Developing a Generative Model Utilizing Self-attention Networks: Application to Materials/Drug Discovery.

Area of Science:

• Artificial Intelligence
• Computational Chemistry
• Drug Discovery

Background:

• Generative models are crucial for molecular design.
• Existing models may require complex parameter tuning.
• Developing versatile and accurate generative models is an ongoing challenge.

Purpose of the Study:

• To introduce a new generative model, the Variational Autoencoding Transformer (VAT).
• To apply the VAT model to the task of molecular generation.
• To demonstrate the model's accuracy and versatility without heuristic settings.

Main Methods:

• Combining Variational Autoencoder (VAE) networks with Transformer architectures.
• Training the VAT model on molecular datasets.
• Implementing fine-tuning and latent space mixing strategies for directed generation.

Main Results:

• The VAT model accurately reproduces input molecules.
• The VAT model generates novel molecules with high fidelity from a prior.
• The model demonstrates optimal performance without heuristic settings, applicable to diverse datasets.

Conclusions:

• The Variational Autoencoding Transformer is a powerful and adaptable tool for molecular generation.
• The demonstrated strategies offer practical pathways for materials and drug discovery.
• The VAT model's ease of use and high accuracy make it a valuable asset for researchers.