基于有向学习策略的整合多组学数据的基因组预测统计模型研究

基因组预测(genomic prediction, GP)联合利用全基因组所有遗传变异，通过构建准确的统计模型来预测生物体表型。它在人类复杂疾病的早期诊断与精准治疗、植物分子辅助育种方面具有重要科学意义和广阔市场前景。当前，仅基于遗传变异的GP模型在刻画位点间的互作信息方面显示出瓶颈劣势；如何有效整合遗传信息传递经过的表观组、转录组、代谢组等中间组学信息，并从生物系统层面反映互作信息被普遍认为是突破瓶颈的关键。本项目拟创新性地提出一种包含多层统计模型的有向学习仿生策略：我们将遗传变异或DNA甲基化视为工具变量，拟在中间组学数据的监督下依次学习到能反映互作信息的中间遗传特征，并以此作为中间纽带进行表型预测。另外，我们拟自洽引入反向传播技术以迭代循环方式进一步提高预测准确性。最后，我们拟将其应用于人类癌症预后预测和杂交水稻全基因组选择两个领域中，期望能为人类精准医疗和植物分子育种提供有力支撑。

Genomic prediction jointly employs genome-wide genetic variants to construct accurate and valid statistical models for predicting phenotypes. It has great scientific significance and broad market prospect on human precision medicine and marker-assisted plant breeding. With the rapid development of advanced biological experiments, intermediate omics including epigenomics, transcriptome and metabolome have been intensively studied. The GP model only based on genetic variants has met a bottleneck on its defect for inadequate characterization of gene interaction information. Then how to integrate intermediate omics to reflect gene interaction information is regarded as the key for breaking through the bottleneck. In this project, we propose an innovative directed learning strategy including multiple layers of statistical models. We will take genetic variants as instrumental variables, and iteratively learn multiple layers of genetic features supervised by observed intermediate omics, and then we will employ these genetic features to predict phenotypes. Additionally, we will introduce back-propagation into the above directed learning strategy to further improve model predictability. Finally, we will apply it on cancer prognosis prediction in human and genome selection in hybrid rice. We hope that it will provide helpful supports to human precision medicine and marker-assisted plant breeding.