Molecular Property Prediction and Molecular Design Using a Supervised Grammar Variational Autoencoder.

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Machine learning (ML) algorithms are commonly used for predicting molecular properties and designing novel molecules.
• These applications are typically handled by separate models, leading to fragmented approaches in computational chemistry.

Purpose of the Study:

• To develop a unified machine learning framework that combines molecular property prediction and de novo molecule generation.
• To enhance generative models by incorporating property information directly into the training process.

Main Methods:

• Modification of the grammar variational autoencoder (GVAE) model to create a supervised GVAE (SGVAE).
• Incorporation of property information into the SGVAE training procedure.
• Utilizing the QM9 dataset for training and evaluating the model's performance.

Main Results:

• The SGVAE successfully predicts molecular properties and generates novel molecules with desired characteristics.
• The model achieved high accuracy in predicting dipole moment and atomization energies, comparable to chemical accuracy.
• Performance surpassed existing ML models that exclusively predict properties using SMILES representations.

Conclusions:

• The proposed SGVAE approach offers a viable method for integrating property prediction into generative molecular models.
• This unified approach enhances the generation of novel molecules with predictable and desirable properties.
• The SGVAE framework advances the field of generative chemistry by enabling accurate property prediction alongside molecule design.