RFD4-Proteina

RFD4-Proteina webserver overview

RFD4-Proteina 0.0.3 is an all-atom flow-matching model from the Baker Lab and NVIDIA for designing proteins around a stated condition: a binder for a target protein, a scaffold around an enzyme motif, or a pocket around a small molecule. It samples backbone, sequence and side chains together, so every design arrives with its own sequence. ProteinIQ offers a task for each common design problem, writes the condition specification from a few fields, and returns each design as an annotated CIF file.

Three workflows are available. Generate designs builds new chains from noise, Partial diffusion re-noises a supplied structure part-way and refines it, and Inverse folding keeps a supplied backbone and resamples its sequence and side chains.

The tool is in beta and runs the design step only. Metrics, refolding with the source's AlphaFold2, AlphaFold 3 and OpenFold 3 oracles, reward-guided search, and the antibody and nanobody campaign configuration are not offered. Other design models on ProteinIQ include RFdiffusion3 and Proteina-Complexa.

How RFD4-Proteina works

Each residue is described by the position of its C-alpha atom and a vector of 32 latent channels that carries its identity and side-chain atoms. An autoencoder, trained first and then frozen, defines those latents and decodes them back into atoms, so the design model works on a compact description instead of every atom.

Generation is flow matching. Both the C-alpha positions and the latents start as Gaussian noise and move toward a design over 60 sampling steps. At each step a 178-million-parameter transformer reads the noisy state and the conditions, then predicts a velocity for the positions, a velocity for the latents and a sequence readout, and an integration step applies them.

The transformer's trunk can run several times on its own output within a step. The Recycling schedule sets the passes at each of the 60 steps, and the presets total 120 to 300 passes.

After the last step the decoder turns the final positions and latents into residue types, all-atom coordinates and the set of atoms present. Each run therefore returns a sequence and a full-atom structure together, with no separate sequence design step. The Temperature settings rescale the score during integration, Sampling noise chooses stochastic or deterministic sampling, and Guidance weight scales classifier-free guidance.

Conditions enter as annotations on the starting CIF. Where one is set, the model receives the value and keeps to it; elsewhere it designs the value. Fixing a side-chain coordinate also fixes the residue's identity, because the pattern of occupied atom slots reveals the amino acid, but fixing only the backbone atoms N, CA, C and O does not. That is how a backbone is held while its sequence is redesigned.

Only protein can be designed at present. Ligands, metals and modified residues in the input are context the model is given and never generates. Partial diffusion starts the loop part-way along the trajectory from the supplied structure with noise added back, so Sampling steps is unchanged, and Inverse folding holds the backbone and resamples only the latents.

Pricing

Runs cost 22 credits per minute of measured runtime. Completed runs are charged in proportion to elapsed time, rounded up to a whole credit, and unused reserved credits are returned. Charged runtime is capped at 3,300 seconds, which is 1,210 credits. Set a spending limit before submission.

The quote shown before submission is an estimate of that charge, starting at 41 credits for one design at 60 sampling steps. It allows 60 seconds for startup and model loading plus 50 seconds per design at 60 steps, scaled in proportion to Sampling steps. These figures are provisional while runtimes are measured on production hardware.

The estimate is calculated from Designs per input and Sampling steps only. Target size, contig length, workflow and recycling schedule do not change it, although they change the measured runtime that is charged. Steps other than 60 need the No schedule recycling option. Twenty designs at 200 steps reach the runtime cap, so the quote stops at 1,210 credits.

Inputs

Each task takes one CIF structure, writes the condition specification from the fields below it, and follows the model's own recipe for that task. The specification it wrote is returned as input/spec.yaml, so a run can be repeated or edited under Custom conditions. A structure is a plain .cif of at most 50 MiB, uploaded or fetched from RCSB. Compressed CIF and binary CIF (.cif.gz, .bcif, .cif.zst) are not accepted; decompress them first. Account upload limits may be lower.

Residue numbers follow the numbering the model reads: label_seq_id for protein residues and the author residue number for ligands. Some viewers show author numbering for protein residues, which can differ; in PyMOL, set cif_use_auth, off shows the label numbers. A range such as A17-131 needs every residue of it to be present in the structure. A task refuses a structure it cannot express and names the residue, atom or ligand at fault; Custom conditions covers those cases.

Protein binder

The task keeps the target residues, samples a binder chain of the length range you give, and conditions the target as the model's PD-L1 example does: the target backbone as one rigid coordinate group, the target sequence, and the hotspot atoms you list. Target residues takes one contiguous range per chain, such as A17-131; blank keeps every protein chain whole, which needs chains without gaps.

Hotspot atoms are written as chain and residue, a colon and atom names, such as A56:CG,OH A115:CG,SD A123:CD2,OH, separated by spaces. The model's example uses two side-chain atoms for each of three residues.

Small-molecule binder

The task keeps one ligand with its coordinates and designs one protein chain around it, the layout of the model's SAM benchmark input. The structure must contain the ligand; a SMILES string is not enough. Ligand residue names it as a chain and the author residue number, such as B386, and is needed only when the structure holds more than one ligand. Some crystallization additives, such as formate, are dropped by the model's loader and cannot be kept.

Motif scaffolding

The task keeps the motif residues from one chain and samples a designed segment of the given length range before, between and after them. With Keep motif side chains on, every motif atom and its residue type are fixed. Off fixes only the backbone atoms N, CA, C and O, so the model designs the motif sequence. Each motif range must be contiguous in the structure, because the model treats a kept range as consecutive residues.

Enzyme design

The task follows the layout of the model's M-CSA benchmark inputs. The catalytic residues from one chain are kept with their side chains, each as its own unindexed group, so the model places them anywhere in a designed scaffold of the given length range. Ligand residues keeps ligands, cofactors and metals with their coordinates, each as a chain of its own, written as chain and author residue number such as C476 D477.

Existing structure

The task runs the chosen workflow on the whole structure. Partial diffusion re-noises it part-way, and the partial diffusion settings choose how far. Inverse folding keeps the backbone and resamples the sequence and side chains. Turn on Structure is an RFD4-Proteina design for a design from an earlier run, so that its condition annotations are kept.

Custom conditions

Custom conditions passes the specification you write and up to 20 structures to the model unchanged. Use it for what the tasks do not cover, such as motifs from several chains, a gap inside a kept range, or conditions beyond those above. The specification a task wrote is a starting point.

A specification names each design problem and states what to keep, what to build and what to condition. Every top-level key except global_args is one design problem, and Designs per input applies to each of them. An unrecognized field under a design raises an error.

A global_args key holds any of these fields as the default for every design in the file, and a value under a design's own key overrides it.

The contig is one comma-separated string.

Atoms outside a kept segment are dropped, and output residues are numbered from 1 in each chain. Residue numbers in contig and conditions refer to the original label_seq_id of the input CIF, which can differ from the author numbering that viewers such as PyMOL show. Turn author numbering off in the viewer (set cif_use_auth, off in PyMOL) before reading residue numbers from it.

Atoms and residues are chosen with select: strings. Path syntax is sel('A/*/[10-20]'), written as chain, residue name, residue number and atom name with trailing fields defaulting to *. Query syntax is a Python boolean expression over annotations such as is_ligand or chain_id, and the two can be mixed.

Each annotation maps a value to a list of selections. Overlapping entries for the same annotation resolve in declaration order, and the last one wins.

The Protein binder task writes the following specification for PD-L1, chain A residues 17 to 131 of PDB 5O45, with three hotspot residues.

pdl1_binder:
  input: 5o45_cropped.cif
  contig: "50-120, /0, A17-131"
  conditions:
    C_CRDGRP: {1: [{select: "sel('A/*/*/[N,CA,C,O]') and not has_nan_coord()"}]}
    C_SEQ: {true: [{select: "sel('A')"}]}
    C_HOT: {true: [{select: "sel('A/*/56/[CG,OH]')"}, {select: "sel('A/*/115/[CG,SD]')"}, {select: "sel('A/*/123/[CD2,OH]')"}]}

The contig designs a chain of 50 to 120 residues, starts a second chain with a break, and keeps the target residues. The target backbone forms coordinate group 1, so its internal geometry is fixed while its position relative to the binder is open. C_SEQ keeps the target sequence, and C_HOT marks the side-chain atoms the binder should engage.

With the cropped structure uploaded as the target, Binder length 50 to 120, the three hotspots above, Designs per input 2 and Recycling schedule medium/front_loaded, the estimate is 59 credits.

Under Custom conditions, a CIF that already carries condition annotations, such as a design returned by an earlier run, can be submitted without a specification. Each file becomes one design problem named after the file, and the model reads its annotations as given.

Use this for monomer design and for other annotated inputs, such as the model's benchmark structures. A specification can load the same kind of file with input_as_annotated_cif: true.

Settings

Design

Sampling

The recycling schedule sets how many trunk recycling passes run across the 60 steps. Schedule names combine a level and a shape.

thinking_low, thinking_medium and thinking_high are front loaded with 120, 180 and 300 passes. The interface suggests medium/front_loaded for protein binders, medium/midpeak for small-molecule binders and enzymes, and medium/flat for inverse folding. A schedule other than none cannot be combined with a different Sampling steps or a nonzero Recycling depth, and the form refuses the combination.

Temperature

Partial diffusion

These settings appear only when Workflow is partial_diffusion. Set either the fraction or the corruption levels, not both.

Trajectories

Turn C-alpha only trajectories off whenever Save trajectory every N steps is above 0. The model version in use fails while writing C-alpha trajectories, so a run that keeps the default writes no trajectory and ends in an error.

Outputs

Viewer displays the designed structures. Files lists every download.

Understanding results

A finished run returns structures with no verdict attached. For a PD-L1 binder job, the files hold the target chain and a new designed chain, and nothing in them says whether the binder folds on its own or sticks to PD-L1. The model returns no confidence scores, and per_example_metrics.csv records settings rather than quality. A design is a candidate, and the checks that rank candidates come afterwards.

Predict the designed sequence on its own with ESMFold2 or AlphaFold2 and compare it with the designed backbone using RMSD; a design that its own sequence does not fold back into is unlikely to hold. For a binder, predict the complex with Boltz-2 and read ipTM and ipSAE at the interface. pLDDT from those predictions says how much of the designed chain the predictor trusts.

Each designed CIF records where its atoms came from in ordinary _atom_site columns. annotation_src-component_1_atom gives the source contig label of each atom, such as A17-131, and . for atoms with no source. annotation_src-coordinate-conditioned-region_1_atom is . for atoms whose coordinates were not conditioned. annotation_src-index-group_1_atom and is_unindexed_insertion mark motifs that the model placed without a fixed position in the chain.

These columns locate a motif in a design, because output chains are renumbered from 1 and a kept motif no longer sits at its input number. For enzyme design, check that the coordinate-conditioned region matches the motif you specified before reading anything into the scaffold.

The replicate index in a file name is not a rank. Design 0 is not better than design 7, and sorting by name sorts by index only. Trajectory frames are intermediate states of the sampling process, not molecular dynamics, so they show how a design formed and say nothing about motion or binding.

For de novo protein design in general, generate several designs per target and filter them with an independent predictor, as above, before choosing candidates for the bench.

Sources

• Corley N, Didi K, Kazman MW, Brent RI. "Biomolecular design through latent reasoning with RFD4-Proteina." Preprint (2026). RFD4-Proteina.pdf
• RosettaCommons. RFD4-Proteina source code and documentation, BSD 3-Clause. github.com/RosettaCommons/RFD4-Proteina