基于图对比学习的单细胞多组学癌症亚型识别方法
王永旺 , 梁潇 , 刘威 , 骆嘉伟
天津大学学报(自然科学与工程技术版) ›› 2026, Vol. 59 ›› Issue (11) : 1208 -1218.
基于图对比学习的单细胞多组学癌症亚型识别方法
Graph Contrastive Learning-Based Approach for Cancer Subtype Identification Using Single-Cell Multi-Omics Data
随着单细胞测序技术的快速发展,单细胞多组学数据在癌症异质性研究中展现出巨大潜力.癌症亚型识别作为癌症异质性研究中的关键任务,对于揭示肿瘤的发展具有重要意义.传统的癌症亚型识别方法多依赖于 bulk 测序数据,虽取得一定成效,但由于其仅反映细胞群体的平均表达水平,难以揭示肿瘤微环境中细胞间的多样性与异质性.此外,现有多组学融合方法往往未能充分兼顾组学内部的特征关联与组学间的协同整合,导致融合后的特征未能充分捕获组学间的复杂关系.针对上述问题,提出了一种基于图对比学习的单细胞多组学癌症亚型识别方法(scMOGCL).该方法为不同组学分别构建了模态特定的图注意力网络,自适应地捕捉细胞之间的相互作用关系并提取潜在表示.随后,通过引入自监督对比学习策略,在每种组学中构造原始样本与扰动样本之间的对比机制,提升模型在无监督情境下的表征能力与鲁棒性.同时,引入跨模态一致性约束机制,实现各组学之间的协同训练,从而获取更具一致性的多组学综合表示.该方法在 3 个真实的单细胞多组学癌症数据集上进行了系统实验,结果表明,相比于其他 6 种基准测试方法,scMOGCL 展现出具有竞争力的性能,验证了其在癌症亚型识别与单细胞多组学整合任务中的有效性与应用前景.
With the rapid technological development of single-cell sequencing,such multi-omics data have shown great potential in cancer heterogeneity research and tumor pathogenesis. A key challenge in such research is subtype identification,the traditional methods for which mostly rely on bulk sequencing data. Although these methods have achieved a certain degree of success,they reflect only average expression intensities at the cell population level but not the diversity or heterogeneity of individual cells within the tumor microenvironment. In addition,integrating existing multi-omics often fails to fully consider the feature correlations within each omics and the collaborative integration across different omics,resulting in a combination of features that are ineffective in capturing complex relationships across omics. Considering these issues,we propose scMOGCL,a graph contrastive learning-based method for cancer subtype identification using single-cell multi-omics data. This method constructs modality-specific graph attention networks for various omics,which can adaptively capture cell-cell interactions and extract latent representations. A self-supervised,contrastive learning strategy is subsequently introduced to construct contrastive pairs of original and perturbed samples within each omics,thereby enhancing the representative ability and robustness of the model under unsupervised conditions. Meanwhile,a mechanism constraining cross-modal consistency is also incorporated to coordinate training across multi-omics and thus represent them more consistently and comprehensively. The proposed method is systematically evaluated for applicability by employing three real-world datasets. Compared with six benchmark methods,scMOGCL exhibited competitive performance,validating its effectiveness and application potential in cancer subtype identification utilizing single-cell multi-omics data.
| [1] |
|
| [2] |
|
| [3] |
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
Baul S,Ahmed K T,Filipek J, |
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
|
| [16] |
|
| [17] |
Wang B, Mezlini A M, Demir F, |
| [18] |
|
| [19] |
|
| [20] |
|
| [21] |
|
| [22] |
|
| [23] |
|
| [24] |
|
| [25] |
|
国家自然科学基金资助项目(62372165)
国家自然科学基金重点资助项目(62032007)
/
| 〈 |
|
〉 |