Three questions, three model roles
For a researcher choosing a design strategy, the first question is what must be generated. RFdiffusion generates protein backbones and supports conditional design, including motif scaffolding and binder-design settings. ProteinMPNN proposes amino-acid sequences conditioned on a supplied backbone. Boltz predicts biomolecular structures from specified entities; Boltz-2 additionally offers a separate small-molecule affinity module. These responsibilities are complementary, but a chain of model outputs is a design hypothesis.
rf-README.md · mpnn-README.md · boltz-docs/prediction.md
Choose by the missing information
| Research need | Candidate role | Remaining question |
|---|---|---|
| New scaffold subject to geometric constraints | RFdiffusion | Are the generated geometry and constraints acceptable? |
| Sequences for a selected backbone | ProteinMPNN | Is chain/residue mapping preserved? |
| A predicted structure or complex for specified sequences/entities | Boltz | What do local and interface confidence indicate? |
This is an editorial decision framework, not a ranking. If the sequence is already fixed, backbone generation is not automatically the next task. If function is the goal, define an experimental endpoint before selecting computational filters.
The handoff is a scientific contract
RFdiffusion documents PDB structure outputs and accompanying design metadata; ProteinMPNN uses parsed backbone information and writes designed sequences in FASTA outputs. Boltz YAML distinguishes protein, DNA, RNA, and ligand entities. Ligands use SMILES or CCD identifiers, and protein MSA handling must be explicit. File extension alone does not establish compatibility. Retain design identifiers, chain IDs, residue numbering, fixed positions, omitted residues, and constraint settings; verify them before preparing the next input.
rf-README.md · mpnn-README.md · boltz-docs/prediction.md
Separate computational checks from biological evidence
Inspect preserved motifs and chain boundaries, plausible geometry, and the predicted relationship between a designed sequence and its intended backbone. Treat these as proposed quality checkpoints. A high confidence score does not measure expression, folding yield, binding, or function. In particular, Boltz-2 affinity is scoped to small molecules binding protein targets; it must not be repurposed as protein–protein binder affinity. Record code revision, checkpoint identity and rights, seeds, settings, and all excluded designs. Using a remote MSA service introduces a separate sequence-transfer decision.
Related resources and reading
RFdiffusion · ProteinMPNN · Boltz
Protein Backbone to Sequence with RFdiffusion and ProteinMPNN · Boltz-2 Complex Prediction and Small-Molecule Affinity Analysis
Protein & Structural Biology AI Workflow · A checklist for assembling chemistry model workflows · AI for Drug Discovery: Agents, MCP Servers and Platforms
Sources and evidence boundary
Sources were inspected on 2026-10-08. Revision-pinned project documentation supports capability statements. Evaluation choices are editorial proposals. This article reports no executed workflow, measured performance, or experimental validation.