(E-182) Behavioral movement classification of optogenetically-evoked motor output in awake, Thy1-ChR2 mice using B-SOiD

Introduction:: Stroke is the leading cause of adult disability in the US. Though many stroke survivors show some degree of spontaneous recovery, specific therapeutic targets to improve recovery are lacking due to an incomplete understanding of underlying mechanisms involved in systems-level brain repair. Motor mapping in animal models allows for determining macroscopic organization of cortical motor circuits and complex neuromechanics of motor control. In conjunction with DeepLabCut, a marker-less pose estimation toolbox driven by deep learning algorithms, optogenetic motor mapping provides a powerful tool for examining evolving relationships between motor circuit reintegration and motor recovery following focal ischemia. Using this technique, we can quantify motor output following optogenetic stimulation and demonstrate regional specific contributions to coordinated movements and, ultimately, behavior. One current limitation, however, is the inability to automate this behavioral classification. Here, we used B-SOiD, an unsupervised behavioral identification algorithm, to identify behaviorally relevant patterns of movement in response to optogenetic stimulation of the left cortex in healthy, awake mice. Applying this technique will provide an automated method to evaluate the region-level contributions to coordinated motor output and to characterize the behaviors encoded within the cortex.

Materials and Methods:: Optogenetic stimulation was performed across 378 cortical sites of the left hemisphere in nine, healthy, Thy1-ChR2. Concurrently, limb movements in response to stimulation were captured via four CMOS cameras at 200 Hz. DeepLabCut was used to extract pose estimations of several anatomical features of each of the mouse’s fore- and hindlimbs. These estimations were then triangulated and co-registered into a shared 3D space. The B-SOiD algorithm, provided by the Yttri Lab at Carnegie Mellon University, was used to train a classifying model using 3 pose estimation files from different sessions of the same mouse to label DLC pose estimations. Feature extraction was performed with a training input fraction set at 1.00 (default). Group labels for each pose estimation were clustered to identify unique behavioral patterns. To identify clusters, the range for minimum cluster size was set to 0.19% to 1.00%. This range is equivalent to 5 seconds for the smallest cluster, with reference to overall session length. To evaluate B-SOiD's performance, a ground truth was manually created, where known behaviors of the mouse were assigned to specific video frames and the labeled DLC data was compared to verify whether those same video frames contained the same clusters of groups.

Results, Conclusions, and Discussions:: The output of B-SOiD includes clusters of behaviorally relevant movement groups of pose estimations, showing a specific type of behavior. These groups were graphically represented by an HBDSCAN graph. HDBSCAN is a hierarchical clustering algorithm that segments pose estimation patterns based upon the densities of the input distribution. The HDBSCAN graph showed how distinct groups were similar by assigning them a shared color scheme. Video snippets from the training video were generated where each snippet was indicative of corresponding group, showing how B-SOiD used similar pose estimation data to categorize specific poses of the forelimbs. The classifying model, after being applied to an unlabeled dataset, output an excel file where each pose estimation was labeled a group number. Specific clusters of the labeled groups represented distinct behaviors. A synchronized video of the representative forelimb movements and the corresponding clustered groups was generated from B-SOiD. These video clips were segments of the session video that were known to contain coordinated mouse movements representative of a particular behavior. Similar movements were observed among cluster-wide movies made for behaviorally relevant epochs in the testing data. Thus, B-SOiD demonstrated the automation of behavior classification from DLC pose estimations. Clustering output from the algorithms was matched qualitatively to observed patterns of coordinated movement manually scored. B-SOiD categorized similar pose estimations to groups of distinct poses of the mouse. Using a trained model, the algorithm found clusters of the groups that were indicative of behaviors. Results of the algorithm show that this method for behavioral clustering can be applied to DLC pose estimation data. The automation of behavior classification from coordinated motor movements allows us to map behavior of the mouse to regions of the brain. These behavioral maps can be generated following focal ischemia, and a comparison can be drawn between the control and stroke-induced mice to further understand the relationship of motor recovery and motor circuit reintegration following a stroke. Future steps include training another model by inputting more pose estimation data to improve generalizability and avoid overfitting, with further refinement of hyper parameters, such as minimum cluster size.

Acknowledgements (Optional): :

References (Optional): : Hsu, A. I., & Yttri, E. A. (2021). B-SOiD, an open-source unsupervised algorithm for identification and fast prediction of behaviors. Nature communications, 12(1), 5188. https://doi.org/10.1038/s41467-021-25420-x