Fig 1.
A. scRNA-seq count matrix are downloaded and preprocessed using linnorm. B. LSH based sampling is performed on the preprocessed data to obtain a subsample of features. C. A cell neighbourhood graph is constructed using copula correlation. D. A three layer graph convolution neural network is learned with adjacency matrix and node feature matrix as input. It aggregates information over neighbourhoods to update the representation of nodes. The final representation obtained is called graph embedding which is utilized for cell clustering.
Table 1.
Performance of GCN on networks created from four datasets.
First two columns of the table shows total number of edges and number of nodes of the four networks. The rest of the columns show ROC and average precision score for validation and test edges. V. ROC and V. AP refer to validation ROC and validation average precision score, whereas T. ROC and T. AP refer to the same for test set.
Fig 2.
Performance of different embedding algorithms on four datasets.
Kl divergence (KL div) is computed by rerunning embedding algorithms 50 times.
Table 2.
Comparison with state-of-the-art: Adjusted Rand Index (ARI) and Average Silhouette Width (ASW) are reported for seven competing methods on four datasets.
Fig 3.
Correlation score between two distance matrices, defined on original and reduced dimension.
Figure shows the comparisons among the competing methods based on the correlation scores (Kendall τ) obtained from four different scRNA-seq datasets.
Table 3.
Execution time in minute for eight competing methods.
Table 4.
A brief summary of the dataset used here.