眼见未必为实:理解基于甲基化的肿瘤分类中的UMAP与t-SNE
本文是一篇技术综述,阐述在基于DNA甲基化的肿瘤分类研究中,UMAP和t-SNE等降维可视化工具的本质、解释要点及报告规范。作者强调这些工具并非聚类或分类算法,其外观高度依赖样本选择、数据预处理和参数选择,并推荐了一套报告标准以提高研究的可重复性和科学严谨性。
要点
- UMAP和t-SNE是可视化工具,而非聚类或分类算法。
- 可视化图形高度依赖样本选择、预处理和参数设置。
- 推荐报告预处理全透明、可视化与定量分析配对及独立验证分组。
核心结论
- UMAP和t-SNE图不能直接作为聚类或分类证据,其外观受样本选择、预处理和参数影响大。
- 基于甲基化的肿瘤研究应遵循透明报告标准以提高可重复性和科学严谨性。
研究设计与样本
- 本文为技术综述,聚焦于甲基化肿瘤研究中UMAP和t-SNE可视化的解释与报告,而非全面比较分析方法。
- 附带的在线笔记本允许读者在公开肿瘤甲基化数据集上探索不同设置对UMAP和t-SNE图的影响。
关键数值与推荐
- 推荐报告标准包括:预处理和算法参数完全透明。
- 推荐将可视化与定量聚类或分类分析配对呈现。
- 推荐用独立的临床、组织学和分子数据验证观察到的分组。
临床怎么用
- 在诊断环境中解释甲基化可视化时,应认识到其非聚类本质及对参数和样本的依赖性。
- 病理研究采用甲基化分类时应遵循所推荐的报告标准以确保可重复性。
局限性
- 本文不提供对甲基化分析方法、分类工具或聚类算法的全面比较。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
DNA methylation profiling has become an important tool for tumor classification across multiple subspecialties of pathology. To visualize relationships among samples, many studies employ dimensionality-reduction techniques such as uniform manifold approximation and projection and t-distributed stochastic neighbor embedding. Although these techniques create visually compelling plots that summarize complex molecular data, they are visualization tools, not clustering or classification algorithms, and their appearance depends heavily on sample selection, data preprocessing, and parameter choices. This technical review outlines key considerations for the interpretation and reporting of visualization and classification methods in methylation-based tumor studies. We recommend a set of reporting standards covering full transparency of preprocessing and algorithmic parameters, pairing visualizations with quantitative clustering or classification analyses, and validating observed groupings with independent clinical, histologic, and molecular data. An accompanying online notebook allows readers to explore different settings on public tumor methylation data sets to visualize their effects on uniform manifold approximation and projection and t-distributed stochastic neighbor embedding. We focus on how to interpret methylation-based visualizations and classifier output in diagnostic settings, rather than providing a comprehensive comparison of profiling methods, classification tools, or clustering algorithms. Understanding these principles will help ensure the reproducibility and scientific rigor of pathology research incorporating methylation-based tumor classification.
引用
- Emert B. When Seeing Is Not Believing: Understanding Uniform Manifold Approximation and Projection (UMAP) and t-Distributed Stochastic Neighbor Embedding (t-SNE) in Methylation-Based Tumor Classification. Mod Pathol. 2026. PMID: 42285440. doi:10.1016/j.modpat.2026.101023
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。