Scholarlink

Single-cell RNA sequencing (scRNA-seq) technique may offer expression profiles of individual cells and open a new area of biological research. Effective identification of distinct cell populations requires addressing three data-specific challenges: high-dimensional, nonlinear, and noisy. These constraints drive demand for advanced analytical frameworks optimized for single-cell omics. In this paper, on the basis of the low-rank representation model in kernel space, we propose a new clustering method, LRMKLS. The LRMKLS framework combines the multiple kernel learning technique, local structures learning, and adaptive similarity preserving with the low-rank representation model in kernel space to build a fused affinity matrix for clustering. Specifically, first, to well handle the nonlinear structure, the low-rank method is extended to the kernel space. Also, the multiple kernel learning technique is introduced to improve the learning capability of the method. Second, to enable the model to retain crucial similarity information from multiple top gene sets, an adaptive similarity preserving term is incorporated. Finally, local linear and nonlinear manifold structure is extracted from data in the original space using the distance fusion approach. Following LRMKLS decomposition, a consensus affinity matrix can be constructed, allowing for downstream analysis. Extensive validation on actual datasets demonstrates that LRMKLS outperforms several widely used clustering techniques in terms of clustering accuracy.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Low-Rank Multiple Kernel Model Based on Local Structures Learning and Adaptive Similarity Preserving for scRNA-seq Data Clustering

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Juan Wang,
  • Tianjing Qiao,
  • Zhenduo Zhang,
  • Chunhou Zheng,
  • Shasha Yuan

摘要

Single-cell RNA sequencing (scRNA-seq) technique may offer expression profiles of individual cells and open a new area of biological research. Effective identification of distinct cell populations requires addressing three data-specific challenges: high-dimensional, nonlinear, and noisy. These constraints drive demand for advanced analytical frameworks optimized for single-cell omics. In this paper, on the basis of the low-rank representation model in kernel space, we propose a new clustering method, LRMKLS. The LRMKLS framework combines the multiple kernel learning technique, local structures learning, and adaptive similarity preserving with the low-rank representation model in kernel space to build a fused affinity matrix for clustering. Specifically, first, to well handle the nonlinear structure, the low-rank method is extended to the kernel space. Also, the multiple kernel learning technique is introduced to improve the learning capability of the method. Second, to enable the model to retain crucial similarity information from multiple top gene sets, an adaptive similarity preserving term is incorporated. Finally, local linear and nonlinear manifold structure is extracted from data in the original space using the distance fusion approach. Following LRMKLS decomposition, a consensus affinity matrix can be constructed, allowing for downstream analysis. Extensive validation on actual datasets demonstrates that LRMKLS outperforms several widely used clustering techniques in terms of clustering accuracy.