基于优化片段搜索和残基接触预测的全新蛋白质结构从头预测算法设计

In order to perform their physiological roles, macromolecules like proteins have to fold into their unique native conformations, which renders the importance of the structural studies of proteins. However, the experimental protein structural determinations, albeit well developed, are lagging far behind the derivation of amino acid sequences of proteins. Protein structure prediction, the computational technique that utilizes the amino acid sequences to predict the tertiary structures, can effectively fill the gap between sequencing and structural determination. Unlike the relatively mature protein structure prediction techniques including the homologous modeling and threading, the ab initio prediction methods predict the tertiary structures from the amino acid sequences purely based on first principles and are thus independent of the presence of homologous templates in the structural database. Unfortunately, none of the available ab initio prediction algorithms can predict the protein tertiary structures reliably. In this project, we plan to improve the accuracy and reliability of the present ab initio methods. On one hand, we will optimize the searching of fragments with low homology, which may greatly improve the efficiency of the subsequent fragment assembly that predict the protein structures by randomly assembling the identified fragments. On the other hand, we will further improve the present algorithm for predicting the contacts between amino acid residues, by effectively combining the respective information derived from the structure and sequence databases. The predicted residue contacts can further facilitate the protein structure prediction by reducing the sampling space. Combining the above two improvements, we will develop a novel algorithm for the ab initio protein structure prediction.

蛋白质需要折叠到其天然态构象行使生理功能，因此蛋白质的结构研究非常重要。虽然实验测定蛋白质结构的方法发展很快，但是仍远远落后于氨基酸序列测定的速度。蛋白质结构预测通过理论计算，根据氨基酸序列预测三级结构，因此能有效地填补结构和序列测定间的鸿沟。在蛋白质结构预测方法中，不同于较为成熟的同源建模法和穿线法，从头预测法完全根据物理化学规律进行预测，因此不依赖于结构数据库中是否存在同源模板。但是，目前没有一种从头预测法能可靠地预测蛋白质结构。本项目中，我们计划提高从头预测法的准确度和可靠性。一方面，我们优化远同源片段的搜索，进一步提高使用片段组装法通过拼接这些片段模板来预测蛋白质结构的效率。另一方面，我们通过有效结合得自于结构和序列数据库的信息，优化氨基酸残基间接触的预测。预测所得的残基接触信息可以缩减采样空间，从而进一步辅助结构预测。结合以上改进，我们计划发展一种全新的蛋白质结构从头预测算法。

蛋白质需要折叠到其天然态构象行使生理功能，因此蛋白质的结构研究非常重要。虽然实验测定蛋白质结构的方法发展很快，但是仍远远落后于氨基酸序列测定的速度。蛋白质结构预测通过理论计算，根据氨基酸序列预测三级结构，因此能有效地填补结构和序列测定间的鸿沟。在蛋白质结构预测方法中，不同于较为成熟的同源建模法和穿线法，从头预测法完全根据物理化学规律进行预测，因此不依赖于结构数据库中是否存在同源模板。但是，目前少有从头预测法能可靠地预测蛋白质结构。本项目中，我们开发了多种算法从多个角度提升从头预测方法的准确率和可靠性。特别是DeepFragLib、AmoebaContact+GDFold和GANProDist等三种算法或流程，不仅在算法的设计思路上具有高度的创新性，而且在性能上也至少达到了与领域内主流算法的持平的效果。