MichiGAN: sampling from disentangled representations of single-cell data using generative adversarial networks.

Area of Science:

• Computational Biology
• Genomics
• Artificial Intelligence

Background:

• Deep generative models like variational autoencoders (VAEs) and generative adversarial networks (GANs) excel at image generation and manipulation.
• These models have potential applications in analyzing complex biological data, such as single-cell gene expression.

Purpose of the Study:

• To systematically evaluate the strengths and weaknesses of VAEs and GANs for single-cell gene expression data analysis.
• To develop an improved generative model, MichiGAN, that integrates VAE and GAN capabilities for enhanced data representation and generation.
• To explore the utility of MichiGAN in understanding cellular identity and predicting responses to drug treatments.

Main Methods:

• Comparative analysis of VAEs and GANs on single-cell RNA sequencing (scRNA-seq) datasets.
• Development and implementation of MichiGAN, a novel neural network architecture.
• Learning disentangled representations from three large-scale scRNA-seq datasets.
• Utilizing MichiGAN for sampling from learned representations and manipulating cellular identity features.

Main Results:

• MichiGAN effectively combines the complementary strengths of VAEs and GANs.
• The model successfully learns disentangled representations from complex scRNA-seq data.
• MichiGAN enables manipulation of distinct cellular identity aspects.
• The model demonstrates capability in predicting single-cell gene expression responses to drug treatments.

Conclusions:

• MichiGAN offers a powerful new approach for analyzing and generating single-cell gene expression data.
• The model facilitates deeper understanding of cellular heterogeneity and drug effects.
• This work highlights the potential of integrated deep generative models in advancing single-cell genomics research.