Deep Generative Models and Single Cell Data

Deep Generative Models and Single Cell Data

Deep generative models are a series of unsupervised methods to learn any kind of data distribution and generate new dataset based on the learned distribution. These methods are developed in computer vision community and can be used to generate new creative images. The well known approaches are Variational Autoencoders (VAE) and Generative Adversarial Networks (GAN). The success of these methods has sparkled the interests of the researchers from other domains like bioinformatics, coinciding with the recently development of single-cell technology. In biological domain, single cell technology basically means two things: single cell RNAseq and mass cytometry; regardless of the intention and logics behind the experiment setup, both technology delivers multiple-sample-multiple-dimension datasets, which are the natural match to these generative models, at least to the outward appearances.

I set up such a project to critically investigate and examine the application of those generative models to the single cell datasets. There are multiple focuses: philosophy behind the data analysis and data interpretation, the architecture of the methods, the consistency of the methods (newer ones vs classical ones).

Project duration: December 2018 - January 2020

Current progress:

  1. Project A: Single-Cell Data Analysis Using MMD Variational Autoencoder
  2. Project B: ongoing.
Chao (Cico) Zhang
Thinker, Mindfulness Meditator, Mathematician, PhD Candidate

With mindfulness and philosophy, think about the meaning of being and doing.


MMD-VAE is a better option for single cell data analysis than Vanilla VAE, because MMD-VAE retains more information in its latent …