Optimal marker gene selection for cell type discrimination in single cell analyses

Abstract

Single-cell technologies characterize complex cell populations across multiple data modalities at unprecedented scale and resolution. Multi-omic data for single cell gene expression, in situ hybridization, or single cell chromatin states are increasingly available across diverse tissue types. When isolating specific cell types from a sample of disassociated cells or performing in situ sequencing in collections of heterogeneous cells, one challenging task is to select a small set of informative markers that robustly enable the identification and discrimination of specific cell types or cell states as precisely as possible. Given single cell RNA-seq data and a set of cellular labels to discriminate, scGeneFit selects gene markers that jointly optimize cell label recovery using label-aware compressive classification methods. This results in a substantially more robust and less redundant set of markers than existing methods, most of which identify markers that separate each cell label from the rest. When applied to a data set given a hierarchy of cell types as labels, the markers found by our method improves the recovery of the cell type hierarchy with fewer markers than existing methods using a computationally efficient and principled optimization.

Fig. 1. scGeneFit identifies markers associated with a flat partition of cell type labels when applied to a wide range of synthetic datasets.

A Proof of concept inspired by ref. ; cells are color coded with labels. In simulated high-dimensional data, for each sample, two dimensions (x- and y-axes) are drawn from concentric circles, and the remaining dimensions are drawn from white noise. The underlying structure is not apparent from the data (A-i). Considering each dimension in isolation, marker selection fails to capture the true structure (A-ii,iii). In contrast, scGeneFit recovers the correct dimensions as markers, and is able to recapitulate the label structure (A-iv). B Discriminative markers were correctly recovered by scGeneFit for simulated samples drawn from mixtures of Gaussians corresponding to two distinct label sets with three (B-ii), and four (B-iii) labels, respectively. Each row is a single sample and each column is a single feature or gene. Only 1000 genes of 10,000 are visualized, representing all the types simulated. The yellow lines correspond to the markers selected by scGeneFit. C t-SNE visualizations of results from the functional group synthetic data (C-i–iv). ROC curves comparing the performance of one-vs-all and scGeneFit in distinguishing cell labels following dimension reduction. scGeneFit outperforms one-vs-all in most cell labels when using the same number of markers (C-v).

Fig. 2. scGeneFit applied to scRNA-seq data with input cell labels both unstructured (flat) and hierarchical.

A, B Results from single-cell expression profiles of cord blood mononuclear cells (CBMC) given a flat partition of labels. A Mean accuracy and variance of scGeneFit as a function of the number of allowed markers. B t-SNE visualization of scGeneFit with 15 marker genes distinguishing 13 distinct cell populations. C Hierarchical clustering of brain scRNA-seq data and a t-SNE plot of the cell labels at the first level of the hierarchy: interneurons, pyramidal S1 cells, pyramidal CA1 cells, oligodendrocytes, microglia, endothelial–mural cells, and astrocytes. D Hierarchical labeling of the data with respect to the 30 markers chosen by scGeneFit. scGeneFit achieves predictive accuracy comparable to the full gene set, using 40% fewer markers than one-vs-all.

Fig. 3. Example of hierarchical partition explaining the notation.

In this example, we have three classes (T₁, T₂, and T₃) at the first level of the hierarchy. At the second level of the hierarchy, T₁ is divided into three classes (T₁₁, T₁₂, and T₁₃), and T₂ is divided in two classes (T₂₁ and T₂₃).