How to Build Regulatory Networks from Single-Cell Gene Expression Data
Nearly twenty methods have been developed to infer gene regulatory networks (GRNs) from single-cell RNA-seq data. An experimentalist seeking to analyze a new dataset faces a daunting task in selecting an appropriate inference method since there are no widely accepted ground-truth datasets for assessing algorithm accuracy and the criteria for evaluation and comparison of methods are varied. The first half of the presentation focuses on BEELINE, a systematic evaluation framework for algorithms that infer GRNs from scRNA-seq data. We find that the accuracies of the algorithms are moderate to poor. Techniques that do not require pseudotime-ordered cells are generally more accurate. Based on these results, we develop recommendations on GRN algorithms to end users.
The approaches evaluated in BEELINE rely on unsupervised or association based strategies. In the second half of the presentation, we propose a novel graph convolutional neural network (GCN) based autoencoder to infer new regulatory edges in a supervised manner from a known GRN and scRNA-seq data. By systematically evaluating our method on different mouse scRNA-seq datasets, we show that our GCN-based autoencoders have considerably better cross validation performance than existing supervised learning algorithms for GRN inference. Finally, we demonstrate an analysis that reveals that our GCN-based autoencoder can predict cell-type specific regulatory edges, even when trained using non-cell specific regulatory networks.