GLDM: hit molecule generation with constrained graph latent diffusion model

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Discovering novel molecules with specific biological activity is crucial for drug development.
• Generative AI, particularly Diffusion Models (DM), shows promise for de novo molecular design.
• Existing methods face challenges in efficiently generating molecules with targeted biological properties.

Purpose of the Study:

• To propose a Graph Latent Diffusion Model (GLDM) for efficient and targeted molecular generation.
• To leverage autoencoders and latent space diffusion for improved molecular design.
• To generate molecules with desired biological activity based on gene expression profiles.

Main Methods:

• Developed an autoencoder to create low-dimensional latent representations of molecular data.
• Trained a Diffusion Model (DM) in the latent space for molecule generation.
• Focused diffusion processes in the latent space to enhance efficiency and avoid complex reconstruction.

Main Results:

• GLDM demonstrated outstanding performance on molecular generation benchmarks.
• Generated molecules exhibited optimal chemical properties.
• The model successfully produced molecules with the potential to induce desired biological activity.

Conclusions:

• GLDM offers an efficient approach for generating novel molecules with targeted biological activities.
• Latent space manipulation in diffusion models improves training efficiency for molecular design.
• This method advances computer-aided drug discovery by enabling directed generation of bioactive compounds.