GRACE: Graph autoencoder based single-cell clustering through ensemble similarity learning

Abstract

Recent advances in single-cell sequencing techniques have enabled gene expression profiling of individual cells in tissue samples so that it can accelerate biomedical research to develop novel therapeutic methods and effective drugs for complex disease. The typical first step in the downstream analysis pipeline is classifying cell types through accurate single-cell clustering algorithms. Here, we describe a novel single-cell clustering algorithm, called GRACE (GRaph Autoencoder based single-cell Clustering through Ensemble similarity larning), that can yield highly consistent groups of cells. We construct the cell-to-cell similarity network through the ensemble similarity learning framework, and employ a low-dimensional vector representation for each cell through a graph autoencoder. Through performance assessments using real-world single-cell sequencing datasets, we show that the proposed method can yield accurate single-cell clustering results by achieving higher assessment metric scores.

Fig 1. Graphical overview of the proposed single-cell clustering algorithm.

GRACE includes three main steps to derive the accurate single cell clustering. First, the proposed method constructs the ensemble cell-to-cell similarity network, where it can effective represent the similarities between cells based on the multiple similarity measurements through different feature genes. Second, a graph autoencoder derives a low-dimensional vector representation for each node (i.e., cell). Finally, GRACE predicts the optimal number of clusters and yields the accurate single-cell clustering based on the low-dimensional vector representation for each node.

Fig 2. Purity scores for each clustering algorithm.

Note that we performed 10 trials and visualize scattering points for each trial and the bar plot represents the averaged purity scores for all trials.

Fig 3. Jaccard index scores for each single-cell clustering algorithm.

Note that scattering points represent the Jaccard index scores for 10 trials, and the bar plot represents the averaged Jaccard index score for all trials.

Fig 4. Adjusted rand index for each algorithm.

Note that scattering points represent the ARI scores for 10 trials, and the bar plot represents the averaged ARI score for all trials.

Fig 5. Normalized mutual information for each clustering algorithm.

Note that scattering points represent the NMI scores for 10 trials, and the bar plot represents the averaged NMI score for all trials.

Fig 6. Comparison of the true and predicted number of clusters for each algorithm.

Fig 7. Low dimensional visualization of clustering results.

Low-dimensional coordinates are derived through t-SNE and predicted clusters are annotated through different colors.

Fig 8. Computation time of each algorithm.

All experiments were performed on Intel i5 processor with 12 cores, 48GB system memory, and NVIDIA GTX 1060 GPU. Note That the base clock frequency of the CPU is 4.10 GHz.