DMT-Dens:面向生物学数据的密度保持流形可视化
文章背景与核心概要
在单细胞测序以及各类高维生物学数据的探索中,低维嵌入(Low-dimensional embeddings)是分析细胞状态异质性的标准方法。然而,传统的降维方法虽然在保持局部邻域结构方面表现出色,却往往会扭曲观测样本的表观采样密度。这种畸变改变了致密区域与稀疏区域之间的视觉对比度,从而给罕见细胞群、过渡态细胞群或连续细胞状态群体的解释带来了困难。
为了解决这一痛点,本文介绍了 DMT-Dens——一种基于潜在标记(latent-token)Transformer 编码器的参数化流形可视化方法。该方法将基于排序的流形对齐(rank-based manifold alignment)与硬对聚合(hard-pair aggregation)相结合,并通过在 \(k\) 近邻对数半径估计上优化基于皮尔逊相关性(Pearson correlation)的损失函数,从而在复杂生物数据集上实现了稳健的密度保持,同时维持了极具竞争力的标签可分性。
摘要 (Summary)
低维嵌入是探索单细胞及其他高维生物学数据中细胞状态异质性的标准方法。尽管许多方法能够很好地保持局部邻域,但它们可能会扭曲处理后观测值的表观采样密度,从而改变致密区域与稀疏区域之间的视觉对比度,并使罕见、过渡期或连续细胞状态群体的解释变得复杂。
Low-dimensional embeddings are a standard approach for exploring cell-state heterogeneity in single-cell and high-dimensional biological data. While traditional methods preserve local neighborhoods well, they often distort the apparent sampling density of observations. This distortion alters the visual contrast between dense and sparse regions, making it difficult to interpret rare, transitional, or continuous cell-state populations.
为了应对这一问题,DMT-Dens 引入了一种构建于潜在标记 Transformer 编码器之上的参数化流形可视化方法。该模型将基于排序的流形对齐与硬对聚合相结合。为了保持密度,它优化了一种基于输入端与二维嵌入空间中 \(k\) 近邻对数半径估计之间皮尔逊相关性的损失函数。基准评估表明,该方法在生物学数据集上展现出了强大的密度保持能力,同时保留了具有竞争力的标签可分性。
To address this, DMT-Dens introduces a parametric manifold-visualization method built on a latent-token Transformer encoder. By integrating rank-based manifold alignment with hard-pair aggregation and optimizing a Pearson correlation-based loss over \(k\)-nearest-neighbor log-radius estimates, DMT-Dens achieves robust density preservation on complex biological datasets while maintaining competitive label separability.
元数据与出版详情 (Metadata & Publication Details)
- arXiv ID: arXiv:2608.17571 [q-bio.QM]
- 学科分类: 定量方法 (
q-bio.QM), 人工智能 (cs.AI) - 提交时间: 2026年8月18日
- 作者: Ruizhe Wang, Yixuan Dong, Bolin Yang, Bingo Wing-Kuen Ling, Fuji Yang, Zelin Zang
- 文档信息: 22页,5张图表,2个表格,包含补充材料。
- arXiv ID: arXiv:2608.17571 [q-bio.QM]
- Subject Categories: Quantitative Methods (
q-bio.QM), Artificial Intelligence (cs.AI)- Submitted On: August 18, 2026
- Authors: Ruizhe Wang, Yixuan Dong, Bolin Yang, Bingo Wing-Kuen Ling, Fuji Yang, Zelin Zang
- Document Info: 22 pages, 5 figures, 2 tables, includes supplementary material.
摘要详情 (Abstract)
动机: 低维嵌入被广泛用于探索单细胞和其他高维生物学数据中的细胞状态异质性。尽管许多方法保持了局部邻域,但它们可能会扭曲处理过的观测值的表观采样密度,从而改变致密区和稀疏区之间的视觉对比度,并使罕见、过渡或连续细胞状态群体的解释变得复杂。
Motivation: Low-dimensional embeddings are widely used to explore cell-state heterogeneity in single-cell and other high-dimensional biological data. Although many methods preserve local neighborhoods, they may distort the apparent sampling density of processed observations, altering the visual contrast between dense and sparse regions and complicating the interpretation of rare, transitional, or continuous cell-state populations.
结果: 我们提出了 DMT-Dens,这是一种构建在潜在标记 Transformer 编码器上的参数化流形可视化方法。该模型将基于排序的流形对齐与硬对聚合结合起来。为了保持密度,它优化了一种基于处理后输入和二维嵌入空间中 k 近邻对数半径估计之间的皮尔逊相关性的损失。基准评估证明了其强大的密度保持能力(特别是在生物数据集上),同时保留了具有竞争力的标签可分性。
Results: We present DMT-Dens, a parametric manifold-visualization method built on a latent-token Transformer encoder. The model integrates rank-based manifold alignment with hard-pair aggregation. To preserve density, it optimizes a loss based on the Pearson correlation between k-nearest-neighbor log-radius estimates in the processed input and two-dimensional embedding spaces. Benchmark evaluations demonstrate strong density preservation, particularly on biological datasets, while retaining competitive label separability.
资源与链接 (Resources & Links)
- 源码与数据脚本: GitHub 仓库 (Ruizhe-wang/DMT-Dens)
- 全文访问:
- 查看 PDF
- HTML 版本 (实验性)
- TeX 源码
- 许可协议: 知识共享署名 4.0 国际 (CC BY 4.0)

- Source Code & Data Scripts: GitHub Repository (Ruizhe-wang/DMT-Dens)
- Full-Text Access:
- View PDF
- HTML Version (Experimental)
- TeX Source
- License: Creative Commons Attribution 4.0 International (CC BY 4.0)
引用与参考文献 (Citations & References)
- DOI: 10.48550/arXiv.2608.17571
- External Indices:
- NASA ADS
- Google Scholar
- Semantic Scholar