cellstruct: Metrics scores to quantify the biological preservation between two embeddings

Abstract

Single-cell transcriptomics (scRNA-seq) is extensively applied in uncovering biological heterogeneity. There are different dimensionality reduction techniques, but it is unclear which method works best in preserving biological information when creating a two-dimensional embedding. Therefore, we implemented cellstruct, which calculates three metrics scores to quantify the global or local biological similarity between a two-dimensional and its corresponding higher-dimensional PCA embeddings at either single-cell or cluster level. These scores pinpoint cell populations with low biological information preservation, in addition to visualizing the cell-cell or cluster-cluster relationships in the PCA embedding. Two study cases illustrate the usefulness of cellstruct in exploratory data analysis.