Speaker: Ali Heydari (Applied Mathematics Graduate Student)
Title: Generating Realistic Single-Cell RNA-seq Data
Short Abstract: This talk is about generating realistic synthetic single-Cell RNA-seq data.
Part i) A review of existing deep learning methods
Part ii) A preview of our novel method, which is based on Introspective Variational Autoencoders
Long Abstract: A fundamental problem present in biomedical research is the low number of single-cell observations, primarily due to experimental costs, ethical reasons, or unavailability of patient samples. The absence of sufficient data thus leads to a lack of accuracy and generalizability of many current treatment prediction models. Generating realistic single-cell RNA sequencing data (scRNA-seq) or augmenting existing datasets could lead to more robust analyses and higher reproducibility rates. To address this problem, we will first review deep learning methods for producing in silico scRNA-seq with an emphasis on a recent method that employs Generative Adversarial Networks (GANs). This new approach, called conditional single-cell GANs, has shown promising results but faces challenges in training stability and sampling diversity due to the nature of GANs. We will then discuss our novel (and not yet fully developed) method of using Introspective Variational Autoencoders for realistic data generation, which addresses the instability issues of GANs while preserving the advantages of adversarial training.