Enhancing Molecular Representations Via Graph Transformation Layers.

Area of Science:

• Computational chemistry
• Machine learning for chemistry

Background:

• Molecular representation learning is crucial for tasks like property prediction and generation.
• Graph neural networks (GNNs) are effective but often lack flexibility in learning multi-granularity molecular information.
• Existing models can be overly complex and rigid for diverse molecular learning tasks.

Purpose of the Study:

• To introduce a flexible and simple graph transformation layer, LineEvo, for enhancing molecular representation learning in GNNs.
• To enable GNNs to learn molecular information from multiple perspectives and granularities.
• To improve the performance and expressiveness of GNNs in molecular tasks.

Main Methods:

• Proposed LineEvo, a plug-and-use graph transformation layer for GNNs.
• Utilized a line graph transformation strategy to convert fine-grained molecular graphs into coarse-grained ones.
• Treated molecular edges as nodes, generating new connections, atom features, and positions.
• Stacked LineEvo layers to facilitate multilevel information learning from atom-level to coarser granularities.

Main Results:

• LineEvo layers improved traditional GNN performance on molecular property prediction benchmarks by an average of 7%.
• Demonstrated that LineEvo layers enhance the expressive power of GNNs beyond the Weisfeiler-Lehman graph isomorphism test.
• Showcased the flexibility and simplicity of LineEvo as a module for GNNs.

Conclusions:

• LineEvo offers a flexible and effective approach to enhance molecular representation learning.
• The proposed layer enables GNNs to capture multilevel molecular information, leading to improved performance.
• LineEvo contributes to developing more powerful and versatile GNN models for molecular science.