ReStateMPNN: State-Retentive Modeling for Multi-State Protein Inverse Folding

Protein inverse folding is typically formulated for a single target structure, yet many functional proteins must remain compatible with multiple conformational states. Existing multi-state methods largely treat this setting as an information-fusion problem, combining state-conditioned predictions or aggregating state-indexed representations before sequence decoding. This formulation obscures structural changes between conformations and compresses state-specific variation before autoregressive sequence decisions are made. We introduce ReStateMPNN, a multi-state inverse-folding framework that jointly models cross-state structural relationships and preserves state-specific information throughout sequence generation. ReStateMPNN conditions cross-state communication on global conformational relations and residue-local contact rearrangements, while retaining state-specific residuals for sequence-context-dependent retrieval during decoding. Compared with ProteinMPNN${}_{\mathrm{MSD}}$, ReStateMPNN improves sequence recovery from 38.0% to 45.0%, increases worst-state TM-score from 0.574 to 0.832, and reduces RMSD from 5.84 to 1.78 , with controlled ablations supporting complementary contributions from relation-conditioned communication and state-retentive decoding.