.png){kind=icon}
.png){kind=icon}
Return
Motif-driven molecular graph representation learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.126484.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Graph Neural Networks (GNNs) have emerged as powerful tools for molecular graph analysis. Subgraphbased GNNs focus on learning high-level local patterns beyond simple node interactions to improve GNN expressiveness. However, current subgraph-based methods lack a unified scheme for incorporating molecular motifs that can be applied consistently across various GNN frameworks. To address this, we propose UniMotif, a universal molecular motif integration approach that enhances GNNs' expressive power. Specifically, we decouple a motif into functional encoding and learnable structural encoding, where functional encoding serves as a unique identifier for each motif and structural encoding provides local structural context for nodes within the same motif. We further analyze the effects of seven motif extraction techniques on model performance and provide an in-depth evaluation. Experimental results demonstrate the effectiveness of UniMotif in improving GNN expressive power as well as its compatibility with various GNN architectures. Code is available at https://github.com/GraphMoLab/Uni-Motif.
Keywords:
Graph neural network
Expressive power
Decoupled motif representation
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W