Mature Without Breaking: Doob-Twisted Sequential Monte Carlo to Guide Discrete Diffusion for Constrained Antibody Affinity Maturation

Affinity maturation must strengthen antigen binding while preserving the properties that made a parent antibody viable, a multi-objective problem under a tight mutation budget. Existing inference-time steering of discrete generative models either pursues a single objective or silently changes the distribution it claims to sample, so reported gains may reflect a redefined target. We present Doob-twisted Sequential Monte Carlo, a general framework to steer any pretrained discrete diffusion generator toward a reward-tilted design law over budget-feasible variants without retraining it. At each residue reveal it scores every candidate substitution by the future reward mass reachable through it, twists the proposal accordingly, and applies exact sequential and terminal correction, so the twist governs search efficiency alone, never the sampled law. We also pretrain a masked discrete-diffusion prior over paired antibody repertoires and train antigen-conditioned affinity and antibody-only developability oracles as the base generator and reward. We demonstrate its effectiveness in generating variants co-optimized across affinity and developability properties, returning diverse co-improving fronts for every parent from one fixed configuration and beating the strongest post-proposal baseline on every antigen tested. In total, Doob-twisted Sequential Monte Carlo proves to be an auditable route to multi-objective antibody affinity maturation.