Tackling Molecule Assembly with Graph Diffusion

ICLR 2026 Conference Submission20256 Authors

Keywords: graph completion, diffusion models, molecule assembly, ligand design, graph neural network

Abstract: A common starting point for drug design is to find small chemical groups or "fragments" that form interactions with distinct subregions in a protein binding pocket. However, once suitable fragments are identified, assembling these fragments into a high affinity drug with desirable pharmacological properties is difficult. This "molecule assembly" task is particularly challenging because, initially, fragment positions are known only approximately, and the combinatorial space of potential connectivities is extremely large. Even if the individual fragments form favorable interactions with regions of the pocket, a poor assembly of these fragments can drastically compromise the molecule's druglikeness and hinder its binding affinity. In this paper, we present EdGr, a new graph diffusion framework tailored for the molecule assembly task. EdGr can handle both fragments and atoms, and predicted candidate edge likelihoods influence node position updates during the diffusion denoising process, allowing connectivity cues to guide spatial movements, and vice versa. EdGr substantially outperforms previous methods on the molecule assembly task and stays robust even as confidence in fragment placement decreases.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 20256