改进傅里叶域转换的分子性质预测方法仿真

Simulation of Molecular Property Prediction Method Based on Improved Fourier Transform

生物信息分子交换过程存在复杂的量子互斥相互作用,使得分子性质预测难以实现。以图神经网络为基础,提出一种基于分子结构图数据的分子性质预测方法。将由顶点-边结构组成的不同种类图形数据作为神经网络输入,构建图神经网络;结合分子结构图的不规则特征与复杂性,利用傅里叶变换方法在图神经网络中引入谱卷积,完成节点的傅里叶域转换,得到图卷积神经网络。通过融合分子结构图、聚合原子邻域信息、更新原子结构,实现分子性质的预测。实验结果证明了所提方法具有高预测精度,且规模数据集处理能力较强,分子预测的泛化性与迁移性的优越性显著。

Complex mutual quantum exclusion during biological information molecular exchange makes it difficult to predict the molecular properties. Based on the graph neural network, a method of predicting molecular properties based on molecular structure diagram data was proposed. Different types of graph data composed of vertex-edge structures were used as the input of the neural network, thus constructing a graph neural network. Combined with the irregularity and complexity of molecular structure, spectral convolution was introduced into the graph neural network, and then Fourier transform method was adopted to transform the Fourier domain of nodes, and thus to obtain a graph convolution neural network. Finally, the prediction for molecular properties was achieved by integrating molecular structure, aggregating atomic neighborhood information and updating atomic structure. Experimental results prove that the proposed method has high prediction accuracy, strong ability of processing large-scale data sets. Significantly, the generalization and migration of molecular prediction are improved.