Generative Design Comparison
RFdiffusion vs ProteinMPNN vs BindCraft: Protein Design Compared (2026)
Last updated: 2026-04-16
De novo protein design has gone from science fiction to routine workflow in three years. RFdiffusion generates novel protein backbones, ProteinMPNN designs sequences for those backbones, and BindCraft offers a one-shot alternative that combines both steps. These tools serve different roles in the design pipeline — they're often complementary, not competitive.
RFdiffusion
Baker Lab / IPD (University of Washington)
ProteinMPNN
Baker Lab / IPD (University of Washington)
BindCraft
EPFL / MIT (Pacesa, Ovchinnikov, Correia)
Head-to-Head
Structured comparison across key dimensions.
| Dimension | RFdiffusion | ProteinMPNN | BindCraft |
|---|---|---|---|
| What it does | Generates novel protein backbones (3D structures) | Designs amino acid sequences for a given backbone | One-shot binder design (backbone + sequence jointly via AF2 hallucination) |
| Input | Target protein structure + design constraints | Protein backbone structure (PDB) | Target protein structure |
| Output | Novel backbone structures (PDB) — needs ProteinMPNN for sequences | Amino acid sequences (FASTA) predicted to fold into the input backbone | Designed binder sequences + predicted structures ready for experimental testing |
| Experimental success rate | Highly target-dependent; requires screening 100s-1000s of designs | High sequence recovery (~50% native); experimental folding rate high for designed backbones | 10-100% of designs bind experimentally (Nature 2025), often with nM affinity |
| Use together? | Standard pipeline: RFdiffusion → ProteinMPNN → AlphaFold2 validation | Almost always used downstream of a backbone generator | Self-contained; doesn't need RFdiffusion or ProteinMPNN |
| License | MIT | Open source (Baker Lab) | MIT |
| On Platform | Yes | Yes | Unconfirmed |
| Maturity | Research+ — Nature 2023, widely adopted | Production — standard component of all design pipelines | Research+ — Nature 2025, high-impact but newer |
| Key limitation | Generates backbones only; low per-design hit rate requires screening | Cannot generate backbones; not antibody-specific (use AntiFold for CDRs) | Computationally intensive per design; not antibody-specific |
When to Use Each
RFdiffusion
You want to generate novel protein backbones with specific properties (binders, enzymes, symmetric assemblies). You want maximum control over the design process. You'll pair it with ProteinMPNN for sequence design.
ProteinMPNN
You already have a backbone structure (from RFdiffusion, experiment, or homology model) and need to design sequences for it. This is a component of nearly every protein design pipeline.
BindCraft
You want a one-shot binder design pipeline. You want high experimental success rates (10-100%) without needing high-throughput screening. You want simplicity over control.
Practitioner Verdict
RFdiffusion + ProteinMPNN is the standard two-step pipeline: generate backbone, then design sequence. Use this when you need maximum control over the design process. Use BindCraft when you want a simpler one-shot approach with high experimental success rates and don't need the intermediate backbone control. All three are open-source and available available on the platform.
Stay updated on these tools
Weekly briefing on AI tool releases, benchmarks, and what works in drug discovery.