青岛大学学报（自然科学版）2025，Vol.38Issue(1)：30-36,7.DOI:10.3969/j.issn.1006-1037.2025.01.05
基于联合损失和3D-3D对比学习的分子活性预测模型
Molecular Property Prediction Model Based on Joint Loss and 3D-3D Contrastive Learning
摘要
Abstract
The current molecular property prediction methods have insufficient generaliza-tion ability for small datasets,lack sufficient utilization of molecular spatial geometric structure information,and the existing molecular property prediction methods based on contrastive learning lack high-dimensional spatial interconnections in the fine-tuning process.A self supervised learning model based on graph contrastive learning was pro-posed,which simultaneously used 3DGCN and SchNet based on 3D molecular graphs.In downstream tasks,a joint loss function consisting of contrastive learning loss and label prediction loss was introduced to optimize the feature space of the model and strengthen the connection between the high-dimensional space of the model.The experimental results show that the model achieves the best performance on the ESOL,FreeSolv,and QM7 datasets,compared with the second best model,the performance is improved by 10.95%,4.60%,and 29.26%,respectively,and the model has interpretability and can encode mo-lecular features reasonably.关键词
空间几何结构/对比学习/联合损失Key words
spatial geometric structure/contrastive learning/joint loss分类
信息技术与安全科学引用本文复制引用
张强,张贝祎,张连伟,牛东江,李臻..基于联合损失和3D-3D对比学习的分子活性预测模型[J].青岛大学学报（自然科学版）,2025,38(1):30-36,7.基金项目
国家自然科学基金(批准号:12371491)资助. （批准号:12371491）
