偏最小二乘判别分析(PLS-DA)筛选诊断性细胞因子

研究思路

自闭症的早期诊断标志物这篇推文简单介绍了这类研究的基本思路。

研究思路

统计分析

原文An Exploratory Examination of Neonatal Cytokines and Chemokines as Predictors of Autism Risk: The Early Markers for Autism Study中的统计方法如下

Partial least squares discriminant analysis (PLS-DA) was
performed to examine whether different combinations of
multiple cytokines could be used to differentiate between
child developmental outcomes. Initially, linear regression
analysis was performed on each transformed immune marker
individually using the covariates stated above to generate
residuals for use in the PLS-DA. Eotaxin-2, epithelial
neutrophil-activating protein 78, granulocyte macrophage
colony-stimulating factor, eotaxin-1, interferon-g (IFN-g),
IL-4, monocyte chemoattractant protein 4 (MCP-4), and IL-13
all violated assumptions of linearity in the linear regression
model and were therefore excluded from the PLS-DA. The
PLS-DA was computed using the web-based MetaboAnalyst
software in accordance with the protocol by Xia and Wishart
(24). Analysis was performed using leave-one-out cross-
validation and prediction accuracy performance measure for
determining the number of latent variables. The permutation
statistic was performed using prediction accuracy during
training with 2000 permutations.

采用偏最小二乘判别分析(PLS-DA)检验是否可以使用多种细胞因子的不同组合来区分儿童发育结果。最初，使用上述协变量对每个转化后的免疫标记分别进行线性回归分析，以生成残差用于PLS-DA。Eotaxin-2、上皮中性粒细胞活化蛋白78、粒细胞巨噬细胞集落刺激因子、eotaxin-1、干扰素-g (IFN-g)、IL-4、单核细胞趋化蛋白4 (MCP-4)、IL-13均违反线性回归模型的线性假设，被排除在PLS-DA之外。PLS-DA是由Xia和Wishart(24)根据协议使用基于web的MetaboAnalyst软件计算出来的。采用无遗漏交叉验证和预测精度性能指标进行分析，以确定潜在变量的数量。排列统计采用2000个排列的训练预测精度进行。（机译）

偏最小二乘判别分析(PLS-DA)

偏最小二乘判别分析(PLS-DA)是一种用于判别分析的多变量统计分析方法。判别分析是一种根据观察或测量到的若干变量值，来判断研究对象如何分类的常用统计分析方法。其原理是对不同处理样本（如观测样本、对照样本）的特性分别进行训练，产生训练集，并检验训练集的可信度。
偏最小二乘回归（Partial least squares regression）与主成分回归相关，但不是寻找响应变量和自变量之间最大方差超平面，而是通过投影分别将预测变量和观测变量投影到一个新空间，来寻找一个线性回归模型。因为数据X和Y都会投影到新空间，PLS系列的方法都被称为双线性因子模型（bilinear fator models）。当Y是分类数据时称为偏最小二乘判别分析（Partial least squares Discriminant Analysis， PLS-DA）。
我的理解：建立一个线性回归模型来预测分类。

R语言如何进行PLS-DA

ropls: PCA, PLS(-DA) and OPLS(-DA) for multivariate analysis and feature selection of omics data

使用R包ropls进行PLS-DA