MotifRole-Diff boosts molecular generation with role-aware corruption

Molecular graph generation with masked discrete diffusion traditionally uses a uniform corruption schedule, treating all molecular components equally. But different parts of a molecule (like atoms, bonds, and functional groups) vary in difficulty to reconstruct and impact on the final structure. Researchers from UCF have developed MotifRole-Diff, which assigns different masking rates to each token role based on empirical denoising difficulty and perturbation impact. This role-aware strategy is formulated as a risk-optimal allocation of a fixed masking budget, proven to minimize role-weighted residual risk.

Under matched architecture, training budget, and sampling compute, MotifRole-Diff significantly outperforms uniform schedules. On the QM9 dataset, validity improved from 0.905 to 0.944, while Fréchet ChemNet Distance (FCD) dropped from 1.701 to 1.609. On the larger MOSES dataset, validity increased from 0.920 to 0.938, and FCD improved from 2.125 to 1.850. Role-wise diagnostics show better reconstruction across all molecular graph token categories.

The method's key advantage is that it improves performance without changing the underlying model architecture, clean sequence space, or lossless decoder. This means existing models can be swapped to role-aware corruption with minimal engineering overhead, potentially accelerating drug discovery and materials science applications where high-validity molecular generation is critical.