MotifRole-Diff Improves Molecular Graph Generation with Role-Aware Corruption.

Masked discrete diffusion models, commonly used for generating molecular graphs, typically apply a uniform corruption schedule across all tokens in their graph-to-sequence representation. This approach implicitly assumes that all molecular components are equally difficult to reconstruct and equally important to the final molecule's structure. However, different parts of a molecule, or "token roles," vary significantly in their denoising difficulty and their influence on the decoded molecule. To address this, researchers developed MotifRole-Diff, a novel role-aware corruption process. This method intelligently allocates masking rates by considering the empirically measured denoising difficulty and the impact of perturbations at the graph level for each token role. The core idea is to optimize the allocation of a fixed masking budget across these roles to minimize risk, while maintaining the model architecture, clean sequence space, and lossless molecular-graph decoder. Empirical evaluations show that MotifRole-Diff substantially improves molecular generation performance. On the QM9 dataset, it increased validity from 0.905 to 0.944 and reduced FCD from 1.701 to 1.609. Similarly, on MOSES, validity improved from 0.920 to 0.938, and FCD dropped from 2.125 to 1.850. These results, achieved with matched computational resources, demonstrate that a structurally informed corruption strategy is more effective than uniform masking for serialized molecular graph diffusion.