Scholarlink

<p>Cancer subtyping aims to discover subgroups of cancer patients exhibiting clinically significant phenotypic differences, enabling the development of subtype-specific treatments. Subtype identification is a challenging unsupervised learning task due to the absence of the identity labels. An intuitive solution is to employ pseudo labels to train a deep learning model to cognize the discriminativity among samples. However, an inevitable issue of these methods is that noise labels produced during the clustering process can propagate into subsequent cycles, which adversely affect training accuracy. Meanwhile, the varying size of the cluster set will lead to the updating inconsistency. To address these challenges, a novel deep learning model known as Memory Guided Contrastive Learning (MGCL) is proposed, in which a memory dictionary is designed after clustering to smooth and mitigate the effects of clustering noise. Subsequently, a consistency strategy is employed for the generation of a balancedly distributed dataset. Further, it computes cluster-to-instance contrastive loss between query instances and cluster representatives to obtain optimized feature space to achieve superior clustering results. Experiments on multi-omics cancer datasets show that the proposed method outperforms state-of-the-art methods, achieving superior results in 5 out of 8 datasets based on the log-rank p-value and in 6 out of 8 datasets based on the C-index. Additionally, it enriches the largest number of clinical parameters across all datasets, with an average of 3.4 parameters, demonstrating its robustness and clinical relevance. Furthermore, the findings indicate that the method is capable of identifying clinically significant subgroups across diverse cancer types. The source codes are available at <a href="https://github.com/Dereck2025/MGCL.">https://github.com/Dereck2025/MGCL.</a></p>

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

Memory guided contrastive learning for cancer subtype identification based on multi-omics data

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Chengang Liu,
  • Ying Chen,
  • Xi Wu

摘要

Cancer subtyping aims to discover subgroups of cancer patients exhibiting clinically significant phenotypic differences, enabling the development of subtype-specific treatments. Subtype identification is a challenging unsupervised learning task due to the absence of the identity labels. An intuitive solution is to employ pseudo labels to train a deep learning model to cognize the discriminativity among samples. However, an inevitable issue of these methods is that noise labels produced during the clustering process can propagate into subsequent cycles, which adversely affect training accuracy. Meanwhile, the varying size of the cluster set will lead to the updating inconsistency. To address these challenges, a novel deep learning model known as Memory Guided Contrastive Learning (MGCL) is proposed, in which a memory dictionary is designed after clustering to smooth and mitigate the effects of clustering noise. Subsequently, a consistency strategy is employed for the generation of a balancedly distributed dataset. Further, it computes cluster-to-instance contrastive loss between query instances and cluster representatives to obtain optimized feature space to achieve superior clustering results. Experiments on multi-omics cancer datasets show that the proposed method outperforms state-of-the-art methods, achieving superior results in 5 out of 8 datasets based on the log-rank p-value and in 6 out of 8 datasets based on the C-index. Additionally, it enriches the largest number of clinical parameters across all datasets, with an average of 3.4 parameters, demonstrating its robustness and clinical relevance. Furthermore, the findings indicate that the method is capable of identifying clinically significant subgroups across diverse cancer types. The source codes are available at https://github.com/Dereck2025/MGCL.