Joined Diffusion Models for Nanoparticle Generation

This research proposes a novel approach to utilizing Generative AI (GenAI) in nanomedicine by treating nanoparticles as collections of individual molecule-like substructures, rather than as whole entities. By coordinating multiple pretrained models for each substructure using a copula, this method overcomes data scarcity issues commonly faced in nanochemistry. The versatility of this approach allows for applications across various fields, including chemistry and computer vision.

"Some nanoparticles exhibit unique properties that make them ideal drug candidates for treating antibiotic-resistant infections and evasive cancers. However, their unique size, structure, and composition also make them challenging to characterize and design. Generative AI (GenAI) models have led to unprecedented breakthroughs in the design and modeling of images, text, proteins, and small molecules. Unfortunately, such models require enormous training datasets, which renders them useless for nanochemistry. Instead of directly training diffusion models on nanoparticles, I propose a new approach that views a nanoparticle as a set of individual molecule-like substructures and coordinates multiple pretrained models (one for each substructure) using a copula. This compositional approach enables the use of GenAI in the data-scarce and previously-intractable regime of nanomedicine. However, the generality of my approach means that it is applicable to chemistry, computer vision, and countless other fields. I will discuss this work in progress, as well as my preceding published research."