Phenotype Anomaly Detection for Biological Dynamics Data Using a Deep Generative Model

Eisuke Ito, Takaya Ueda, Ryo Takano, Yukako Tohsato, Koji Kyoda, Shuichi Onami, Ikuko Nishikawa

Abstract: Unsupervised anomaly detection is applied to biological dynamics data using a deep generative model. Phenotype anomaly detection for gene function identification has been successfully used in the reverse genetics, where the anomaly is conventionally characterized by a large number of pre-defined handcrafted features. The latest database of three-dimensional cell division data for the early development process of a model animal Caenorhabditis elegans (C. elegans) enables the present data-driven approach. A variational auto-encoder (VAE) was trained by 59 wild type (WT) data to acquire the normal individual features, and used to detect phenotypic anomalies in individuals resulting from selective RNA interference (RNAi) gene knockdown. Morphological anomalies were detected using the reconstruction error, whereas temporal anomalies were characterized by the time development trajectory in the VAE latent space. RNAi data corresponding to 97 essential genes on chromosome III in the two-cell period were studied by computational experiments, and several genes were suggested to be responsible for morphological or temporal anomalies, including genes with well-known functions in asymmetric cell division or cytoskeleton organization.

External IDs:doi:10.1007/978-3-031-15931-2_36