结合分子图与分子指纹对比的分子性质预测  

Integrating Molecular Graphs and Fingerprints through Contrastive Learning for Enhanced Molecular Property Prediction

作　　者：慈磊 蒋林华[1] CI Lei;JIANG Linhua(School of Information Engineering,Huzhou University,Zhejiang,313000,China)

机构地区：[1]湖州师范学院信息工程学院,湖州313000

出　　处：《网络新媒体技术》2025年第2期40-49,共10页Network New Media Technology

基　　金：国家自然科学基金(编号:62175037)。

摘　　要：分子性质预测已经成为新药研发过程中十分重要的一个过程,然而还存在单一分子表示信息不完全、标签分子数量稀缺的问题。为了解决这些问题,提出一种结合分子图与分子指纹的对比学习框架。根据拓扑信息将无标签分子进行聚簇,得到这些簇中的中心分子,接着利用中心分子的2种不同表示形式与数据集中的数据进行对比学习。在对比学习过程中,根据2种不同的分子表示信息,在化学空间中拉近相似结构的分子,疏远结构上不同的分子,同时能够丰富分子的表示信息。实验结果显示,所提出的方法能够有效提高分子表示的特征信息,同时在6个公开数据集上平均性能比目前先进的模型高1.1%。Molecular property prediction has become a crucial process in the development of new drugs.However,challenges still exist,such as incomplete information from single molecular representations and the scarcity of labeled molecules.To address these issues,we propose a contrastive learning framework that integrates molecular graphs and molecular fingerprints.We first cluster unlabeled molecules based on their topological information,identifying the central molecules within each cluster.Subsequently,we use two different representations of these central molecules to perform self-supervised pretraining on the dataset.During the self-supervised training process,molecules with similar structures are drawn closer together in chemical space,while those with dissimilar structures are pushed apart,based on their topological information.Experimental results demonstrate that the proposed method effectively enhances the feature information of molecular representations and achieves an average performance improvement of 1.1%over state-of-the-art models across six public datasets.

关 键 词：分子性质预测 预训练 对比学习 分子指纹 多种表示学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]