FEATURE SCREENING FOR CLUSTERING ANALYSIS OF COUNT DATA WITH AN APPLICATION TO SINGLE-CELL RNA-SEQUENCING

• 成果类型: Article
• 署名作者: Wang, Changhu; Chen, Zihao; Xi, Ruibin
• 署名单位: Peking University; Peking University
• 刊物名称: ANNALS OF APPLIED STATISTICS
• ISSN/ISSBN: 1932-6157; 1941-7330
• DOI: 10.1214/25-AOAS2102
• 发表日期: 2025-12
• 页码: 2738-2758
• 关键词: Clustering analyses; feature screening; homogeneity test; EM-test; single-cell RNA-sequencing; feature-selection; variable selection; mixture; homogeneity; likelihood

摘要

Motivated by single-cell RNA sequencing (scRNA-seq) applications, we consider feature screening for ultrahigh-dimensional clustering of count data. A critical problem in scRNA-seq studies is cell type identification, typically achieved by clustering high-dimensional gene expression data. However, many genes are cluster-irrelevant, and their inclusion can significantly impact clustering accuracy. Effective screening of cluster-relevant genes, known as the feature screening problem in cluster analysis, is thus critical for scRNAseq analyses. Observing that the marginal distribution of any feature is a mixture of its conditional distributions across different clusters, we propose a computationally efficient method, EM-test, to screen cluster-irrelevant features by independently evaluating the homogeneity of each feature's mixture distribution. Under general parametric settings, EM-test achieves the sure independence screening property and even selection consistency. Simulations demonstrate that EM-test accurately screens cluster-relevant features and significantly improves clustering. In an application to an immune single-cell dataset from 31 patients, EM-test outperformed other methods in detecting known cell subtypes and identified a novel monocyte subtype associated with immune responses to viral infection.