Graph Drawing-based Dimensionality Reduction to Identify Hidden Communities in Single-Cell Sequencing Spatial Representation

Abstract

With the rapid growth of single cell sequencing technologies, finding cell communities with high accuracy has become crucial for large scale projects. Employing the current commonly used dimensionality reduction techniques such as tSNE and UMAP, it is often difficult to clearly distinguish cell communities in high dimensional space. Usually cell communities with similar origin and trajectories cluster so closely to each that their subtle but important differences do not become readily apparent. This creates a problem for clustering, as clustering is also performed on dimensionality reduction results. In order to identify such communities, scientists either perform broad clustering and then extract each cluster and perform re-clustering to identify sub-populations or they over-cluster the data and then merging the clusters with similar gene expressions. This is an incredibly cumbersome and time-consuming process. To solve this problem, we propose K-nearest-neighbor-based Network graph drawing Layout (KNetL, pronounced like nettle) for dimensionality reduction. In our method, we use force-directed graph drawing, whereby the attractive force (analogous to a spring force) and the repulsive force (analogous to an electrical force in atomic particles) between the cells are evaluated, and the cell communities are organized in a structural visualization. The coordinates of the force-compacted nodes are then extracted, and we employ dimensionality reduction methods, such as tSNE and UMAP to unpack the nodes. The final plot, a KNetL map, shows a visually-appealing and distinctive separation between cell communities. Our results show that KNetL maps bring significant resolution to visualizing and identifying otherwise hidden cell communities. All the algorithms are implemented in the iCellR package and available through the CRAN repository. Single (i) Cell R package (iCellR) provides great flexibility at every step of the analysis pipeline, including normalization, clustering, dimensionality reduction, interactive 2D and 3D visualizations, batch alignment or data integration, imputation, and interactive cell gating tools, which allow users to manually gate around the cells.