Phenotype Anomaly Detection in Early C. elegans Embryos by Variational Auto-Encoder

Variational auto encoder (VAE) is used to detect and quantify the phenotype anomaly in the nuclear division of the early embryo of C. elegans. The latent space of VAE, on which the normal data distribution is obtained through the training, is used to characterize not only the morphological anomaly, but also the temporal anomaly of the time series data, based on the position in the latent space. The proposed method is applied to the time series of three dimensional DIC data of nuclear division process during two-cell stage of C. elegans. Wild type data is used as the normal data for the training, and then an anomaly is evaluated on an embryo, for which one of the lethal genes is silenced by RNAi. First, Morphological anomaly is quantified by the reconstruction error. Then, for the well-reconstructed data, the trajectory in the latent space corresponding to the input time series is used to characterize the time development of the division process. Anomaly score is defined based on the normal time distribution in the latent space, and the proposed method successfully obtains a list of lethal genes, which cause the temporal anomaly by the knocking down.