Versatile Molecular Editing via Multimodal and Group-optimized Generative Learning

Generating molecules with specific constituents and structures that exhibit desired properties is a crucial yet challenging task in the computer-aided design of functional molecules. This challenge arises from the discrete nature of the vast design space of molecules, which is subject to additional physical constraints such as symmetries. Exploration and optimization within this constrained discrete space pose difficulties for most machine learning methods. We approach these challenges from a novel multimodal perspective, and propose a multimodal generative method, MolEdit, to explicitly model both the discrete atomic constituents and their continuous positions in 3D space. MolEdit is based on a group-optimized score matching algorithm which preserves the symmetry of atomic positions on top of a joint model which generates the molecular constituents. Therefore, MolEdit simultaneously resolves the discrete and continuous optimization problems encountered in molecular generation. As a compelling feature, MolEdit can flexibly generate precisely edited molecular graphs subjected to various prompts which specify conditional information about molecular constituents as well as substructures. Overall, MolEdit not only advances the frontier of prompt-based molecular design, but also provides a template for transacting powerful modern generative methods for domain-specific scientific data in a multimodal way