DiffER$^2$: Diffusion Ensembles for Retrosynthesis Prediction with SMILES Adapted Particle Guidance

ICLR 2026 Conference Submission13716 Authors

Keywords: Discrete Diffusion, Diffusion Guidance Methods, Categorical Diffusion, Ensemble Methods, Molecule Generation, Retrosynthesis Prediction

TL;DR: We develop discrete diffusion ensembles with SMILES adapted particle guidance to improve the diversity of retrosynthesis prediction.

Abstract: Computational retrosynthesis prediction has the potential to reduce development time for newly discovered drugs by automatically generating potential reactions for a target product. Data-driven approaches commonly treat this as a sequence-to-sequence generation task on SMILES strings. In this work, we construct ensembles of discrete-time and continuous-time diffusion models for molecular generation and incorporate guidance mechanisms for improved output diversity. We propose an adaptation of particle guidance to SMILES sequence generation which significantly improves the number of unique molecules generated by diffusion ensembles while increasing top-k accuracy. These results further expand the efficacy of discrete diffusion for SMILES generation, and our empirical analyses offer new insights into the capabilities of diffusion models for chemical retrosynthesis.

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

Submission Number: 13716