RFdiffusion protein design now available on ProteinIQ

Dr. Matic BrozDr. Matic Broz4 min readNovember 24, 2025
RFdiffusion protein design now available on ProteinIQ

We have integrated RFdiffusion, the state-of-the-art protein structure generation model from the Baker Lab, directly into the ProteinIQ web platform. You can now design protein binders, scaffold functional motifs, and generate symmetric oligomers directly in your browser, eliminating the need for local GPU setups or command-line workflows.

RFdiffusion revolutionized protein design upon its launch in Nature (July 2023). Achieving an 18% success rate on the grand challenge of binder design, it proved 5x to 214x more effective than Rosetta-based approaches. For monomer generation exceeding 100 amino acids, RFdiffusion outperformed hallucination methods, where success rates tend to deteriorate as size increases.

However, the traditional RFdiffusion workflow is complex. It requires generating backbones, designing sequences with ProteinMPNN, validating with AlphaFold2, and filtering scores—all while managing dependencies and GPU access. Research teams want to design proteins, not debug CUDA drivers.

Backbone generation in your browser

We have embedded RFdiffusion directly into ProteinIQ. Simply upload a PDB file (or fetch one from RCSB), configure your parameters, and submit, while our backend handles the backbone generation, returning ranked poly-glycine structures ready for the next step.

Note: Because RFdiffusion outputs backbone coordinates without sequences, you will need to submit your generated structures to ProteinMPNN (also available on ProteinIQ) for sequence optimization. While we are building automated workflows to chain these steps, you currently control each stage of the pipeline.

Supported design modes

The tool supports all five RFdiffusion design modes:

  • Binder design: Generates protein binders targeting specific chains. It includes configurable length ranges (10-200 residues) and optional hotspot biasing to guide interface formation.
  • Motif scaffolding: Extends functional motifs N- and C-terminally while preserving structural integrity. You can specify residue ranges to maintain and set terminal extension lengths (0-200 residues per terminus).
  • Partial diffusion: Redesigns specific regions within existing structures. By defining diffused residue ranges and controlling noising timesteps (10-50 steps), you can balance structural diversity with preservation.
  • Unconditional generation: Designs proteins de novo from a specified length (10-500 residues). It supports cyclic (Cn), dihedral (Dn), tetrahedral (T), octahedral (O), and icosahedral (I) symmetries, with configurable orders from 1-12 subunits.
  • Custom design: Allows advanced users to specify Hydra contig syntax directly for complex scenarios requiring precise control over chain composition and connectivity.

Guiding potentials for design control

We have exposed RFdiffusion's experimental guiding potentials as optional settings, allowing you to bias the diffusion process toward specific structural properties.

Fair warning: start with default settings and experiment carefully. These potentials are powerful but require a clear understanding of the design context.

  • Monomer radius of gyration (RoG): Minimizes RoG to generate compact structures (for unconditional, scaffolding, and partial diffusion).
  • Monomer contacts: Encourages internal contact formation within monomers.
  • Oligomer contacts: Promotes interface contacts in symmetric multimers (unconditional symmetric only).
  • Substrate contacts: Biases binding site formation around ligands found in the input PDB (scaffolding only).
  • Binder potentials: Encourages compact binder structures (RoG) and binder-target interface contacts (binder design only).

Processing configuration

Each RFdiffusion job runs at a base cost of 150 credits per design. Processing is asynchronous, so you can submit your work and monitor its progress via the job history panel.

  • Timesteps: Configure from 20 to 200 (Default: 50). Tip: higher timesteps improve quality but increase runtime. For most applications, 50 timesteps offer the best balance between speed and structural accuracy.
  • Designs per job: Generate 1 to 50 designs (Default: 10).

Design modes in practice

  • Benchmarks: RFdiffusion demonstrates impressive capabilities on challenging topologies where traditional methods often struggle. For example, TIM barrel designs achieved a 42.5% in silico success rate, while NTF2 folds reached 54.1%.
  • Symmetric Oligomers: We support both AnAnaS auto-detection and manual specification of symmetry type and order. RFdiffusion has successfully designed C3-symmetric trimers that rigidly position three binding domains to match ACE2 binding sites on the SARS-CoV-2 spike protein.
  • Validation: In the binder design workflow, the standard pipeline combines RFdiffusion backbones with ProteinMPNN sequences and AlphaFold2 validation (pAE < 10, RMSD < 2Å).
  • Real-world results: Although recent studies note limitations regarding low-affinity designs and expression variability, proper filtering and experimental iteration can consistently produce nanomolar-affinity binders. One study scaled down from 10,000 in silico designs to five recombinant candidates, resulting in one nanomolar lead.
  • Comparisons: Motif scaffolding significantly outperforms Constrained Hallucination and RFjoint Inpainting. It effectively handles single motifs or multiple disconnected segments, ensuring complex active sites are preserved.