MotifRole-Diff: Risk-Optimal Role-Aware Corruption for Masked Molecular Graph Diffusion

Short summary

MotifRole-Diff introduces a role-aware corruption process for masked molecular graph diffusion that allocates masking rates based on empirically measured denoising difficulty and graph-level perturbation impact. Under matched compute budgets, it improves validity on QM9 from 0.905 to 0.944 and on MOSES from 0.920 to 0.938 while reducing FCD scores. The results show that structurally informed corruption outperforms uniform masking schedules for serialized molecular graph diffusion.

• •Role-aware corruption allocates masking rates by denoising difficulty and perturbation impact rather than uniform schedules
• •QM9 validity improved from 0.905 to 0.944; MOSES validity improved from 0.920 to 0.938 under matched compute
• •Schedule selection formulated as risk-optimal allocation of fixed masking budget across token roles

Generated with AI, which can make mistakes.