Mlost Jakub, Dawli Rame, Liu Xuan, Costa Ana Rita, Pollak Dorocic Iskra
Science for Life Laboratory, Department of Biochemistry and Biophysics, Stockholm University, Stockholm, Sweden.
Patterns (N Y). 2025 Apr 22;6(5):101237. doi: 10.1016/j.patter.2025.101237. eCollection 2025 May 9.
Analyzing animal behavior is crucial for decoding brain function, modeling neurological disorders, and assessing therapeutics. Recent advances in pose-estimation tools like DeepLabCut and SLEAP have revolutionized behavioral analysis by enabling precise tracking of animal body movements. However, these tools do not automate behavioral classification. Unsupervised learning algorithms address this gap by identifying clusters of recurring behavioral motifs from pose-tracking data without requiring pre-labeled datasets, reducing observer bias and uncovering novel patterns. This study compares four recent unsupervised learning algorithms-B-SOiD, BFA, VAME, and Keypoint-MoSeq-analyzing their methodological approaches, clustering efficiency, and ability to produce meaningful behavioral classifications. By offering a qualitative and quantitative evaluation, this paper aims to aid researchers in selecting the most suitable tool for their specific research needs.
分析动物行为对于解读大脑功能、模拟神经疾病以及评估治疗方法至关重要。像DeepLabCut和SLEAP这样的姿态估计工具的最新进展，通过能够精确跟踪动物身体运动，彻底改变了行为分析。然而，这些工具并没有实现行为分类的自动化。无监督学习算法通过从姿态跟踪数据中识别重复行为模式的聚类来填补这一空白，无需预先标记的数据集，减少了观察者偏差并揭示了新的模式。本研究比较了四种最新的无监督学习算法——B-SOiD、BFA、VAME和关键点序列分析（Keypoint-MoSeq）——分析它们的方法、聚类效率以及产生有意义行为分类的能力。通过提供定性和定量评估，本文旨在帮助研究人员为其特定研究需求选择最合适的工具。
