基于图卷积网络的分子气味印象预测

Molecular Odor Impression Prediction Based on Graph Convolutional Networks

摘要: 自20世纪90年代以来，嗅觉技术越来越受欢迎，并以各种方式进入了商业用途。从化妆品到洗发水，以及带有香味的博物馆和主题公园，嗅觉消费产品已经陡然流行起来，消费者不仅虚心接受，甚至积极寻求嗅觉产品。然而，目前关于嗅觉的研究大多来自于气味分子的电子鼻数据和质谱数据的角度，而这些数据的获取需要耗费大量的人力和时间。因此，我们从一个新的角度出发，将气味分子的结构视为一个由节点和边组成的图，并引入图卷积网络作用于这个图结构来预测气味分子的气味印象。我们在公开的气味数据集上进行了模型训练，预测了气味分子的气味愉悦度、强度和熟悉度得分，均取得了较好的结果，其中气味愉悦度得分预测的平均绝对误差MAE = 8.532，皮尔逊相关系数为r = 0.520 (p < 0.0000001)，证实了将气味分子的结构视为图结构而获得的分子信息能够预测气味分子的气味印象。

Abstract: Since the 1990s, olfactory technology has grown in popularity and entered commercial use in a variety of ways. From cosmetics to shampoos, to scented museums and theme parks, olfactory consumer products have exploded in popularity, with consumers not only humbly accepting but actively seeking them. However, most of the current research on olfaction comes from the electronic nose data and mass spectrometry data of odor molecules, and the acquisition of these data requires a lot of manpower and time. Therefore, from a new perspective, we treat the structure of odor molecules as a graph consisting of nodes and edges, and introduce a graph convolutional network to act on this graph structure to predict the odor impression of odor molecules. We trained the model on the public odor data set, and predicted the odor pleasantness, intensity and familiarity scores of odor molecules, and achieved good results. The mean absolute error of odor pleasantness score prediction was MAE = 8.532, and Pearson’s correlation coefficient was r = 0.520 (p < 0.0000001), confirming that the Molecular information obtained from the structure of odor molecules as a graph structure can predict the odor impression of odor molecules.