基于多时点动态贝叶斯网络的宫颈癌化疗敏感性代谢组学研究

Dynamic variations of endogenous small-molecules occur in cervical cancer (CC) patients after chemotherapy, which provide rich information on research of selecting biomarkers for chemosensitivity. Based on model for multi-time metabolomic data constructed by dynamic Bayesian network (DBN), the current study would explore the metabolomic pathways of chemosensitivity in CC and seek potential biomarkers. The main contents of the research include study of DBN model for high-dimensional data, algorithms and parameter estimation for DBN model, feature selection in DBN model and statistical inference, method of information fusion for DBN model, and validation of metabolomic pathways for the selected biomarkers. A novel algorithm of Random Forest-dynamic Bayesian network (RF-DBN) will be proposed in this research, which can automatically select potential biomarkers and metabolites even with complex interactions for high-dimensional metabolomic data with multi-time dynamic measurements. Meanwhile, RF-DBN can make full use of time-delay effects of different metabolites and present causal relationships of biomarkers in the regulatory networks. In addition, a new model of two-state DBN (TS-DBN) containing hidden variables will be offered in this study, which can effectively utilize biological prior information and analyze data through information infusion from plasma and urine. TS-DBN will much accurately predict in early stage whether the CC patients are resistant to chemotherapy and provide corresponding biological explanation.

宫颈癌患者经药物干预后，其内源小分子水平的变化为筛选化疗敏感性标志物的研究提供了丰富的信息。本项目通过构建多时点代谢组学数据的动态贝叶斯网络（DBN）模型，探索宫颈癌化疗敏感性反应的代谢通路，筛选相应的生物标志物。研究内容包括：高维数据的动态贝叶斯网络模型、模型的参数估计与算法、模型变量的筛选与统计推断、贝叶斯网络模型的信息融合方法、生物标志物代谢通路的验证。本研究预期提出一种RF-DBN算法，这种算法能够对多时点动态检测的高维代谢组学数据，实现自动筛选潜在生物标志物和具有复杂交互作用的代谢物质；并能充分利用调控网络中代谢物质的时间延迟效应，给出不同生物标志物的网络调控关系。本研究还将给出一种新的含有隐变量的双层动态贝叶斯网络（TS-DBN）模型，这种模型能够有效地利用生物学先验信息，融合血浆和尿液不同来源的生物样本数据进行分析，更准确地对患者是否具有耐药性进行早期预测并给出生物学解释。

贝叶斯网络(BN)是一种概率图形模型，它以有向无环图的形式反映一组变量之间潜在的依赖和独立关系，其在代谢组学数据分析中有着重要的作用。随机森林(RF)通过自助法重采样技术，生成分类树组成随机森林，对数据进行分类，具有在处理高维数据时不会产生过拟合现象，在分类的同时能够给出变量的重要性评分的特点。本项目给出了RF-BN算法，并对四种偏最小二乘（PLS-DA）诊断统计量、PLS-DA交叉验证方法进行了深入研究；针对卵巢癌早期诊断和佝偻病诊断的实际医学问题，利用代谢组学数据，应用给出的统计方法及相关的代谢组学统计方法，发现了卵巢癌患者中存在的潜在生物标志物，确定了与佝偻病诊断有关的重要代谢生物标志物, 对这两种疾病的发病机制和病理生理研究提供了新的依据。