Unsupervised Representation Learning of C. elegans Poses and Behavior Sequences From Microscope Video Recordings

Caenorhabditis elegans (C. elegans) is an important model system for studying molecular mechanisms in disease and aging. The nematode can be imaged in highly parallel phenotypic screens resulting in large volumes of video data of the moving worm. However converting the rich, pixel-encoded phenotypical information into meaningful, quantitative description of behavior is a challenging task. There is a range of methods for quantification of the simple body shape of C. elegans and the features of its motion. These methods however are often multi-step and fail in the case of highly coiled and self-overlapping worms. Motivated by the recent development of self-supervised deep learning methods in computer vision and natural language processing, we propose an unbiased, label-free approach to quantify worm pose and motion from video data directly. We represent worm posture and behavior as embedding vectors and visualize them in a unified embeddings space. We observe that the vector embeddings capture meaningful features describing worm shape and motion, such as the degree of body bend or the speed of movement. Importantly, using pixel values directly as input, our method captures coiled worm behaviors which are inaccessible to methods based on keypoint tracking or skeletonization. While our work focuses on C. elegans, the ability to quantify behavior directly from video data opens possibilities to study organisms without rigid skeletons whose behavior is difficult to quantify using keypoint-based approaches.