AbstractCurrently, single-cell RNA sequencing (scRNA-seq) allows high-resolution views of individual cells, for libraries of up to (tens of) thousands of samples. In this study, we introduce the use of graph neural networks (GNN) in the unsupervised study of scRNA-seq data, namely for dimensionality reduction and clustering. Motivated by the success of non-neural graph-based techniques in bioinformatics, as well as the now common feedforward neural networks being applied to scRNA-seq measurements, we develop an architecture based on a variational graph autoencoder with graph attention layers that works directly on the connectivity of cells. With the help of three case studies, we show that our model, named CellVGAE, can be effectively used for exploratory analysis, even on challenging datasets, by extracting meaningful features from the data and providing the means to visualise and interpret different aspects of the model. Furthermore, we evaluate the dimensionality reduction and clustering performance on 9 well-annotated datasets, where we compare with leading neural and non-neural techniques. CellVGAE outperforms competing methods in all 9 scenarios. Finally, we show that CellVGAE is more interpretable than existing architectures by analysing the graph attention coefficients. The software and code to generate all the figures are available at https://github.com/davidbuterez/CellVGAE.