CellMarkerPipe: cell marker identification and evaluation pipeline in single cell transcriptomes

Abstract

Assessing marker genes from all cell clusters can be time-consuming and lack systematic strategy. Streamlining this process through a unified computational platform that automates identification and benchmarking will greatly enhance efficiency and ensure a fair evaluation. We therefore developed a novel computational platform, cellMarkerPipe ( https://github.com/yao-laboratory/cellMarkerPipe ), for automated cell-type specific marker gene identification from scRNA-seq data, coupled with comprehensive evaluation schema. CellMarkerPipe adaptively wraps around a collection of commonly used and state-of-the-art tools, including Seurat, COSG, SC3, SCMarker, COMET, and scGeneFit. From rigorously testing across diverse samples, we ascertain SCMarker's overall reliable performance in single marker gene selection, with COSG showing commendable speed and comparable efficacy. Furthermore, we demonstrate the pivotal role of our approach in real-world medical datasets. This general and opensource pipeline stands as a significant advancement in streamlining cell marker gene identification and evaluation, fitting broad applications in the field of cellular biology and medical research.

Figure 1

Schematic overview of the pipeline. The cellMarkerPipe comprises three key steps: preparation, selection, and evaluation for marker genes.

Figure 2

Testing scenarios for various methods. (a) Evaluation of methods on Zeisel data with varying numbers of selected marker genes. (b) Assessment of methods on Jurkat data with fluctuations in cell type proportions. (c) Examination of methods on PBMC data with varying numbers of cells in the input dataset, and (d) varying numbers of genes post dimension reduction. (e) COSG results showcased for the plant (Arabidopsis root) data, highlighting specific expressions with the top 10 genes in each cluster.

Figure 3

Performance of COSG on single-cell data. (a) Evaluation of COSG performance on human gut single-cell data. (b) Evaluation of COSG performance on mice gut single-cell data. (c) Top 10 gene expressions per cell type identified by COSG for human gut samples. (d) Top 10 gene expressions per cell type identified by COSG for mice gut samples.

Figure 4

Running time measurement in seconds for varied input parameters. (a) Variation in running time with changes in the number of genes. (b) Variation in running time with changes in the number of cells in the input data. These computational tasks were executed on a computing node with 10G of allocated memory, running a single process on standard CPUs at the Holland Computing Center, University of Nebraska Lincoln.

Figure 5

Comparative performance in blood samples. (a) Performance of methods in unedited and modified blood samples from the same patient. (b) Top 10 marker gene expressions per cell type identified by COSG for unedited and modified blood samples. (c) Top 10 gene expressions for B cells. (d) Influence of clustering method on recall and precision.