Hitting stride by degrees: Fine grained molecular generation via diffusion model

Abstract

The application of generative methods in the field of drug design, specifically in the de novo generation of molecules, has gained progress. However, most deep molecular generation models suffer from low value of novelty (a general metric of evaluating generation effect) and weak ability to balance trade-offs among multiple performances. Furthermore, they are unable to effectively generate molecules with user-specified properties and scaffolds, which is required for the actual discovery of drug-like compounds. To solve the problems, a diffusion-based molecule generation model, referred to as DIFFUMOL, is proposed. DIFFUMOL selectively adds Gaussian noise to retain conditional features as guidance information and learns to directly reverse the diffusion process as a Markov chain. The preserved guidance information will be used to guide the generation of molecules with specific property values and scaffolds. By optimizing a variational lower bound to the (conditional) likelihood, DIFFUMOL effectively trains the entire framework in an end-to-end fashion. Unlike one-step molecular generation methods such as auto-encoders and generative adversarial networks, DIFFUMOL employs stepwise noise addition and subsequent denoising to achieve finer control over the generation process. Experiments show that DIFFUMOL achieves pretty good results in fine-grained molecule generation and generates user-specified molecules.