Multiresolution graph transformers and wavelet positional encoding for learning long-range and hierarchical...

Area of Science:

• Computational Chemistry
• Machine Learning
• Materials Science

Background:

• Current graph learning models struggle with large molecules due to their inability to capture hierarchical atomic interactions crucial for molecular properties.
• Macromolecular properties are intrinsically linked to complex, multi-scale structural information.

Purpose of the Study:

• To introduce the Multiresolution Graph Transformers (MGT), a novel graph transformer architecture designed for learning representations of large molecules at multiple scales.
• To develop a new positional encoding method, Wavelet Positional Encoding (WavePE), for improved spectral and spatial domain localization.

Main Methods:

• The proposed Multiresolution Graph Transformers (MGT) architecture processes molecular data at various resolutions, learning representations for individual atoms and their groupings.
• Wavelet Positional Encoding (WavePE) is integrated to enhance the model's understanding of molecular structure in both spectral and spatial domains.
• The model was evaluated on diverse datasets including polymers, peptides, protein-ligand complexes, and drug-like molecules.

Main Results:

• MGT achieved competitive results across multiple macromolecule and drug-like molecule datasets.
• The model demonstrated superior performance compared to state-of-the-art methods, achieving chemical accuracy in predicting key molecular properties like HOMO, LUMO, and their gap for polymers.
• Visualizations confirmed MGT's ability to capture long-range and hierarchical structures in macromolecules.

Conclusions:

• Multiresolution Graph Transformers (MGT) offer a powerful new approach for learning representations of large molecules, addressing limitations of existing graph learning methods.
• The integration of Wavelet Positional Encoding (WavePE) further enhances the model's structural understanding.
• This methodology shows significant promise for advancing computational chemistry and materials science applications.