Adversarial Training of Single-Cell RNA-seq: New Machine Learning Framework for Data Normalization & Visualization

To develop a new machine learning framework based on adversarial training to normalize and visualize single-cell RNA-seq data.

The Zou lab describes a metric to quantify residual confounding: taking as input any dataset (e.g. scRNA-seq) that has been adjusted for batch correction or alignment, the metric is based on training a classifier (the “adversary”) to try to re-identify the original batches. The intuition is that if this classifier is successful, then there is still substantial residual confounding left in the data and downstream analysis could be biased. Code for this work is openly available.