ClustAssess: tools for assessing the robustness of single-cell clustering
The transition from bulk to single-cell analyses refocused the computational challenges for high-throughput sequencing data-processing. The core of single-cell pipelines is partitioning cells and assigning cell-identities; extensive consequences derive from this step; generating robust and reproducible outputs is essential. From benchmarking established single-cell pipelines, we observed that clustering results critically depend on algorithmic choices (e.g. method, parameters) and technical details (e.g. random seeds). We present ClustAssess, a suite of tools for quantifying clustering robustness both within and across methods. The tools provide fine-grained information enabling (a) the detection of optimal number of clusters, (b) identification of regions of similarity (and divergence) across methods, (c) a data driven assessment of optimal parameter ranges. The aim is to assist practitioners in evaluating the robustness of cell-identity inference based on the partitioning, and provide information for choosing robust clustering methods and parameters. We illustrate its use on three case studies: a single-cell dataset of in-vivo hematopoietic stem and progenitors (10x Genomics scRNA-seq), in-vitro endoderm differentiation (SMART-seq), and multimodal in-vivo peripheral blood (10x RNA+ATAC). The additional checks offer novel viewpoints on clustering stability, and provide a framework for consistent decision-making on preprocessing, method choice, and parameters for clustering.