计算机科学
对偶(语法数字)
先验概率
人工智能
适应(眼睛)
机器学习
模式识别(心理学)
事先信息
算法
理论计算机科学
透视图(图形)
作者
Gaokai Wang,Junyu Lin,Shucheng Li,Mei Feng,Yazhou Ren
标识
DOI:10.1109/tmm.2026.3726260
摘要
Molecular property prediction serves as a critical task in computational chemistry and drug discovery, which often employs multi-view learning to leverage large-scale molecular data for encoder pre-training. However, existing methods tend to primarily focus on the internal structure of molecular graphs and use complex structures to enhance the local feature extraction ability of the encoder, while neglecting the influence of individual nodes on external structure information. It is of vital importance to balance the prior requirements of the external structure with the increasing computational costs brought about by the complexity of the model. To this end, we propose Mol-LRDA, a parameter-efficient learning approach with dual structural attention priors, driven by a new strategy of low-rank adaptation for multi-view molecular property prediction. Specifically, we extract latent chemical structure and topological features from SMILES representations and 2D molecular graphs respectively, and construct an adaptive dual structural attention mechanism to enhance the perception of molecular nodes. Moreover, we propose constraint low-rank decomposition and parameter freezing strategy, which effectively reduces the number of trainable parameters while maintaining model performance in supervised learning settings. Experiments on molecular property prediction benchmarks demonstrate our method achieves competitive performance, while reducing computational overhead.
科研通智能强力驱动
Strongly Powered by AbleSci AI