pyNVR: investigating factors affecting feature selection from scRNA-seq data for lineage reconstruction

Abstract

Motivation: The emergence of single-cell RNA-sequencing has enabled analyses that leverage transitioning cell states to reconstruct pseudotemporal trajectories. Multidimensional data sparsity, zero inflation and technical variation necessitate the selection of high-quality features that feed downstream analyses. Despite the development of numerous algorithms for the unsupervised selection of biologically relevant features, their differential performance remains largely unaddressed.

Results: We implemented the neighborhood variance ratio (NVR) feature selection approach as a Python package with substantial improvements in performance. In comparing NVR with multiple unsupervised algorithms such as dpFeature, we observed striking differences in features selected. We present evidence that quantifiable dataset properties have observable and predictable effects on the performance of these algorithms.

Availability and implementation: pyNVR is freely available at https://github.com/KenLauLab/NVR.

Supplementary information: Supplementary data are available at Bioinformatics online.

Fig. 1.

Evaluation of pyNVR performance. (A) Fold difference in runtime between the Python and R implementations of NVR. (B) Gene set similarity given different datasets. (C) Gene set similarity and its relationship with cell number. (D) Gene set similarity and its relationship with closeness threshold-imposed sampling. (E) Representative p-Creode graphs generated using genes selected from closeness-thresholded samples. Heatmap overlay and gating depicts Myc and putative stem-like cell states, respectively