A). We also calculated effect sizes, which consistently revealed divergence between SCI and sham groups following injury (Baseline: 0.0474; 1 dpi: 0.6989; 7 dpi: 0.5424; 14 dpi: 0.7964; 21 dpi: 0.9656; 28 dpi: 0.8645; 35 dpi: 0.7315; 42 dpi: 0.8247). Therefore, by consolidating multiple metrics, PCA enabled us to uncover key recovery patterns that were not apparent when analyzing individual features in isolation. This approach underscores the importance of multi-metric data integration in capturing nuanced behavioral changes and advancing our understanding of spinal cord injury recovery.

Keypoint MoSeq facilitates precise identification and analysis of complex motifs

We implemented Keypoint MoSeq [42] to identify nuanced, recurrent behaviors and extract deeper behavioral patterns from Blackbox-derived tracking data. This analytical approach provided an additional layer of behavioral detail, enabling the identification of subtle yet functionally significant changes in locomotion and motor patterns following SCI. MoSeq analysis enabled syllable-to-behavior mapping, through which we identified several distinct behavioral motifs (syllables) that were significantly less frequent among experimental groups (Control: sham & baseline SCI; 1 dpi SCI; 7–42 dpi SCI). These included forward acceleration, right and left turns and grooming behavior (Fig 5). Syllables 0, 1, 4, and 19 corresponded to forward acceleration trajectories, syllables 9, 10, and 11 to turn trajectories, and syllables 14 and 22 to grooming trajectories (Fig 5B). The following syllables were significantly decreased between Control and 1 dpi mice (syllable 0: p < 0.0001; syllable 1: p < 0.0001; syllable 4: p < 0.0001; syllable 10: p = 0.0145; syllable 11: p = 0.0284), and the following between the Control and 7–42 dpi groups (syllable 0: p < 0.0001; syllable 1: p < 0.0001; syllable 4: p < 0.0001; syllable 9: p = 0.0103; syllable 10: p = 0.0025; syllable 11: p < 0.0001; syllable 22: p = 0.0383). In addition, MoSeq can pick up on subtle behavioral differences between seemingly similar behavioral motifs. For example, syllables 14 and 22 are both designated as “grooming.” However, their syllable trajectories are slightly different (Fig 5B), resulting in syllable 22 being significantly different while syllable 14 is not. Similarly, syllables 0, 1, and 4 (forward acceleration) are significantly changed, while syllable 19 is not. The pattern across motifs indicates that animal motor activity is significantly impaired at 1 dpi, and despite apparent recovery, significant deficits persist into the later stages. By integrating behavioral motif frequency analyses, we uncovered distinct injury-related changes in movement execution.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Keypoint MoSeq analysis reveals behavioral motifs and motor function patterns following SCI.

(A) Frequency of selected behavioral motifs (syllables) in control, 1 dpi, and 7–42 dpi groups. The x-axis represents individual syllables corresponding to distinct movement patterns, while the y-axis represents the frequency of each syllable occurrence. (B) Syllable trajectory plots and syllable-to-behavior mapping. The left panel shows visualized movement trajectories associated with key syllables, depicting characteristic motion patterns. Different colored points represent different body parts, with orange indicating mouse head, red indicating central body parts and tail, and yellow indicating paws. The right panel presents a table mapping each syllable to its corresponding behavioral classification. Data are presented for n = 8 animals per group. Statistical significances using a two-way ANOVA with Tukey post hoc correction are shown for comparisons between sham Control and SCI groups at specific time points.

https://doi.org/10.1371/journal.pone.0343998.g005

Blackbox data analysis application

To provide users of Blackbox with an external application for more complex data analyses, we implemented the Shiny framework [43] using the R programming language [44] to develop a user-friendly, web-based application (Fig 6, https://helenlailab.shinyapps.io/BlackboxDataAnalysisPlots/). Our data visualization and analysis tool provide a comprehensive suite of functionalities for the three types of Blackbox output data (Tracking, Features, and an aggregate summary .csv file). It enables the visualization of body-part coordinate-derived metrics such as speed plots (Fig 6A), feature-derived time series raster plots (Fig 6B), mouse trajectory visualizations plots (Fig 6C), and cage occupancy heatmaps (Fig 6D). Additionally, the tool facilitates simultaneous alignment and overlaying of these plots and allows users to download generated plots and data. Moreover, users can batch extract speed and feature metric averages, construct summary bar plots, and generate adaptable PCA plots (Fig 6E) for dimensionality reduction and visualization purposes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Blackbox Data Analysis App enables detailed, multifaceted behavioral data analysis.

(A) Time series speed plot showing frame-by-frame analysis of animal speed over time. The x-axis represents time (s), and the y-axis represents speed (cm/s). (B) Time series feature plot displaying the progression of a selected behavioral feature across frames. The x-axis represents time (s), and the y-axis represents the feature value. (C) Mouse trajectory plot, visualizing movement paths within the testing environment. The x- and y-axes represent spatial dimensions (box length and width in cm), while color denotes the recording frame (blue = early in session, red = later in session). (D) Cage occupancy heatmap, showing spatial distribution of time spent in different areas of the enclosure. The x- and y-axes represent cage dimensions (cm), and the color scale represents occupancy density (blue = less time, red = more time). (E) Graphical user interface (GUI) of the Blackbox Data Analysis App, illustrating user input options for data selection, metric extraction, and visualization tools. This figure highlights the functionality of the Blackbox app (https://helenlailab.shinyapps.io/BlackboxDataAnalysisPlots/) for in-depth movement tracking and behavioral quantification.

https://doi.org/10.1371/journal.pone.0343998.g006

By leveraging these comprehensive features, our tool enables researchers to seamlessly transform Blackbox system body-part tracking data and behavioral time-series data into intuitive visual representations without requiring prior coding experience. It allows for rapid assessment of mouse movement over the time course, serving as both a quality control measure and a way to identify novel features. This tool can enhance the interpretation of dynamic data patterns while facilitating deeper exploration of experiments.

Thermal and mechanical sensitivity returns to baseline after SCI

To assess whether the postural adaptations detected via the Blackbox were due to pain, we tested thermal and mechanical sensitivity following SCI. Thermal sensitivity was evaluated using the Hargreaves assay at baseline and weekly from 7 to 42 days post-SCI. SCI-treated animals exhibited a significant increase in thermal withdrawal latency at 7 and 14 dpi in the average of left and right paw data, suggesting reduced thermal sensitivity, although motor impairment likely drives the increased withdrawal latency (p < 0.0001; Fig 7A). This increase in withdrawal latencies was observed bilaterally in the hindpaws, with the left paw having significantly higher latencies through 21 dpi and the right paw having a higher latency at 7 dpi (Fig 7B-C). The bilateral reduction in thermal withdrawal latency is consistent with the bilateral contusion model used (Fig 1). By 21 dpi, thermal withdrawal latency returned to baseline levels and remained stable through 42 dpi, indicating a full recovery of thermal sensitivity and ability to withdraw their hindpaw in SCI animals.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Recovery of Thermal and Mechanical Sensitivity Following SCI.

(A, B, C) Thermal sensitivity recovery, measured by withdrawal latency (seconds) using the Hargreaves assay. (A) Average withdrawal latency, (B) left paw withdrawal latency, and (C) right paw withdrawal latency are shown over the recovery period. The x-axis represents days post-injury (dpi), and the y-axis represents latency in seconds. (D, E, F) Mechanical sensitivity recovery, measured by withdrawal thresholds (grams) using the simplified up-down (SUDO) von Frey method. (D) Average withdrawal threshold, (E) left paw withdrawal threshold, and (F) right paw withdrawal threshold are shown over the recovery period. The x-axis represents days post-injury (dpi), and the y-axis represents withdrawal threshold in grams. Data represent n = 8 animals per group, with mean ± SEM displayed. Statistical significances using two-tailed unpaired Student’s t-tests are shown for comparisons between Sham and SCI groups at specific time points.

https://doi.org/10.1371/journal.pone.0343998.g007

Mechanical sensitivity was assessed using the Simplified Up-Down (SUDO) von Frey method at the same timepoints [15]. Similar to thermal sensitivity, SCI-treated animals displayed a significant increase in withdrawal threshold to mechanical stimuli (Fig 7D), suggesting reduced mechanical sensitivity. However, similar to the thermal sensitivity assay, this increased withdrawal threshold is likely driven by motor impairment. At 7 dpi, the average withdrawal threshold of both hindpaws is significantly increased in SCI mice (p = 0.0004, Fig 7D), which can be seen individually in both the left (p = 0.0033, Fig 7E) and right hindpaws (p = 0.0550, Fig 7F). In contrast to thermal withdrawal latencies, mechanical withdrawal thresholds were variable over the time course of recovery.

Advantage of quantitative motor behavior assessments

Our findings build upon previous research utilizing conventional SCI assessment tools such as the BMS locomotor scale and CatWalk system. While these approaches provide valuable insights into gross motor function, they often lack the sensitivity to capture nuanced movement deficits [5,10]. The BMS scale is limited by inter-rater variability and an inability to detect subtle locomotor compensations [45], whereas CatWalk is limited by its constrained pathway, which may not reflect natural movement dynamics [46]. By employing high-speed video tracking, machine learning-based body part estimation, and weight-bearing metrics, we overcame these limitations and identified sustained postural adaptations and gait abnormalities that persisted beyond the expected recovery period. Our findings are consistent with recent studies showing that deep learning-based kinematic analysis can reveal subtle locomotor deficits overlooked by traditional assessments such as the BMS [47,48]. Ascona et al. found that the average mouse speed decreased after SCI, similar to our findings, and Eisdorfer et al. used MoSeq to identify changes in behavioral motifs during the SCI recovery period. Indeed, quantitative motor assessments after SCI will soon become the gold standard. Our study complements these studies by leveraging body-part tracking data to quantitatively capture logical postural and weight-bearing metrics that change post-SCI.

Sensory recovery post-SCI

Human patients can experience a loss of sensation and the development of allodynia and hyperalgesia after SCI [2,3]. While thermal and tactile allodynia has been reported in rats post-SCI [49–51], studies in mice yield conflicting results, with some strains exhibiting allodynia and others not [52,53]. Mouse strain-dependent variability and injury severity may explain some of the disagreement in sensory recovery. BALB/c mice develop tactile allodynia and spontaneous below-level pain behaviors post-SCI [54]. C57BL/6 mice are also reported to develop thermal and tactile allodynia; however, one study found tactile allodynia developed only after severe contusion injury [19,53–55].

We find that thermal and mechanical sensitivity recovers to baseline without subsequent hypersensitivity in C57BL/6J female mice. Even though we do not detect hypersensitivity after SCI, our results align with previous findings that thermal sensitivity typically recovers more reliably than mechanical sensitivity, which exhibits greater variability [55]. This difference could arise from distinct neural mechanisms, particularly within the spinal cord, that differentially influence the recovery of thermal and mechanical sensitivity following SCI [52,56]. Notably, studies have shown that persistent mechanical hypersensitivity following SCI may result from maladaptive plasticity within dorsal horn circuits, including disinhibition and reorganization of excitatory and inhibitory neuronal networks [57].

These findings underscore the need for comprehensive, multimodal sensory assessments to fully capture post-injury sensory dysfunction and highlight the limitations of evoked pain behaviors as reliable indicators of neuropathic pain in mouse SCI models. Furthermore, studies using mouse grimace scales (MGS) have suggested transient pain-like behaviors primarily within the first week post-injury, with minimal chronic hypersensitivity [58,59]. These uncertainties evaluating pain perception in mice reinforce the necessity of refining behavioral pain assessments and integrating additional physiological measures, such as calcium imaging and electrophysiological recordings, to provide a more comprehensive evaluation of pain processing after SCI.

Behavioral motif analysis

Beyond traditional motor and sensory analyses, we leveraged advanced computational techniques, including Keypoint MoSeq [42] and PCA-based population analyses, to extract deeper behavioral insights. MoSeq revealed distinct shifts in locomotor motifs across injury groups, such as decreased frequency of forward accelerations and turns, highlighting persistent post-injury changes. This aligns with previous studies demonstrating that rodents with SCI modify their movement patterns to maintain stability and mitigate functional loss [60]. Interestingly, in our study, even though average acceleration, maximum acceleration, and acceleration standard deviation were unchanged, MoSeq identified that the forward acceleration motifs were less frequent. Again, highlighting the benefit of machine-learning algorithms to identify subtle behavioral changes. Meanwhile, PCA revealed that despite improvements in gross locomotion, SCI-treated animals retained distinct motor profiles separate from sham controls, underscoring persistent postural and neuromuscular differences [61]. This is consistent with prior reports showing that, although behavioral improvements are observed following SCI, underlying motor control deficits often persist, indicating incomplete functional recovery.

A recent study by Eisdorfer et al. [48] also applied MoSeq-based behavioral decomposition to evaluate recovery after T9 spinal cord injury. However, direct comparison is limited by methodological differences: Eisdorfer et al. employed depth MoSeq, which classifies behavioral syllables using 3D body-shape dynamics, whereas our study used Keypoint MoSeq, which operates on 2D skeletal coordinate trajectories. Despite this difference, both studies converge on the observation that locomotor-related behaviors remain impaired after SCI, particularly in turning and forward-motion states. Notably, Eisdorfer et al. reported increased grooming after injury, which they attributed to sensory hypersensitivity, while our C57BL/6J female contusion model showed a mild reduction in one grooming-associated syllable (syllable 22) and no change in another (syllable 14) with no evidence of hypersensitivity (i.e., full recovery of thermal and mechanical sensitivity by 21 dpi). Altogether, the difference in grooming behaviors between the two studies (increased in Eisdorfer and decreased in ours) is not easily explained but could be attributed to strain- (C57BL/6N vs. C57BL/6J) [62–64] or model-specific differences (50 kDyn vs 60 kDyn contusion). Together, these comparisons emphasize that cross-study MoSeq syllable interpretations require caution, as unsupervised clustering in distinct feature spaces (2D vs. 3D) can yield divergent yet complementary insights into SCI-induced behavioral adaptations.

Study considerations

While our study provides a comprehensive evaluation of SCI recovery, several considerations should be acknowledged. First, our analysis was conducted in a single mouse strain and injury severity, which may limit generalizability to other models. Future studies should explore whether these findings hold across different genetic backgrounds and injury paradigms [65]. Investigating whether similar postural adaptations and compensatory mechanisms occur across diverse injury paradigms, including different severities of contusion, hemisection, and transection SCI models, will enhance the generalizability of our findings [66]. Second, all experiments were performed in female mice, which allowed for consistent postoperative care and minimized complications associated with urinary obstruction and bladder management that are more frequent in males. However, accumulating evidence suggests that sex influences inflammatory responses, neuroplasticity, and locomotor recovery after SCI [67,68]. Future work directly comparing both sexes will determine whether postural adaptations and gait deficits quantified by the Blackbox differ between males and females. Additionally, expanding analyses to different mouse strains and sexes will provide insight into potential genetic and sex-specific differences in recovery [69–71]. While the Blackbox platform excels in detecting postural and gait abnormalities, integrating complementary techniques such as electromyography (EMG) could further elucidate neuromuscular recovery dynamics [60]. Lastly, long-term follow-up beyond 42 dpi would provide insight into whether these postural adaptations eventually normalize or persist indefinitely [72].

Transparency, rigor, and reproducibility summary

This study was not formally registered because it is a mouse study and not a clinical study. The analysis plan was also not formally pre-registered.

The current research employed a controlled and managed experimental protocol to assess long-term postural and motor deficits following SCI in animal models. Female C57BL/6J mice (n = 8 in each group) were randomly distributed into sham or SCI groups. A sample size of 6 subjects per group was estimated based on an expected effect size of 2 with a standard deviation of 1 for the forepaw distance parameter and calculated to yield 80% power using a Student’s t-test with a p-value < 0.05. Therefore, we chose a sample size of 8 mice in each group which exceeds our estimated sample size and is consistent with published literature that had demonstrated replicability of behavioral and histological results with eight mice in each group [13]. All of the surgical procedures, behavioral testing, and histological analysis were performed according to the protocols approved by the IACUC at UT Southwestern Medical Center.

Blinding was employed during behavioral testing and data analysis. The researcher administering behavioral tests and quantifying image data was blind to treatment group allocations. To restrict potential confounding variables, mice were maintained under standard environmental conditions and were tested at the same times daily.

We integrated various complementary behavioral and analytic approaches to address a wide range of locomotor and postural functions. Motor behavior was quantified with the Blackbox One system integrating high-speed videography, pose estimation, and FTIR-based weight-bearing measurement. For reproducibility, tracking data were analyzed using publicly available software (Analysis, Palmreader) and custom scripts checked, which are available on GitHub (https://github.com/helenLaiLab/BlackBox_Analysis). Tracking-derived features and pose estimation quality were manually checked. We also utilized Keypoint MoSeq for detecting motor syllables linked to behavior changes induced by SCI. All of the motion data were analyzed with version-controlled code using documented settings and parameters.

Histological analyses were performed using standard immunohistochemical techniques. Validated antibodies were utilized, purchased from commercial suppliers with proper RRIDs. Confocal imaging was performed using consistent acquisition parameters in all animals. ImageJ was used to measure injury lesion area and marker expression. Statistical analyses were performed with GraphPad Prism and R software. Temporal and group differences were compared by two-way ANOVA, with the respective post hoc corrections. PCA and MANOVA were employed for the analysis of behavioral data that are defined by more than one characteristic. All p-values and effect sizes are presented. Data are given as mean ± SEM. All raw data, statistical calculations, and analysis pipelines applied in this research are provided upon request or via GitHub to enable transparency and ensure reproducibility.