Conditional Molecule Generation with 2D Latent Diffusion

Abstract

Diffusion-based generative deep neural networks have achieved excellent results in sound and image generation, with promising applications in various fields, including drug discovery. In addition to already existing 3D approaches, we propose a new method, MoLD (Molecular Latent Diffusion) that utilizes 2D latent diffusion to achieve both unconditional and conditional molecule generation. With both sampling methods, the model is able to create numerous novel and valid molecules not present in the training dataset. Furthermore, comparing the distribution of unconditionally and conditionally sampled molecules based on the control property indicates that the diffusion model effectively influences the molecule formation process. The effectiveness of our approach demonstrates that the conditional molecule generation process can be modularly modified by substituting or retraining only the diffusion network responsible for conditioning. This modification can be achieved without restructuring the latent space, leaving the possibility for further research open.