图值神经网络体系结构及在蛋白质预测中的应用

Architecture of Graph-valued Neural Network and its Application in the Protein Folding Prediction

提出了一种新的蛋白质结构预测模型——图值神经网络。该模型以半边图理论为依据,用可结合半边图模型表示原子间和原子团间相互的结合作用,从而将对蛋白质的正确折叠起关键作用的大分子相互作用因素和环境融合进图值神经网络预测模型。此外,针对全原子计算量大的缺点,在图值神经网络中引入基团,从而在一定程度上减小了计算量。通过对模拟蛋白质折叠过程中能量变化的分析,实验表明这种融合了原子间以及原子团间作用力的模型是完全可行的。

Graph-Valued Neural Network is based on the theory of half-link graph, it can not only be used to represent the force between the atoms, but also the force between the atomic groups. Thus, the factors which play a key role in protein folding can be taken into Graph-Valued Neural Network. In addition, given the massive amount of calculation, we also bring ＂Group＂ into Graph-Valued Neural Network to reduce the amount of calculation to some extend. After analyzing changes of energy in the process of protein folding simulation, we come to a conclusion that the prediction model which combined with the force between the atoms and the force between the atomic groups is feasible.