3D Molecular Representations Based on the Wave Transform for Convolutional Neural Networks.

Area of Science:

• Computational chemistry
• Machine learning
• Structural bioinformatics

Background:

• Convolutional neural networks (CNNs) are effective for 3D data but struggle with sparse molecular representations.
• Direct voxel-based 3D molecular structures lead to suboptimal CNN performance.

Purpose of the Study:

• To develop a novel, more effective 3D molecular representation for CNNs.
• To improve the performance of CNN-based autoencoders and classification tasks.

Main Methods:

• Proposed a new method to extend atom representations to fill nearby voxels using wave transforms.
• Applied the novel representation to CNN-based autoencoders and molecular classification tasks.
• Experimented on a dataset of 4.5 million molecules from the Zinc database.

Main Results:

• The proposed wave transform-based representation significantly outperformed traditional voxel-based and Gaussian blur methods.
• Achieved enhanced performance in CNN-based autoencoders using the new molecular representation.
• Successfully applied the representation to MACCS fingerprint prediction.

Conclusions:

• The novel wave transform-based 3D molecular representation enhances CNN performance.
• This approach offers a more robust method for analyzing 3D molecular structures computationally.
• The improved representation has practical applications in chemical classification and drug discovery.