Forked Diffusion for Conditional Graph Generation

Abstract

We introduce a novel score-based diffusion framework that incorporates forking for conditional generation. In this framework, a single parent diffusion process is associated with a primary variable (e.g., structure), while multiple child diffusion processes are employed, each dedicated to a dependent variable (e.g., property). The parent process guides the co-evolution of its child processes towards segregated representation spaces. This approach allows our models to manage conditional information flow effectively, uncover intricate interactions and dependencies, and ultimately unlock new generative capabilities. Our experimental results demonstrate the significant superiority of our method over contemporary baselines in the context of conditional graph generation, highlighting the potential of forking diffusion for enhancing conditional generation tasks and inverse molecular design tasks.