基于稀疏概率图模型的多癌症多组学生物网络构建方法研究

The reconstruction of cancer biological networks is important for investigating the mechanisms underlying tumor development and progression. With the increasing availability of large-scale cancer omics data, numerous computational methods have been proposed to reconstruct cancer biological networks. However, these traditional approaches cannot accurately reconstruct the cancer biological networks and explore the similarities and differences between different cancers due to the following reasons: (1) different types of omics data follow different types of probability distributions; (2) the collected omics data usually have high dimensionality and low sample size; (3) a considered type of cancer can often be divided into distinct subtypes which are similar with each other but have important differences; (4) there are some common and specific network structure across different subtypes of different cancer types. In this study, we will comprehensively analyze the characteristics of cancer omics data, and develop new cancer network reconstruction methods based on sparse graphical models. We mainly focus on the following contents: (1) considering the fact that multidimensional omics data contain both continuous and binary variables, we will develop new network reconstruction methods which can model mixed multivariate data to reveal the conditional dependencies among different types of omics variables; (2) to handle the challenges with high dimensionality, we will propose new methods to simultaneously incorporate multiple types of prior knowledge into network inference processes; (3) we will develop joint network reconstruction models to explore the common and specific subnetworks across different subtypes of a considered cancer type; (4) considering the fact that different types of cancers share certain common patterns, new network reconstruction methods which can explore the similarities among different subtypes of different cancer types will be investigated. The results of this study would help us to understand how cancer networks affect tumor development and progression.

构建癌症生物网络对理解肿瘤的发生发展机制至关重要。癌症组学数据具有混合型分布与小样本高维度的特点，同类型癌症存在不同的亚型，而不同类型癌症之间存在一定共有模式，因此，传统方法很难准确地构建癌症生物网络并挖掘不同癌症之间共有与特有的遗传特征。本项目在深入分析癌症组学数据具体特征的基础上，以概率图模型和结构化稀疏学习为理论基础，寻找多癌症多组学生物网络构建的新方法。重点研究以下内容：（1）针对组学数据混合型分布的特点，研究能整合混合型多组学数据的网络构建方法；（2）针对组学数据小样本高维度的特点，研究能整合利用多种先验知识的网络构建方法；（3）针对同一类型癌症存在不同亚型的特点，研究能挖掘同一类型癌症不同亚型之间共有与特有结构的网络构建方法；（4）针对不同类型癌症之间存在共有模式的特点，研究能挖掘不同类型癌症不同亚型之间相似性的网络构建方法。本项目研究有助于理解生物网络在肿瘤发生发展中的作用。

肿瘤异质性是恶性肿瘤的关键特征，与肿瘤的生长、侵袭、转移、耐药和预后密切相关。肿瘤异质性可分为肿瘤间异质性和肿瘤内异质性。瘤间异质性主要是指不同癌种或同一癌种不同亚型之间的差异，而瘤内异质性主要是指同一肿瘤组织内不同细胞之间的差异。揭示肿瘤异质性的分子机制，有助于建立符合癌症患者病理特征的个体化诊疗策略，是癌症精准诊疗面临的关键挑战。. 为探索肿瘤间异质性，基于概率图模型和结构化稀疏学习，我们提出了一系列针对bulk组学数据的单个基因网络构建方法、差异网络分析方法和多个基因网络联合构建方法，为从基因之间直接关系在不同癌种或同一癌种不同亚型之间扰动的视角解析肿瘤间异质性提供了数学方法和应用工具。为探索肿瘤内异质性，基于集成学习、概率图模型和多目标优化等机器学习方法，我们提出了一系列针对单细胞转录组和空间转录组数据的缺失值填补方法、差异表达分析方法、差异网络分析方法和细胞类型反卷积方法，为从肿瘤组织中不同类型细胞之间差异的视角解析肿瘤内异质性提供了数学方法和应用工具。. 在本项目的支持下，我们公开发布了4个基因网络构建和扰动分析软件包，6个单细胞转录组和空间转录组数据分析软件包。项目负责人以第一作者或通讯作者发表学术论文18篇，其中Bioinformatics期刊8篇，Briefings in Bioinformatics期刊2篇，IEEE Transactions汇刊7篇，Pattern Recognition 1篇。培养博士生8人，硕士生23人，其中4人次获得研究生国家奖学金，5人次获湖北省工业与应用数学学会、湖北省生物信息学学会优秀论文奖，5人次获华中师范大学优秀博士学位论文培育计划项目和研究生教育创新资助项目的资助。. 本项目的研究可为探究肿瘤内和肿瘤间的异质性提供数学模型、计算方法和软件工具，有助于理解肿瘤发生发展、复发转移、耐药的分子机制，对癌症的诊断、治疗和预后有重要意义。