目前有很多技术可以预测蛋白质的三级结构，但如何从众多预测模型中挑选出最接近天然结构的模型，是当前蛋白质三级结构预测的瓶颈。在蛋白质模型评估中，二级结构和溶剂可及面积等指标经常被用来评估模型的好坏。但是，二级结构只能粗略描述蛋白质局部构象，而溶剂可及面积不能区分深度和浅度包埋内部残基。本研究拟引入基于蛋白质主链二面角、带侧链信息的半球接触数和基于置信度的经验势函数等结构特征来构建新的蛋白质结构模型评估算法。通过计算结构模型的二面角，局部构象，半球接触数等结构特征与基于序列的预测值之间的一致性，并加入经验势函数作为输入特征，利用深度学习算法进行残基水平上（局部）和蛋白水平上（全局）的模型评估。在此基础上，通过结合其它评估方法建立一个评估结构的meta方法。拟参加CASP竞赛以检验本方法，并将本方法应用到识别构象型B细胞表位的研究中。

Many techniques for protein tertiary structure prediction have been proposed. It is the bottleneck of protein tertiary structure prediction to pick out the model, which is closest to the native structure, from predicted models. In the protein structure quality assessment, the measures are usually based on secondary structure and solvent accessible surface. However, the discrete states by secondary structure is only a coarse description of local structure, and solvent accessible surface can’t discriminate deeply buried residues from those shallowly buried. Here, we propose to design new features based on predicted protein main-chain torsional angles, half sphere exposure and confidence score-based potential energy to assess model. By measuring the consistency of torsional angles, local conformations and Half Sphere Exposure (HSE) between predicted and those derived from structural models and combination of potential energy function as inputs, we will develop new methods using deep neural network techniques to assess protein model on residue-level(local) and protein-level(global). By combining with other state-of-the-art methods, we will set up a meta-server, and will participate in Critical Assessment of protein Structure Prediction (CASP). Furthermore, the methods will be applied to prediction of conformational B-cell epitopes.