Use a fresh Python 3.12 environment and an NVIDIA GPU with a driver compatible with CUDA 12 or CUDA 13.
Install from PyPI (recommended):
pip install "rfd4-proteina[all-cuda13]" -Ugit clone https://github.com/RosettaCommons/RFD4-Proteina.git
cd RFD4-Proteina
pip install -e ".[all-cuda13]"These commands use CUDA 13. For CUDA 12, use all-cuda12 with the matching PyTorch wheels; see the CUDA 12 installation instructions.
pip install diffusers \
git+https://huggingface.co/nvidia/RFD4-Proteina-prepStart with an annotated CIF, or convert a YAML/JSON condition spec to an annotated CIF first.
import torch
from diffusers import RFD4ProteinaPipeline
from rfd4_proteina_prep import prepare_cif, write_cif
pipe = RFD4ProteinaPipeline.from_pretrained(
"nvidia/RFD4-Proteina-Diffusers",
dtype=torch.bfloat16,
).to("cuda:0")
features, atoms = prepare_cif(
"input.cif", seed=42, device=pipe.device,
)required = [
"padding_mask", "is_in_context_conditioning",
"supports_local_latent", "token_type",
"atom37_mask", "coords_nm",
]
output = pipe(
**{name: features[name] for name in required},
conditioning_features={
name: value for name, value in features.items()
if name not in required
},
generator=torch.Generator("cuda:0").manual_seed(42),
)
write_cif(output, features, atoms, "design.cif")Continue in the same Python session. See the Diffusers quickstart for installation details and an example input.
We present RFD4-Proteina, a biomolecular design model that reasons in a shared chemical and conformational latent space covering standard and modified amino acids, sugars, DNA, and RNA. The model uses recurrent refinement to adaptively allocate computation and inference-time search guided by geometric and energy evaluations. Below are the model and its in silico benchmarks, followed by wet-lab validation.
Biomolecular design requires finding chemically and geometrically compatible arrangements of interacting atoms and molecules. The complexity of problems which can be tackled by deep learning design models is limited by both the molecular landscape encompassed by the model and the ability to explore that landscape effectively. Here we introduce RFD4-Proteina, which addresses these limitations through intrinsic and extrinsic reasoning in a shared, physics-informed chemical and conformational latent space. The latent space encompasses standard and modified amino acids, sugars, DNA, and RNA, supervised with force-field energies to favor physically plausible conformations. Within this space, intrinsic reasoning uses recurrent refinement to adaptively allocate computation, while extrinsic reasoning guides search with arbitrary constraints such as energetics or geometric criteria. Across in silico protein design benchmarks, RFD4-Proteina achieves higher design success rates at greater generation throughput than all methods evaluated, with further gains from conformational flexibility, recurrent refinement, and search. Without additional training, the same model performs equivalently to or better than state-of-the-art specialist methods for related tasks, including side-chain packing, fixed-backbone sequence design, conformational sampling, and design refinement. Experimental validation spans specific and cross-reactive binding to therapeutically relevant targets, modified and glycosylated proteins, and catalysis of both bond-breaking and bond-forming reactions, including a first proof-of-concept demonstration of a de novo designed enzyme for active pharmaceutical ingredient synthesis. By uniting diverse molecular chemistry, conformational flexibility, scalable reasoning, and physical modeling with machine-learning generation, RFD4-Proteina expands the frontier of protein design.
RFD4-Proteina makes use of a unified latent space. Select from the tabs below to explore the latent space, a full generation trajectory, and intrinsic and extrinsic reasoning.
An autoencoder represents amino acids, sugars, and nucleotides in a shared 32-dimensional chemical and conformational space.
Full size ↗
An enzyme-design trajectory for cleavage of a peptide–DNA linkage.
Full size ↗60 denoising steps · 120, 180, and 300 trunk evaluations
Trunk evaluations per step
Current low, medium, and high thinking presets allocate more recurrent updates to earlier steps. Medium is the default. Recycling schedule guide.
Different molecular properties become resolved at different points in the generation trajectory. In this example, we select branches first by interface area, which becomes resolved early in generation, followed by hydrogen bonding as finer-scale contacts emerge, and finally by AlphaFold3 (AF3) confidence.
Full size ↗Typically, design methods trade accuracy for speed. For example, hallucination methods such as BindCraft achieve strong pass rates but only produce tens of designs per GPU-hour. Generative diffusion models, meanwhile, produce hundreds to thousands of designs per GPU-hour but exhibit lower per-design quality. On all four benchmarks, RFD4-Proteina achieves both higher pass rates and higher throughput than every other method evaluated, even at its lowest thinking setting.
For detailed results, evaluation criteria, and method comparisons, explore the full benchmark section.
Figure 2. Clustered pass rate against generation throughput (designs per GPU-hour) for protein binding (24 targets), small-molecule binding (77), enzyme active-site scaffolding (37), and DNA binding (3). Right-hand bars show target coverage and passing clusters per GPU-hour including AF3 evaluation. Low, medium, and high thinking denote task-specific recurrent-refinement budgets without structure-predictor feedback.
The structures below are computational models of designs we tested in the lab. Drag to rotate, scroll to zoom, and use the buttons to switch designs or states.
Calmodulin switches conformation when it binds Ca2+, and many of its partners recognize only one state. We asked RFD4-Proteina for binders to the calcium-bound form only.
Figure 4a. Calcium-dependent calmodulin binding: the design concept, the distribution of apparent KD across designs, and the paper’s 69 pM example* with BLI responses in Ca2+ (teal) and EDTA (gray). The interactive viewer above shows a different binder. Full figure
* The paper’s 69 pM apparent affinity comes from a 1:1 fit at the off-rate fitting floor; its precise affinity and ranking remain unresolved.
A binder to a peptide–MHC complex must recognize a few exposed peptide residues while ignoring the conserved MHC around them. Cancerous and infected cells are flagged this way, which is why look-alike peptides from healthy cells are the main risk.
Figure 4b. Binders to MART-1, SARS-CoV-1, and Phox2B peptide–MHC complexes: design models, yeast-display titrations, and binding to the six peptide–MHC targets in the specificity screen. Full figure
Intrinsically disordered regions are involved in neurodegeneration, chronic pain, and cancer, but they have no fixed structure to design against. We targeted eight disordered peptides: five known to fold when bound and three without that tendency.
Figure 4c,d. Binders to four disordered peptides (YAP1, Tau153, N17, PER2) with SPR sensorgrams and 1:1 fits (c), and all-by-all SPR specificity of the best binders against five peptides (d). Full figure
Trk receptors are targets in neuropathic pain and in cancers such as neuroblastoma. Human TrkA is an established benchmark, so we asked for more: binders that avoid mouse TrkA (86% sequence identity over the modeled region) and the human paralogues TrkB and TrkC.
Figure 4e,f. Human-specific and human/mouse cross-reactive TrkA campaigns with SPR and BLI measurements against human and mouse TrkA (e), and the experimental hit rate of unconstrained human TrkA design compared with other campaigns (f). Full figure
N-linked glycans are trimmed by enzymes in the ER and Golgi. Blocking that processing triggers an immune response to the underprocessed glycan, which makes it usable as a vaccine adjuvant. A design has to shield the core GlcNAc while keeping the terminal mannose accessible.
Figure 5a–c. Shielding the core of an N-linked glycan while exposing the terminal mannose (a), design B10 with its glycan composition (b), and the extent of underprocessing across design campaigns (c). Full figure
Phosphorylation regulates molecular interactions across all domains of life. Bauer et al. demonstrated the design of phosphorylation-dependent binders to phosphotyrosine-containing peptides. Here, we targeted three phosphorylated variants of the histone H3 tail: pS10, pT11, and pS28.
Figure 5d–f. Binders to histone H3 pS10 (d) and pT11 (e) with yeast-display titrations of the phosphorylated and unmodified peptides, and selectivity between adjacent phosphorylation sites by NGS recovery (f). Full figure
Paraoxon is an organophosphate pesticide and a potent neurotoxin. RFD3 demonstrated the design of active phosphotriesterases by scaffolding six catalytic residues around two metal ions, achieving kcat/KM up to 4 ± 3 M−1 s−1. Building on this work, we used RFD4-Proteina to scaffold an expanded, eight-residue active site, testing whether additional catalytic interactions could improve activity.
Figure 6a–d. Paraoxon hydrolysis (a), the eight-residue active-site theozyme (b), in silico success rates for RFD3 and RFD4-Proteina (c), and four designs with their backbone and motif RMSDs and catalytic efficiencies (d). Full figure
Lenacapavir is an HIV-prevention drug whose manufacturing cost limits access. One step joins an activated ester to the (S)-enantiomer of a hindered chiral amine, so the enzyme has to favor aminolysis over hydrolysis of the ester.
Figure 6h–n. The lenacapavir amide-forming step (h), the ligase input motif (i), in silico success rates (j), designs 52 and 58 (k), HPLC traces (l), conversion by parent designs and catalytic-Cys knockouts (m), and conversion of the (S)- and (R)-amine (n). Full figure
@article{corley2026rfd4proteina,
title = {Biomolecular design through latent reasoning with {RFD4-Proteina}},
author = {Corley, Nathaniel and Didi, Kieran and Kazman, Maxwell W. and Brent, Rafael I. and others},
year = {2026},
note = {Preprint},
url = {https://rfd4-proteina.org/assets/paper/RFD4-Proteina.pdf}
}