scMoClust: A Residual-Enhanced Deep Learning Framework for Clustering Single-Cell Multi-Omics Data

Authors:

Abstract
Recent advances in single-cell multi-omics enable joint profiling of multiple molecular modalities within the same cell, offering unprecedented opportunities to dissect cellular heterogeneity. However, accurate clustering of paired multi-omics data remains challenging due to modality-specific noise, sparsity, and technical variations, which obscure the shared biological structure underlying cell identities. Existing methods often fail to preserve fine-grained cell–cell relationships while extracting robust cross-modal consensus, leading to unstable or biased clustering.We propose scMoClust, a unified framework that integrates graph-based neighborhood expansion, deep cumulative residual representation learning, and structured low-rank consensus factorization in an end-to-end optimization. Modality-specific graphs are first constructed and expanded to enhance local manifold structure. A deep cumulative residual network then progressively propagates multi-scale features while mitigating information loss across layers. To integrate complementary modalities, scMoClust learns a structured low-rank consensus matrix that captures robust cross-modal relationships and suppresses modality-specific noise. This consensus representation is subsequently used for clustering.Extensive experiments on simulated and real paired datasets demonstrate that scMoClust outperforms state-of-the-art methods in accuracy, robustness, and stability, particularly under high sparsity and heterogeneous conditions. Downstream biological and clinical analyses further confirm that the identified clusters are biologically coherent and clinically relevant. Collectively, these results position scMoClust as an effective and robust solution for integrative clustering of single-cell multi-omics data.

This publication has 31 references indexed in Scilit: