Overview

RFD4-Proteina is a generative protein design model developed by the Baker Laboratory in collaboration with NVIDIA. Where a structure-prediction model like OpenFold3 answers “what does this sequence fold into?”, RFD4-Proteina answers the inverse question: given what a design must satisfy (a backbone length, a motif to scaffold, a target to bind, a framework to graft a new loop onto), it generates a novel protein backbone and sequence that meets it. It combines RFdiffusion-style backbone diffusion with a learned sequence decoder, so a single request returns both a 3D structure and the amino-acid sequence that folds into it.

The NVIDIA NIM for RFD4-Proteina wraps this model behind a single synchronous HTTP endpoint. It can:

• Generate protein backbones and sequences across a wide range of designs, from de novo generation to conditioned scaffolding, binder design, and antibody CDR redesign; refer to Supported Designs for the full list.

• Accept the design in either of two input dialects: a single self-contained annotated mmCIF file, or a condition spec (a structured YAML/JSON document plus the structure files it references).

• Run reward-guided beam search over the generation trajectory, so a request can search for designs that score well against a weighted combination of metrics rather than returning one unguided sample, optionally including an independent structure-prediction oracle’s confidence as part of that reward, scored during generation rather than only after.

• Report the same oracle’s agreement with each finished design as eleven in-silico pass/fail criteria. This NIM ships with no oracle built in; refer to Oracles for both roles.

• Report full reproducibility metadata with every response (checkpoint identity, execution backend, compute dtype, and the exact sampling schedule that ran), so results are comparable across requests and deployments.

Supported Designs

The RFD4-Proteina NIM generates the following designs. They also combine: coordinate, sequence, motif, hotspot, secondary-structure and molecular-context conditions all compose in a single input, so a binder that also preserves a structural motif, for example, is expressed by supplying both at once.

Design	Description
De novo monomer	A new backbone and sequence of the length you ask for, generated from noise with nothing held fixed.
Protein binder	A new protein designed to bind a given target, with the target held fixed and its epitope marked as hotspots.
Peptide-MHC binder	A binder designed against a peptide presented by an MHC complex.
DNA binder	A protein designed to bind a given DNA duplex, with the duplex held by coordinate and sequence.
Nucleic-acid design	Nucleic-acid chains designed under secondary-structure and base-pair conditioning.
Structure prediction	The fold for a sequence you supply, generated without holding any coordinates.
Motif scaffolding	A scaffold built around one or more motif segments held at their given chain positions.
Enzyme active site	A scaffold around a catalytic site whose residues the model places itself, rather than holding them to fixed indices.
Tip-atom scaffolding	The same, where no ligand is present.
Small-molecule binder	A protein designed around a small molecule held at its given pose.
Ligand pocket design	A binding pocket designed around a ligand, preserving any pocket residues you hold.
Antibody CDR design	New CDR loops on a fixed antibody framework, optionally against a fixed antigen.
Nanobody CDR design	New CDR loops on a single-domain framework.
TCR CDR design	New CDR loops on a T-cell receptor framework.
Design refinement	A variation on an existing design, produced by restarting generation partway along the schedule.
Design diversification	A broader spread of variations on a starting structure, from an earlier restart point.
Sequence design	A new sequence and side chains for a backbone you supply, with the backbone held fixed.

Supported Tasks

Task	What it does
generation	Standard generation: generated regions start from noise while any supplied conditions are preserved. Covers de novo monomers and conditional designs alike, including protein and DNA binders.
motif_scaffolding	Build a scaffold around motif segments held at their given chain positions.
enzyme_design	Scaffold an unindexed catalytic site, letting the model choose where each residue lands.
small_molecule_binder	Design a protein around an atomized small molecule held at its given pose.
antibody_design	Redesign CDR loops on a fixed antibody (or TCR) framework, optionally against a fixed antigen.
partial_diffusion	Refine an existing design by restarting the trajectory partway along the schedule.
inverse_folding	Hold an input backbone fixed and resample the sequence and side chains for it.

Note

task selects how the sampler runs the trajectory, not which design problem is being solved. Several problems need no task of their own, because they are expressed by what the input holds fixed: a de novo protein binder against a target, for example, is generation with the target’s atoms held by coordinate and sequence conditioning and its epitope marked as hotspots – the binder is what the model fills in around them. Read the task table as the set of sampler modes, and the conditioning carried in the input as what actually states the problem.

Tooling

The RFD4-Proteina NIM can run the reference (OSS) sampler or an NVIDIA-accelerated trunk built on the BioNeMo Inference Runtime (BioIR), selectable per deployment:

• oss_eager: the upstream reference network, uncompiled.

• oss_torch_compile: the same network under torch.compile, padded to fixed token buckets.

• bioir: an accelerated trunk with its own fused kernels and CUDA-graph capture (the default execution backend).

Every stage besides the trunk evaluation (featurization, the sampling rollout, structure export, and metrics) is identical across the three backends, so results are comparable in kind, though not bit-identical: compilation and BioIR’s kernels both change floating-point numerics. Refer to Performance for measured comparisons between backends.

Citation:

If you use RFdiffusion, the backbone-generation approach RFD4-Proteina builds on, cite:

@article{Watson2023,
  author  = {Watson, Joseph L. and Juergens, David and Bennett, Nathaniel R. and Trippe, Brian L. and Yim, Jason and Eisenach, Helen E. and Ahern, Woody and Borst, Andrew J. and Ragotte, Robert J. and Milles, Lukas F. and Wicky, Basile I. M. and Hanikel, Nikita and Pellock, Samuel J. and Courbet, Alexis and Sheffler, William and Wang, Jue and Venkatesh, Preetham and Sappington, Isaac and Torres, Susana Vázquez and Lauko, Anna and De Bortoli, Valentin and Mathieu, Emile and Ovchinnikov, Sergey and Barzilay, Regina and Jaakkola, Tommi S. and DiMaio, Frank and Baek, Minkyung and Baker, David},
  journal = {Nature},
  title   = {De novo design of protein structure and function with RFdiffusion},
  year    = {2023},
  volume  = {620},
  pages   = {1089--1100},
  doi     = {10.1038/s41586-023-06415-8}
}

Advantages of NIMs

NIMs offer a simple and easy-to-deploy route for self-hosted AI applications. Two major advantages that NIMs offer for system administrators and developers are:

1. Increased productivity: NIMs allow developers to build generative AI applications quickly, in minutes rather than weeks, by providing a standardized way to add AI capabilities to their applications.

2. Simplified deployment: NIMs provide containers that can be easily deployed on various platforms, including clouds, data centers, or workstations, making it convenient for developers to test and deploy their applications.

The RFD4-Proteina NIM provides a generative protein design model behind a consistent API. As part of the broader NVIDIA NIM Ecosystem, RFD4-Proteina can be used alongside other NIMs (for example, refolding a generated design with an OpenFold3 or Boltz-2 NIM as an independent structural check) to build pipelines for protein engineering, enzyme design, and biologic discovery.

You are responsible for ensuring that your use of NVIDIA AI Foundation Models complies with all applicable laws.