基于广义矩量法的自适应套索惩罚线性混合模型用于高维多组学数据的预测分析

An Adaptive Lasso Penalized Linear Mixed Model with Generalized Method of Moments for Prediction Analysis on High-Dimensional Multi-Omics Data

摘要: 在现代精准医学的探索中，对疾病风险的准确预测至关重要。高维多组学数据为此类预测研究提供了前所未有的资源，但其高维性和复杂的内部关系给分析带来了重大挑战。我们提出了一种基于广义矩估计框架方法(MpLMMGMM-AL)的自适应Lasso惩罚线性混合模型，用于使用高维多组学数据预测表型。我们的方法采用自适应Lasso作为惩罚函数，利用随机效应部分的核函数捕获不同组学数据层的各种类型的预测效应，并有效地选择预测组学区域及其相应的效应。通过大量的仿真，我们证明了MpLMMGMM-AL可以同时考虑大量变量，并有效地从各自的组学层中选择具有预测能力的变量。将该方法应用于公开数据集TCGA中的乳腺癌数据，并与MpLMMGMM进行了性能比较。

Abstract: In the exploration of modern precision medicine, an accurate prediction of the disease risk is crucial. For such predictive research, high-dimensional multi-omics data provide unprecedented resources, however, their high dimensionality and intricate internal relationships pose significant analytical challenges. We propose an adaptive Lasso penalized linear mixed model under a generalized method of moments estimation framework (MpLMMGMM-AL) for predicting phenotypes using high-dimensional multi-omics data. Our approach employs adaptive Lasso as the penalty function, utilizes kernel functions in the random effects part to capture various types of predictive effects across different omics data layers, and effectively selects predictive omic regions and their corresponding effects. Through extensive simulations, we demonstrate that MpLMMGMM-AL can simultaneously consider a large number of variables and effectively choose variables with predictive power from their respective omics layers. Our method is applied on a breast cancer data from the publicly available dataset TCGA, and the performance is compared with MpLMMGMM.