Data for evaluation of unsupervised learning algorithms for the classification of behavior

Experiments were carried out in adult C57BL/6J (Charles River) male and female mice. Mice were maintained under standard housing conditions with a 12-hour light cycle and with ad libitum access to food and water. All animal experiments had received approval from the local ethical board, Stockholms Norra Djurförsöksetiska Nämnd, and were performed in accordance with the European Communities Council Directive 2010/63. The open field test was conducted in a 40cm x 40cm arena (Ugo Basil®), with recordings made from a top-down view using a Basler acA1920-155um camera. Each mouse was recorded for 10 minutes, and a total of 54 mice were used for analysis.  For body part tracking we used DeepLabCut (version 2.3.3) [Mathis et al, 2018, Nath et al, 2019]. Specifically, we labeled 5 frames taken from 42 videos/animals (210 frames in total, then 95% (200 frames) was used for training) with 11 body part markers (keypoints): nose, head-center, neck, ear left, ear right, body-center, body-center left, body-center right, hip left, hip right and tailbase. We used a ResNet-50 neural network with default parameters for 200,000 number of training iterations. For the final model the test error was: 3.16 pixels, train: 2.83 pixels (image size was 650 by 600). This network was then used to analyze videos from similar experimental settings.

For VAME analysis we used default settings for  z_dim parameter and set time window to 15 frames corresponding to 500 ms of behavior. For B-SOiD analysis we used default values for minimum cluster size that ranged between 0.5% and 1%, that provided the highest score in random forest accuracy > 0.95. For Keypoint MoSeq we used kappa = 1e6, which also yielded median motif duration of 520 ms. For BFA we applied a default pipeline. To evaluate the performance of the unsupervised classification, we have performed manual labeling of 3000 frames from 3 mice (1000 frames each). Videos were annotated with the following behavioral classification labels: turn right, turn left, walk, stand and sniff, unsupported rear, walk and sniff, supported rear, groom, look up and down, pause. Three experts agreed on the manually-labeled behavioral classification in each frame.

本实验采用成年C57BL/6J（Charles River）品系的雌雄小鼠作为实验对象。小鼠饲养于标准环境中，采用12小时光照循环，且可自由进食饮水。所有动物实验均获得当地伦理委员会——斯德哥尔摩北区动物实验伦理委员会（Stockholms Norra Djurförsöksetiska Nämnd）的批准，并严格遵循《欧洲共同体理事会指令2010/63》执行。

旷场实验在40cm×40cm的实验箱（Ugo Basil®）中进行，采用Basler acA1920-155um相机从俯视视角录制视频。每只小鼠录制10分钟，共计54只小鼠用于后续分析。

身体部位追踪采用DeepLabCut（版本2.3.3）[Mathis等，2018；Nath等，2019]完成。具体而言，研究人员从42段小鼠视频中选取5帧（共计210帧），其中95%（200帧）用于模型训练，共标注11个身体部位标记点（关键点）：鼻尖、头部中心、颈部、左耳、右耳、躯干中心、左侧躯干中心、右侧躯干中心、左髋、右髋以及尾基。本研究采用参数默认的ResNet-50神经网络，训练迭代次数设置为200000次。最终模型的测试误差为3.16像素，训练误差为2.83像素（图像分辨率为650×600）。该训练完成的模型随后被用于分析相同实验条件下的视频数据。

针对VAME分析，本研究采用z_dim参数的默认设置，并将时间窗口设置为15帧，对应500ms的行为时长。针对B-SOiD分析，本研究将最小聚类簇大小的默认值设置为0.5%~1%，该参数下随机森林分类准确率最高，超过0.95。针对Keypoint MoSeq，本研究设置kappa=1e6，该参数下得到的行为基序中位时长为520ms。针对BFA分析，本研究采用默认分析流程。

为评估无监督分类的性能，研究人员对来自3只小鼠的3000帧视频帧（每只小鼠1000帧）进行人工标注。视频的行为分类标签包括：右转、左转、行走、站立嗅探、无支撑后肢站立、行走嗅探、有支撑后肢站立、理毛、上下张望、暂停。三位专家对每帧视频的人工标注行为分类结果达成一致共识。