Abstract
Graph diffusion models, while dominant in graph generative modeling, remain relatively underexplored for graph-to-graph translation tasks like chemical reaction prediction. We show that standard permutation equivariant denoisers cause severe limitations on such tasks, a problem that we pinpoint to their inability to break symmetries present in the noisy inputs. We then propose to \emph{align} the input and target graphs in order to break the input symmetries, while retaining permutation equivariance in the non-matching portions of the graph. We choose retrosynthesis as an application domain, and show how alignment takes the performance of a discrete diffusion model from a mere 5% to a SOTA-matching 54.7% top-1 accuracy.
Type
Publication
In International Conference on Learning Representations
