Prompsy Pacôme, Saichi Mélissa, Raimundo Félix, Vallot Céline
CNRS UMR3244, Institut Curie, PSL Research University, 26 rue d'Ulm, 75005 Paris, France.
Department of Translational Research, Institut Curie, PSL Research University, 26 rue d'Ulm, 75005 Paris, France.
NAR Genom Bioinform. 2024 Dec 18;6(4):lqae174. doi: 10.1093/nargab/lqae174. eCollection 2024 Dec.
The increasing diversity of single-cell datasets require systematic cell type characterization. Clustering is a critical step in single-cell analysis, heavily influencing downstream analyses. However, current unsupervised clustering algorithms rely on biologically irrelevant parameters that require manual optimization and fail to capture hierarchical relationships between clusters. We developed IDclust, a framework that identifies clusters with significant biological features at multiple resolutions using biologically meaningful thresholds like fold change, adjusted -value and fraction of expressing cells. By iteratively processing and clustering subsets of the dataset, IDclust guarantees that all clusters found have significantly different features and stops only when no more interpretable cluster is found. It also creates a hierarchy of clusters, enabling visualization of the hierarchical relationships between different clusters. Analyzing multiple single-cell transcriptomic reference datasets, IDclust achieves superior clustering accuracy compared to state of the art algorithms. We showcase its utility by identifying previously unannotated clusters and identifying branching patterns in scATAC datasets. Using it's unsupervised nature and ability to analyze different -omics, we compare the resolution of different histone marks in multi-omic paired-tag dataset. Overall, IDclust automates single-cell exploration, facilitates cell type annotation and provides a biologically interpretable basis for clustering.
单细胞数据集日益增加的多样性需要系统的细胞类型表征。聚类是单细胞分析中的关键步骤，对下游分析有重大影响。然而，当前的无监督聚类算法依赖于生物学上不相关的参数，这些参数需要手动优化，并且无法捕捉聚类之间的层次关系。我们开发了IDclust，这是一个框架，它使用诸如倍数变化、调整后的 值和表达细胞分数等生物学上有意义的阈值，在多个分辨率下识别具有显著生物学特征的聚类。通过迭代处理和聚类数据集的子集，IDclust保证找到的所有聚类都有显著不同的特征，并且只有在找不到更多可解释的聚类时才停止。它还创建了聚类层次结构，能够可视化不同聚类之间的层次关系。通过分析多个单细胞转录组参考数据集，与现有算法相比，IDclust实现了更高的聚类准确性。我们通过识别以前未注释的聚类和识别scATAC数据集中的分支模式来展示其效用。利用其无监督性质和分析不同组学的能力，我们在多组学配对标签数据集中比较了不同组蛋白标记的分辨率。总体而言，IDclust使单细胞探索自动化，便于细胞类型注释，并为聚类提供生物学上可解释的基础。
