Macromolecular Potential Energy Decoder Based on Graphical Model

了解三级结构对于理解和预测蛋白质功能非常重要。然而，如何从氨基酸序列预测蛋白质的三级结构是一个悬而未决的问题。在这个项目中，切片链最大和（SCMS）算法已被提出。该方法将蛋白质分子的势函数表示为因子图，这是一种图形模型。根据分子在三维空间中的切片，将因子图转换为线性结构图。基于转换后的图，最大和搜索与逐节点局部MCMC采样相结合，通过离散变量近似连续变量。实验结果表明，SCMS比传统的MCMC方法更有效。结果表明，SCMS的改进版本（即SCMS2.0）优于由拟牛顿法增强的MCMC方法。

Knowing tertiary structure is important to understand and predict protein function. However, it is an open question how to predict the tertiary structure of proteins from a sequence of amino acids. In this project, Slice Chain Max-Sum (SCMS) algorithm has been proposed. This method represents the potential function of a protein molecule as a factor graph, which is a kind of a graphical model. The factor graph is converted into a linearly structured one according to a slicing of the molecule in 3D space. Based on the converted graph, max-sum search is performed in combination with node-wise local MCMC sampling that approximates continuous variables by discrete ones. Experimental results show that SCMS is more efficient than conventional MCMC method. It is also shown that improved version of SCMS (i.e. SCMS2.0) outperforms MCMC method that is reinforced by the quasi-Newton method.