Multiscale Hypergraph Masked Autoencoder with Δ-Property Alignment for Novel Molecular Representation Learning

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Molecular representation learning struggles with functional groups and cross-scale interactions.
• Existing methods often focus on local patterns, limiting comprehensive understanding.
• Capturing chemically meaningful relationships is crucial for drug-like molecule analysis.

Purpose of the Study:

• To develop a novel self-supervised framework for drug-like molecular representation learning.
• To enhance the model's ability to capture functional group semantics and cross-scale dependencies.
• To improve the sensitivity of learned representations to structure-property relationships.

Main Methods:

• Proposed the Multiscale Hypergraph Convolutional Masked Autoencoder (MSHG-MAE) using a hypergraph model.
• Incorporated multiscale hypergraph convolutions to capture atomic, substructural, and molecular level dependencies.
• Introduced Δ-Property Alignment (Δ-PropAlign) to link embedding differences with property changes.

Main Results:

• MSHG-MAE demonstrated superior performance on molecular property regression benchmarks compared to baseline methods.
• Achieved significant reductions in RMSE for physicochemical properties like ESOL, FreeSolv, and Lipophilicity.
• Δ-PropAlign enhanced the consistency between embedding and property differences without compromising structural integrity.

Conclusions:

• MSHG-MAE provides a robust framework for chemically meaningful molecular representation learning.
• The proposed methods advance self-supervised learning in cheminformatics.
• This approach holds promise for accelerating drug discovery and development through improved molecular understanding.