How to Use RFOptimization

RFOptimization alternates structure optimization/prediction and sequence redesign. Start with a complex containing design protein chain A and partner chain B. RF3 gradient optimization, Boltz-2 prediction, or mixed cycling produces structures for MPNN redesign. Partner sequences remain fixed. The result includes chronological cycle metrics, chain-A FASTA sequences, predicted complexes, redesigned complexes, and the final optimized design.

Upload one starting complex with complete standard-protein backbone atoms in chain A. Put the interaction partner in chain B and each additional protein, nucleic acid, or ligand entity in a separate chain. Choose Mixed RF3 and Boltz-2, RF3 Only, or Boltz-2 Only, then set the number of cycles. Mixed cycling chooses a branch independently each cycle using RF3 Probability.

RF3 optimizes the chosen confidence objective and fixes chain-A residues within 8 Å of chain B for the subsequent MPNN step. Boltz-2 fixes high-confidence chain-A runs of at least five residues. MPNN redesigns the remaining chain-A positions and carries one sample into the next cycle. Use LigandMPNN for ligand-aware redesign. Cyclic Peptide is available only for Boltz-2 Only.

Sequence constraints: Enter either Fixed Residues to preserve a motif or Designable Residues to permit changes only at selected positions. Leave both empty for the default scope. Use the original input structure IDs, such as A23,A40-55,A42B; ranges include existing residues between the endpoints, and insertion codes are supported as individual IDs. At least one position must remain designable. Constraints follow observed sequence positions if output structures are renumbered. Fixed identities are enforced during RF3 optimization and every MPNN redesign; coordinates are free to move.

MPNN Omit Amino Acids excludes letters globally from user-designable positions. MPNN Per-Residue Omissions accepts JSON such as {"A23":"CP","A42B":"G"}. Per-position keys must identify designable residues. Exclusions are combined, and cannot forbid all standard amino acids at a position. Fixed motifs may retain globally excluded amino acids. When necessary, automatic MPNN confidence/contact fixes are released at user-designable positions to satisfy exclusions. These exclusion rules apply after MPNN redesign; RF3 can temporarily propose excluded amino acids.

Scores and predicted structures describe the structure-prediction step before that cycle's MPNN redesign. The redesigned sequence and final optimized PDB are the outputs of the subsequent MPNN step and have not received an additional final prediction. Redesigned side chains are not repacked. RF3 and Boltz metrics are shown separately because their definitions and scales can differ. Results are chronological, and the final cycle is not necessarily the highest-confidence design. Computational confidence does not establish affinity or biological activity.

• Combines RF3 gradient optimization, Boltz-2 prediction, and MPNN redesign.
• Retains partner sequences while optimizing the design chain.
• Reports each cycle's structures, sequences, and confidence metrics.