RFdiffusion

What is RFdiffusion?

RFdiffusion is a generative AI model for designing novel protein structures with atomic precision. Developed by the Baker Lab at the University of Washington and published in Nature (2023), it uses diffusion models, the same technology behind DALL-E and Stable Diffusion, adapted for protein structure generation. With over 1,000 citations since publication, the method achieves approximately two orders of magnitude improvement over traditional computational protein design approaches.

Independent experimental validation demonstrates 84% confirmation rates for designed binders, with affinities ranging from tens of micromolar to tens of nanomolar. Applications include therapeutic binder design, enzyme active site scaffolding, symmetric oligomer creation, and de novo protein generation.

RFdiffusion works by iteratively denoising random 3D coordinates into structured protein backbones through learned reverse diffusion steps. Unlike traditional methods that optimize physics-based energy functions, it learns directly from the statistical distribution of natural protein structures in the Protein Data Bank.

How does RFdiffusion work?

Architecture

RFdiffusion builds on the RoseTTAFold (RF) structure prediction network architecture, similar to AlphaFold2. The model was initialized with RoseTTAFold's pretrained weights and fine-tuned specifically for structure denoising tasks. It uses SE(3)-equivariant graph neural networks that respect rotational and translational symmetries, improving generalization and data efficiency.

The architecture processes protein structures as rigid-frame representations: coordinates and orientations of backbone atoms (N, Cα, C, O, and virtual Cβ). This representation enables precise control at the residue level while maintaining physical realism.

Diffusion process

The model operates through reverse diffusion on protein backbone structures. Starting from random noise (random 3D coordinates), it iteratively denoises through learned steps toward a structured protein. At each timestep, the model predicts the final structure from the current noised structure, then interpolates from the current coordinates toward the predicted structure.

Training used 200 discrete diffusion timesteps, though the default deployment uses 50 steps for optimal quality-speed balance. Research shows that as few as 20 timesteps achieve equivalent quality with 10x speedup. Higher timestep counts improve quality with diminishing returns beyond 50.

Self-conditioning

A critical innovation is self-conditioning: the model receives its previous prediction as template input, similar to AlphaFold2's "recycling" mechanism. This architectural choice improves prediction quality by allowing the model to iteratively refine its understanding of the emerging structure.

Key innovations

Unlike traditional protein design methods requiring manual parameter tuning and physics-based optimization, RFdiffusion treats design as generative modeling over protein structure space. It requires no evolutionary information or multiple sequence alignments (MSAs). The approach enables diverse design tasks through conditional generation, specifying constraints like binding sites, motifs, or symmetry while allowing the model to generate compatible structures.

Design modes

RFdiffusion v1.1.0 supports six design modes on ProteinIQ for different applications:

Binder design

Creates proteins that bind to a specified target protein. You provide the target structure and optionally specify binding pocket residues (hotspots). The model generates novel binder proteins with specified length ranges that form interfaces with the target. This mode has been experimentally validated with picomolar-affinity binders to therapeutic targets including MDM2, PD-L1, and IL-7Rα.

Motif scaffolding

Builds a protein scaffold around a functional motif of interest. You specify which residues to preserve and how much structure to add at the N- and C-termini. Applications include scaffolding viral epitopes, receptor binding sites, enzyme active sites, and metal-binding motifs. This enables transplanting functional elements into new structural contexts.

Partial diffusion

Partially redesigns an existing protein structure to create variants. By controlling the diffusion timesteps (partial_T), you can tune the diversity: low values create subtle variations, high values generate more radical changes. Useful for exploring structural space around a known fold while maintaining key geometric features.

Unconditional generation

Generates novel protein structures from scratch without any template. You specify the desired length and optional symmetry. The model samples entirely new protein folds from learned structural distributions. This mode enables exploring uncharted regions of protein structure space.

Custom design

Advanced mode for users who want precise control using contig syntax. Contigs define exactly which regions to keep from input structures and where to generate new structure. This enables complex design tasks like inserting domains, creating fusion proteins, or specifying exact topological arrangements.

Fold conditioning

Uses paired secondary-structure and block-adjacency tensors to guide the sampled fold. Upload matching <id>_ss.pt and <id>_adj.pt files produced by RFdiffusion's scaffold preparation scripts. Optional target PDB, target SS, and target adjacency files support fold-conditioned binder design. These tensors are native scientific inputs, separate from protein sequence design.

Input requirements

RFdiffusion v1.1.0 uses different inputs for each design mode. Each uploaded file is limited to 50 MB.

Input	Formats	Requirement
Input structure (protein)	.pdb, .ent, or an RCSB PDB ID	Required for binder design, motif scaffolding, and the partial-diffusion preset; optional for custom contigs and fold conditioning; absent in unconditional generation.
Scaffold tensors (scaffold_tensors)	.pt	Fold conditioning only: 2 to 100 files, supplied as matching <id>_ss.pt and <id>_adj.pt pairs.
Fold target structure (target)	.pdb, .ent	Optional target for fold-conditioned binder design, uploaded separately from the input structure.
Target secondary structure tensor (target_ss)	.pt	Optional secondary-structure vector for the fold target.
Target adjacency tensor (target_adj)	.pt	Optional block-adjacency matrix for the fold target.

Uncropped binder targets are limited to 250 residues. An explicit crop range or a cropped structure is required for larger targets. PDB residue numbers retain their original numbering, including gaps; selected crop endpoints must exist in the native residue map. The partial-diffusion preset requires a single-chain PDB. Custom partial-diffusion contigs support complexes, but their total length must match the input structure. A PDB is required whenever a custom contig references retained residues or partial diffusion starts from an existing structure.

Fold tensors must be plain secondary-structure vectors or adjacency matrices, rather than model objects or TorchScript programs. Preserve the original paired filenames and omit periods from scaffold IDs. The official make_secstruc_adj.py helper generates these files. In workflows, the Tool Files root input accepts .pt files and routes a scaffold pair together to the scaffold tensor collection.

Pricing

Pricing starts at 50 credits per job. The upfront quote scales with the number of designs, diffusion steps, and modeled structure length, up to 480 credits. Binder quotes include the selected target chain or crop and the maximum binder length. Scaffolding and custom contigs use the maximum requested length; partial diffusion uses the full input structure and its partial diffusion steps. Fold-conditioned quotes use scaffold tensor storage length and requested extensions, with a conservative fallback when tensor metadata cannot be read. Checkpoint choice and guiding potentials do not add a separate charge.

The exact quote is shown and charged before submission. Runtime does not change the charge. Each child job in a batch or workflow receives its own quote. Native inference retains its 20-minute timeout.

Settings

Defaults below apply to new runs. Conditional settings are shown for the relevant design mode or when guiding potentials are enabled. Residue ranges use inclusive endpoints, and each minimum must not exceed its corresponding maximum.

Core settings

Parameter	Type	Default	Description
Design mode (design_mode)	enum	binder	binder, scaffolding, partial_diffusion, unconditional, custom, or fold_conditioning.
Number of designs (num_designs)	integer	10	Generate 1 to 50 independent structures in one job.
Timesteps (timesteps)	integer	50	Total diffusion steps, 20 to 200; the slider uses increments of 10.
Total assembly length (protein_length)	integer	100	Unconditional mode only: 10 to 500 residues across the complete assembly; symmetric lengths must divide evenly across subunits.
Use beta model (use_beta_model)	boolean	false	Binder mode only: select Complex_beta_ckpt.pt; incompatible with an explicit different checkpoint.
Checkpoint (checkpoint)	enum	auto	Native automatic selection, or Base_ckpt.pt, Complex_base_ckpt.pt, Complex_beta_ckpt.pt, Complex_Fold_base_ckpt.pt, InpaintSeq_ckpt.pt, InpaintSeq_Fold_ckpt.pt, ActiveSite_ckpt.pt, or Base_epoch8_ckpt.pt; the selected model must support the task.
Deterministic sampling (deterministic)	boolean	false	Seed each design using its native design index.
First design index (design_startnum)	integer	0	Nonnegative starting index for native design numbering and deterministic seeds.

Job name is an optional label for the saved job and does not alter the scientific settings.

Binder design

These settings apply to binder mode.

Parameter	Type	Default	Description
Target chain (binder_input_chain)	string	A	Single PDB chain containing the target.
Binding pocket residue start (binding_pocket_res_start)	integer	1	Inclusive target crop start, at least 1; setting both crop endpoints to 1 uses the full chain.
Binding pocket residue end (binding_pocket_res_end)	integer	1	Inclusive target crop end, at least 1; an explicit crop can begin at residue 1.
Binder length minimum (binder_length_min)	integer	10	Lower bound for generated binder length, 5 to 100 residues.
Binder length maximum (binder_length_max)	integer	20	Upper bound for generated binder length, 5 to 200 residues.
Hotspots (hotspots)	string	optional	Comma-separated target residues, such as A50,A51,A52; range notation such as A50-64 is unsupported.

Hotspots can come from experimentally validated sites or predicted epitopes. A 40 to 80 residue binder is a common starting range for single-domain designs; the default preset samples shorter 10 to 20 residue binders.

Motif scaffolding

These settings apply to scaffolding mode. Setting both extensions at one terminus to 0 omits that extension.

Parameter	Type	Default	Description
Motif chain (motif_input_chain)	string	A	Single PDB chain containing the motif.
Minimum N-terminal extension (motif_min_n_terminal)	integer	5	Lower bound for added N-terminal residues, 0 to 200.
Maximum N-terminal extension (motif_max_n_terminal)	integer	10	Upper bound for added N-terminal residues, 0 to 200.
Minimum C-terminal extension (motif_min_c_terminal)	integer	5	Lower bound for added C-terminal residues, 0 to 200.
Maximum C-terminal extension (motif_max_c_terminal)	integer	10	Upper bound for added C-terminal residues, 0 to 200.
Scaffold start residue (scaffold_start)	integer	1	Inclusive preserved-motif start, at least 1; setting both scaffold endpoints to 1 preserves the full selected chain.
Scaffold end residue (scaffold_end)	integer	1	Inclusive preserved-motif end, at least 1.

Partial diffusion

These settings apply to partial_diffusion mode. Setting both diffused residue endpoints to 1 selects the full chain.

Parameter	Type	Default	Description
Partial diffusion chain (partial_diffusion_chain)	string	A	Chain in the required single-chain PDB; use custom contigs for a complex.
Diffused residue start (diffused_residue_start)	integer	10	Inclusive beginning of the region to redesign, at least 1 and present in the input PDB.
Diffused residue end (diffused_residue_end)	integer	20	Inclusive end of the region to redesign, at least 1 and present in the input PDB.
Partial noising steps (partial_t)	integer	10	Explicit diffuser.partial_T, 1 to 200 and no greater than total timesteps; 10 is a ProteinIQ preset rather than an automatic RFdiffusion default.

Custom contigs and sequence controls

Parameter	Type	Default	Description
Custom contig (custom_contig)	string	conditional	Required in custom mode and optional in fold conditioning; A10-100/0 50-150 retains A10 to A100, adds a chain break, and generates a 50 to 150 residue chain.
Partial diffusion timesteps (custom_partial_t)	enum	none	Custom mode only: none, 10, 20, 30, 40, or 50; a selected noising count must not exceed total timesteps.
Provided sequence (provided_sequence)	string	optional	Custom partial diffusion only: preserve sequence at zero-based inclusive indices, such as 80-109; requires a selected partial-diffusion timestep value.
Sequence masking (inpaint_sequence)	string	optional	Hide sequence identity at PDB residues while retaining their structure, using native syntax such as A10-20/A30.
Sampled contig length (contig_length)	string	optional	Constrain the designed contig length with an integer or inclusive range, such as 100 or 100-120.

Contig syntax and sequence indexing follow the official RFdiffusion reference. The space after /0 marks a chain break. Fold conditioning cannot combine a separate fold target structure with a custom contig.

Noise and motif alignment

Parameter	Type	Default	Description
Coordinate noise scale (noise_scale_ca)	number	1	Nonnegative scale for translational noise; the control uses increments of 0.1.
Frame noise scale (noise_scale_frame)	number	1	Nonnegative scale for rotational noise; the control uses increments of 0.1.
Align motif (align_motif)	boolean	true	Align the motif during native inference.

Symmetry

These settings apply to unconditional and custom modes. Cyclic oligomer symmetry is supported; covalent macrocyclic-chain generation is unavailable in RFdiffusion v1.1.0.

Parameter	Type	Default	Description
Symmetry (symmetry)	enum	none	none, cyclic, dihedral, tetrahedral, octahedral, icosahedral, or t3; fixed symmetries use 12, 24, 60, or 4 subunits respectively.
Symmetry order (symmetry_order)	integer	1	Cyclic Cn or dihedral Dn order, 1 to 12; cyclic assemblies use n subunits, dihedral assemblies use 2n, and fixed symmetries ignore this value.
Recenter symmetric assembly (recenter)	boolean	true	Recenter coordinates for native symmetry generation.
Symmetry radius (symmetry_radius)	number	10	Nonnegative native inference.radius in angstroms; increments of 0.1.
Model neighboring subunits only (model_only_neighbors)	boolean	false	Restrict the symmetry model context to neighboring subunits.
Symmetric self conditioning (symmetric_self_cond)	boolean	true	Apply symmetry during self conditioning.

Guiding potentials

These controls are experimental. The parent Guiding potentials switch must be enabled; disabling it omits every potential and associated control, even if child selections are retained. Preset switches use fixed native weights. Native guiding potentials replaces the preset potential list with supplied native strings.

Parameter	Type	Default	Description
Guiding potentials (use_guiding_potentials)	boolean	false	Enable the potential controls below.
Monomer radius of gyration (potential_monomer_rog)	boolean	false	Add monomer_ROG with weight 1 to encourage compact monomers.
Monomer contacts (potential_monomer_contacts)	boolean	false	Add monomer_contacts with weight 1 to encourage intramolecular contacts.
Oligomer contacts (potential_oligomer_contacts)	boolean	false	Add olig_contacts with intra weight 1 and inter weight 0.1; enables all intra- and inter-subunit contacts unless explicitly overridden.
Substrate contacts (potential_substrate_contacts)	boolean	false	Add substrate_contacts with weight 1; requires a substrate residue name and the corresponding ligand in the input PDB.
Substrate residue name (substrate_name)	string	conditional	Three-letter PDB ligand name, such as NTP; shown when the substrate-contact preset is enabled.
Binder radius of gyration (potential_binder_rog)	boolean	false	Add binder_ROG with weight 1 to encourage compact binders.
Binder & interface contacts (potential_binder_contacts)	boolean	false	Add native interface_ncontacts with weight 1 to encourage binder-target interface contacts.
Native guiding potentials (guiding_potentials_json)	string	optional	JSON array of native potential strings, such as ["type:monomer_ROG,weight:1,min_dist:15"]; replaces preset potential strings and permits native weights and parameters.
Potential guide scale (guide_scale)	number	10	Nonnegative multiplier for potential guidance, in increments of 0.1.
Potential guide decay (guide_decay)	enum	constant	constant, linear, quadratic, or cubic guidance decay.
Oligomer contact matrix (olig_custom_contact)	string	optional	Native pair-contact specification, such as A&A,B!B.
Olig intra all (olig_intra_all)	enum	Native default	Leave unset, or choose enabled (true) or disabled (false) for all intra-subunit contacts.
Olig inter all (olig_inter_all)	enum	Native default	Leave unset, or choose enabled (true) or disabled (false) for all inter-subunit contacts.

Fold conditioning

These settings apply to fold_conditioning mode and require the paired scaffold tensors listed above.

Parameter	Type	Default	Description
Loop insertions (fold_sampled_insertion)	string	0	Sample extra residues in each loop from 0 up to an integer, or within an inclusive range such as 5-15.
N-terminal additions (fold_sampled_N)	string	0	Sample added N-terminal residues from 0 up to an integer, or within an inclusive range.
C-terminal additions (fold_sampled_C)	string	0	Sample added C-terminal residues from 0 up to an integer, or within an inclusive range.
Target crop (fold_contig_crop)	string	optional	Native target contig crop, such as A1-100, using PDB chain and residue numbering.
Secondary structure masking (fold_ss_mask)	integer	0	Nonnegative count of residues to mask at each end of a helix or strand block.
Cycle scaffolds systematically (fold_systematic)	boolean	false	Cycle through the scaffold selection in order; otherwise sample scaffolds randomly.
Mask scaffold loops (fold_mask_loops)	boolean	true	Mask loop conditioning; when disabled, loop insertions and both terminal additions must be 0.
Scaffold selection (fold_scaffold_ids)	string	optional	Supplied scaffold IDs, one per line, in the desired order; omit _ss.pt and _adj.pt, or leave blank to use all supplied pairs.

Fold insertion, masking, and target-crop semantics follow the native fold-input implementation.

Understanding the results

Native output files

The Viewer presents structures in native design order, with design number, filename, and download controls. Data lists primary output files, and Files includes the supporting artifacts. Design numbers indicate generation order rather than a confidence ranking.

RFdiffusion returns its native backbone design files:

• Design PDB files: Final RFdiffusion backbone predictions. Designed residues are output as glycine because RFdiffusion performs backbone generation, not sequence design or side-chain packing.
• Run diagnostics: Complete stdout, stderr, and native Hydra logs/configuration files. If inference fails or times out, the job remains failed and any already-written design files are retained.
• TRB metadata files: Native run metadata for each design, including the sampled contig, RFdiffusion config, pLDDT trajectory values, device, runtime, and residue mapping arrays.
• Trajectory PDB files: Native Xt-1 and pX0 multi-step trajectory files when trajectory output is enabled by RFdiffusion. These show the denoising path and can be inspected in molecular viewers that support multi-model PDBs.

Run ProteinMPNN, LigandMPNN, AlphaFold, or a related validation workflow after RFdiffusion when you need sequence design, side-chain packing, or fold-confidence metrics.

Use cases

Protein binder design

High-affinity binders to therapeutic targets represent the most experimentally validated application. Published examples include nanomolar binders to MDM2 (0.5-0.7 nM vs 600 nM native), influenza hemagglutinin, IL-7Rα, PD-L1, and TrkA. Independent validation by Adaptyv Bio on IL-7Rα demonstrated 84% confirmation rate with 23/27 binders showing measurable affinity (strongest: 40 nM).

Cryo-EM structures of designed binders show near-perfect agreement with computational models. Some designed interfaces achieve picomolar affinity through pure computation without experimental optimization.

Enzyme active site scaffolding

RFdiffusion can scaffold catalytic residues into novel protein folds with specified symmetry. The method enables de novo enzyme design by building NTF2-like folds around functional sites. Applications include designing protein-metal assemblies around coordination sites and transferring catalytic motifs into alternative structural contexts.

Symmetric oligomer design

Generates novel symmetric assemblies validated by electron microscopy across symmetry types, including cyclic (Cn), dihedral (Dn), tetrahedral (T, 12 subunits; t3, 4 subunits), octahedral (O), and icosahedral (I). Examples include C3 symmetric trimers targeting SARS-CoV-2 spike protein. Hundreds of designed metal-binding symmetric proteins have been experimentally characterized.

De novo protein generation

Unconditional mode creates entirely novel protein folds not seen in nature. No template or evolutionary information required. Topology-constrained monomer designs demonstrate the model's ability to explore uncharted regions of protein structure space.

Success rates

Original Nature paper: 55/96 designs (57%) showed detectable binding at 10 μM, representing ~2 orders of magnitude improvement over previous methods. However, recent critical evaluations show variable success for challenging eukaryotic targets, highlighting the need for generating hundreds to thousands of designs for difficult cases.

Newer methods like BindCraft and EvoPro achieve ~50% success rates (>10x improvement over earlier approaches), while specialized tools show varying performance: RFpeptides (macrocycles) 1.72%, Latent-X 8.26%.

Best practices

Getting started

Start with unconditional mode (100-150 residues) to understand outputs and quality metrics. Use default parameters initially: 10 designs and 50 timesteps. Begin with simple design tasks before adding guiding potentials or complex contigs.

Binder design workflow

1. Prepare target structure: Clean PDB, remove waters, ensure proper protonation
2. Identify binding site: Use hotspots if experimentally known
3. Generate 10-20 initial backbone designs with default parameters
4. Run sequence design and fold prediction before filtering by pAE_interaction, pLDDT, ipTM, or related downstream metrics
5. Validate top 3-5 candidates with MD simulations
6. Plan experimental testing: Expect 10-100x designs needed vs confirmed hits

Motif scaffolding strategy

Define motif precisely with exact residue ranges. Allow sufficient N/C-terminal extensions (minimum 5-15 residues) for structural stability. Use substrate_contacts guiding potential if ligand is present. Validate that motif geometry is preserved in outputs by measuring RMSD.

Partial diffusion approach

Start with partial_T=25-30 for moderate diversity around existing fold. Increase to 40-50 only when more radical structural changes are needed. Use provide_seq parameter to retain sequences of critical functional residues. Check RMSD to input structure to ensure appropriate diversity level.

Parameter tuning guidance

Increase timesteps when facing complex topology requirements, poor initial results at 50 steps, or preparing final production runs. Enable beta model when outputs show excessive helices or need better secondary structure balance. Add guiding potentials only after trying without them first; use for specific biophysical requirements informed by experimental feedback.

Validation pipeline

1. Sequence and fold prediction: Run ProteinMPNN or LigandMPNN, then evaluate AlphaFold pLDDT, pAE, ipTM, and pTM metrics
2. Structural: Run MD simulations, Rosetta relaxation
3. Biophysical: Compare AlphaFold2 multimer predictions across candidate sequences
4. Experimental: Expression testing, purification, binding assays
5. High-resolution: X-ray crystallography or cryo-EM for final candidates

Common pitfalls to avoid

Skipping downstream fold-confidence checks can hide weak designs. Design numbers indicate generation order. Compare several designs and apply downstream validation before selecting a candidate. Insufficient sampling (too few designs) reduces statistical power for finding successful candidates. Skipping proper protein preparation leads to artifacts. Over-reliance on any single metric misses important failure modes captured by complementary scores.

Limitations

Technical constraints

Standard RFdiffusion performs backbone-only design without explicit side chain or ligand modeling (use LigandMPNN for post-processing). The target protein is treated as rigid during design. Water molecules and explicit solvation are not modeled. ProteinIQ applies a 20-minute inference timeout and upfront pricing starting at 50 credits per job.

Biological challenges

Success rates vary from 1-50% depending on target difficulty. Designed sequences may show low recombinant expression in standard systems. Some designs exhibit non-specific binding or promiscuity. Affinity ranges cluster in hundreds of nanomolar, often requiring optimization for therapeutic applications.

Applicability limits

Primarily designs with canonical amino acids. Practical size constraints: 10-500 residues. Novel fold generation remains challenging despite improvements over alternatives. Metalloproteins require special consideration. Covalent modifications are not directly supported.

Model uncertainty

Confidence scores are predictions, not guarantees of experimental success. Training on x-ray crystal structures may reduce performance on computational models. Domain generalization struggles with highly novel protein families. Limited training data for macrocycles and non-standard protein architectures.