Statistical test of structured continuous trees based on discordance matrix
Statistical test of structured continuous trees based on discordance matrix
复制标题
基于不一致矩阵的结构化连续树统计检验
DOI：
10.1093/bioinformatics/btz425
复制
发表时间：
2019
期刊：
影响因子：
5.8
通讯作者：
Wan Lin
中科院分区：
文献类型：
--
作者：
Bai Xiangqi;Ma Liang;Wan Lin
动机
细胞命运决定是一个连续的过程，其中一种细胞类型按照分层路径分化为其他细胞类型。单细胞技术的进步提供了揭示细胞发展连续体的机会，形成结构化的连续树。计算算法通常基于对隐藏结构的先验假设，先前已经被提出作为沿着沿着细胞分化过程恢复伪轨迹的手段。然而，目前尚缺乏一个统计框架来评估嵌入在高维基因表达谱中的内在结构。另一方面，单细胞数据的基础是遗传噪声和细胞间变异，这对测试甚至是基本结构（如线性与分叉）都构成了巨大挑战。
结果
在这项研究中，我们提出了一个自适应的统计框架，称为结构化连续树（SCTree），测试的内在结构的高维单细胞数据集。基于度量几何和随机矩阵理论的工具进行SCTree测试。简单地说，通过扩展Gromov-Farris变换，利用半圆定律，我们将连续树结构检测问题转化为信号矩阵检测问题。我们表明，SCTree测试是最强大的信噪比超过一个中等的值时。我们还证明了SCTree能够用模拟数据集和真实的scRNA-seq数据集鲁棒地检测线性、单个和多个分支事件。总的来说，SCTree测试提供了对单细胞数据隐藏结构显著性的统一统计评估。
可用性
SCTree软件可从https://github.com/XQBai/SCTree-test获得。
补充资料
补充信息可在Bioinformatics获得。
MOTIVATION
Cell fate determination is a continuous process in which one cell type diversifies to other cell types following a hierarchical path. Advancements in single-cell technologies provide the opportunity to reveal the continuum of cell progression which forms a structured continuous tree. Computational algorithms, which are usually based on a priori assumptions on the hidden structures, have previously been proposed as a means of recovering pseudo-trajectory along cell differentiation process. However, there still lack of statistical framework on the assessments of intrinsic structure embedded in high-dimensional gene expression profile. Inherit noise and cell-to-cell variation underlie the single-cell data, on the other hand, pose grand challenges to testing even basic structures, such as linear versus bifurcation.
RESULTS
In this study, we propose an adaptive statistical framework, termed structured continuous tree (SCTree), to test the intrinsic structure of a high-dimensional single-cell dataset. SCTree test is conducted based on the tools derived from metric geometry and random matrix theory. In brief, by extending the Gromov-Farris transform and utilizing semicircular law, we formulate the continuous tree structure testing problem into a signal matrix detection problem. We show that the SCTree test is most powerful when the signal-to-noise ratio exceeds a moderate value. We also demonstrate that SCTree is able to robustly detect linear, single and multiple branching events with simulated datasets and real scRNA-seq datasets. Overall, the SCTree test provides a unified statistical assessment of the significance of the hidden structure of single-cell data.
AVAILABILITY
SCTree software is available at https://github.com/XQBai/SCTree-test.
SUPPLEMENTARY INFORMATION
Supplementary Information is available at Bioinformatics.