基于元学习及Lp-Lq正则编码的有监督多组学分析模型及其应用

多组学数据的整合分析为揭示疾病的分子机理提供了全新思路。而生物数据中高噪音、类别严重失衡、高维小样本等特点给现代机器学习框架带来了严重挑战。本课题以有监督多组学分析中样本选择(即样本及组学权重分配)及特征选择技术为研究对象，主要采用元学习和Lp(0<=p<=1)-Lq(1<q<=2)(Lp-Lq)正则编码两种理论，研究可基于数据进行自适应调节的样本及特征选择算法。重点研究如下内容：1)设计能充分挖掘多组学数据同异质性的有监督多组学分析框架，随后研究样本及组学权重与显式加权函数的映射形式。2)推导出Lp-Lq的广义阈值收缩算子和拟合项的梯度算子，随后提出Lp-Lq的梯度迭代算法并研究其收敛性。3)构建基于元学习及Lp-Lq正则编码的有监督多组学分析模型，实现针对不同任务对自身权重函数及模型参数分布进行自适应调节。4)发现类风湿病证诊断的生物标记物，助力建立病证结合诊断类风湿新方案。

The integrated analysis of multi-omics data provides new ideas to reveal the molecular mechanisms of diseases. However, the characteristics of high noise, severe category imbalance, and high-dimensional low-sample-size in biological data pose severe challenges to modern machine learning frameworks. In this study, we focus on sample selection (i.e., sample and omics weights assignment) and feature selection techniques in supervised multi-omics analysis, mainly using two theories of meta-learning and Lp(0<=p<=1)-Lq(1<q<=2) (Lp-Lq) regular coding, to investigate sample and feature selection algorithms that can be adaptively adjusted based on data. The key research is as follows: 1) Designing a supervised multi-omics analysis framework that can fully exploit the homogeneity and heterogeneity of multi-omics data, and subsequently studying the mapping form of sample and view weights to explicit weighting functions. 2) Derive the generalized threshold contraction operator and the gradient operator of the fitted term of Lp-Lq, and subsequently propose the gradient iteration algorithm of Lp-Lq and study its convergence. 3) Construct a supervised multi-omics analysis model based on meta-learning and Lp-Lq regular coding, and realize adaptive adjustment of its own weight function and model parameter distribution for different tasks. 4) Discover biomarkers for the diagnosis of rheumatoid disease and help establish a new scheme for rheumatoid disease diagnosis combined with disease and syndrome.