Keypoint-MoSeq: parsing behavior by linking point tracking to pose dynamics.

Area of Science:

• Ethology
• Computational Neuroscience
• Machine Learning

Background:

• Keypoint tracking quantifies animal movement from videos.
• Parsing continuous keypoint data into discrete actions is challenging due to high-frequency jitter.
• Existing clustering algorithms can misinterpret noise as action transitions.

Purpose of the Study:

• To develop a machine learning-based platform, keypoint-MoSeq, for unsupervised identification of behavioral modules ('syllables') from keypoint data.
• To distinguish between movement noise and actual behavior using a generative model.
• To enable accurate parsing of continuous pose dynamics into discrete behavioral units.

Main Methods:

• Developed keypoint-MoSeq, a generative model-based platform.
• Trained the model to differentiate keypoint noise from behavioral signals.
• Applied the platform to analyze animal movement data from various settings and species.

Main Results:

• Keypoint-MoSeq successfully identifies behavioral syllables with boundaries corresponding to natural pose dynamics discontinuities.
• The platform outperforms standard clustering methods in identifying action transitions, correlating neural activity with behavior, and classifying behaviors.
• Keypoint-MoSeq demonstrates versatility across multiple species (mice, fruit flies) and behavioral timescales.

Conclusions:

• Keypoint-MoSeq provides a robust, unsupervised method for dissecting complex behaviors into modular units from standard video recordings.
• The platform accurately distinguishes behavioral elements from noise, offering a significant advancement in ethological analysis.
• Keypoint-MoSeq makes the modular structure of animal behavior accessible for research using readily available video data.