单细胞多组学数据的多尺度双对齐深度聚类方法

打开文本图片集

关键词：多尺度融合；多组学聚类；对比学习；对齐；协同策略

中图分类号：TP399 文献标志码：A 文章编号：1001-3695（2026）01-012-0102-08

doi：10.19734/j. issn.1001-3695.2025.06.0204

Single-cell multi-omics data multi-scale dual-alignment deep clustering

Jin Zhicheng 1，2 ， Zhang ，Li Yuru1，2，Su Chen1，²，Tian Ye1，²，Wang Yin³，Feng Xi4 （1.SchoolofompuerSeedEnei，ulUeitfoluilnagi5oha;.Guagibt fEmbededTecholodIntellgentSystuiuangi54o，Chia；3ScholofompuerSciencndEngninCtal SouthUnierstofuednsit tions，Chongqing ，China）

Abstract：Single-cellclusteringaalysisplaysacrucialroleindissectingcellarheterogeneity.Existingmethodsfortegating multi-omicsdatafaceseveralchalenges，including insuficientmodelingoflocalandglobalinter-omicsrelationships，featureredundancynoise interference，anddificultiesinconstructingaconsensusclustering space.Toaddressthesesues，his studyproposedanovel single-cellmulti-omicsclusteringmethodcaledmulti-scaledual-alignmentdeepclustering（scMDDC）. scMDDC capturedboth local and globalrelationshipsbetweencells througha multi-scale fusion strategy，which ffectively extractedcomplexintercellularinteractionpaterns.Furthermore，iteducedinter-omicsredundantinformationandhighlights modality-specificsignalsviacontrastivealignmentandcellalignment.Themethodtheniterativelytreateddiferentomics modalitiesasanchorstogudetheclusteringofother modalitiesusing amulti-omicsco-clusteringstrategy，therebyachieving inter-modalitycomplementarityandenhancing consensus.Extensiveexperimentsonmultiplereal-worlddatasetsshowthat scMDDCsignificantlyoutperforms eightbenchmarkmodelsonvariousclusteringevaluationmetrics，including clusteringacuracyand the adjusted Rand index.Thisdemonstrates thatscMDDC notonly provides a new and efective approach forsinglecell multi-omics analysis but also substantially improves the precision of cell type identification.

Key words：multi-scale fusion; multi-omics clustering； contrastive learning；alignment； co-clustering strategy

0 引言

单细胞RNA 测序（single-cell RNA sequencing，scRNA-seq）技术已成为解析组织细胞异质性的关键工具[1～3]。（剩余20573字）