The development of the generative adversarial supporting vector machine for molecular property generation

Area of Science:

• Computational Chemistry
• Artificial Intelligence
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) are powerful for image generation but suffer from large hyper-parameter spaces, complicating training.
• Predicting molecular properties from scratch is crucial for drug discovery and materials science.

Purpose of the Study:

• To develop a novel generative model by integrating Supporting Vector Machines (SVM) into the GAN architecture.
• To reduce the hyper-parameter space and improve the accessibility of training generative models for molecular property prediction.

Main Methods:

• A hybrid generative model combining GANs with SVM was designed.
• The model was trained using molecular structures, energies, and dipole moments of the formic acid dimer (FAD) system as feature vectors.
• Generated data was validated against ab initio calculations.

Main Results:

• The proposed model successfully generated new molecular feature vectors from scratch.
• Generated data demonstrated strong agreement with ab initio values.
• The model avoided the mode collapse problem, ensuring unique data generation.

Conclusions:

• The SVM-enhanced GAN offers a more efficient and accessible approach to generative modeling for molecular properties.
• The model's extensibility allows incorporation of diverse molecular properties, suggesting broad applicability.
• This novel method provides a robust tool for de novo molecular design and property prediction.