Partial Mixture-of-Experts Similarity Variational Autoencoder for Clustering on Single Cell Data

Mixture-of-Experts Similarity Variational Autoencoder (MoE-Sim-VAE) is a novel generative clustering model which can cluster high-dimensional samples well and generalize to multi-modal distributions. However, high-dimensional feature from biological measurement is often incomplete. The common solution is to fill '0's to those missing elements in high-dimensional feature, which may lead to a decrease in accuracy. In this paper, we propose a data optimization strategy which called 'partial VAE' to overcome the issue caused by missing value using 'maxpooling' operation for partial inference. The improved version of MoE-Sim-VAE is called Partial Mixture-of-Experts Similarity Variational Autoencoder (Partial MoE-Sim-VAE). We evaluate the performance of clustering on public datasets including mouse organ-cell and simulated dataset with different proportions of '0's. The experiments demonstrate that Partial MoE-Sim-VAE outperforms MoE-Sim-VAE.