NVIDIA RFD4-Proteina: Open Protein Design Model With Reward-Guided Search, 3.8x Faster Than RFD3

NVIDIA has released a protein design model that treats the job as a search problem and spends extra computing on the designs most likely to work. In a technical tutorial published October 21, 2026, NVIDIA introduced RFD4-Proteina, a BioNeMo open model co-developed with the Institute for Protein Design at the University of Washington School of Medicine. It generates a protein's amino-acid sequence and its all-atom 3D structure together, and it can use scoring models during generation to steer toward better candidates. This post explains the approach, the numbers NVIDIA reports, a worked example against the cancer target PD-L1, and what the tutorial does not claim.

TL;DR: what NVIDIA is claiming

Why is protein design a search problem?

NVIDIA's tutorial opens with the scale of the space: an eight-residue sequence has more than 25 billion possible combinations, and a 100-residue sequence has more possibilities than there are atoms in the observable universe. Nobody searches that exhaustively. Generative models narrow the space by learning patterns from natural proteins, and the quality of a result depends on where compute is spent.

Earlier pipelines generated a backbone structure first and then used a separate inverse-folding model such as ProteinMPNN to choose a sequence that fits. RFD4-Proteina removes that handoff. Building on the partially latent approach of Proteina-Complexa, it combines explicit backbone coordinates with learned latent variables for sequence and atomic detail, and an autoencoder decoder produces the sequence and the full structure together. For related NVIDIA work on running structure-prediction models fast, see our coverage of the BioNeMo Inference Runtime with Boltz, OpenFold and Protenix.

Molecule of rounded nodes with one highlighted, illustrating a candidate binder from RFD4-Proteina protein design

What makes it faster?

NVIDIA lists four reasons. First, it works on a residue-level latent representation rather than an atom-level one, which avoids memory blow-up. Second, engineering of the diffusion process cuts sampling from about 400 steps in Proteina-Complexa to 60, with computation concentrated on the parts that matter. Third, the scoring oracles used in search are CUDA-accelerated. Fourth, the model and kernels were co-designed with the cuEquivariance and BioNeMo Inference Runtime teams.

The headline figure, 3.8x faster generation and 84% less peak reserved memory than RFD3 at 200 residues, comes from NVIDIA's reported generation benchmark. The tutorial adds a second speed-up for deployment: the BioIR backend accelerates the model by 1.37 to 1.73x over the default compiled backend across six tested Hopper GPU configurations, and on an H100 PCIe the total suite time drops from 161 to 111 seconds, about 31 percent. That benchmark uses batch size 1 and sums warm end-to-end timings across 15 sequence lengths from 128 to 1,024 tokens. Treat both as vendor measurements on NVIDIA's chosen hardware and settings.

Two different speed-up figures from NVIDIA

NVIDIA's own materials give different multiples. The technical tutorial reports 3.8x faster generation than RFD3 at 200 residues. NVIDIA's Europe open-models announcement of the same day says RFD4-Proteina "generates designs about 13x faster than RFdiffusion3 in a benchmark of 200 amino acid proteins" and delivers "leading performance across more than 10 protein design tasks." Both refer to 200-residue proteins, but the posts do not say whether the 13x figure includes the search step, a different hardware setup or a different baseline configuration, so the two numbers are not reconciled. Until the manuscript spells out the conditions, quote the tutorial's 3.8x as the figure with a stated measurement, and treat 13x as NVIDIA's headline claim.

How does reward-guided search work?

This is the part of the design that sets it apart from one-shot generators. During inference the model denoises generation trajectories into structures, scores them, retains and branches promising partial states, and continues until the criteria are met. In the PD-L1 example, the search preset uses two branches, a beam width of three, and scoring at steps 20, 40, 50 and 60. The reward is the OpenFold3 pLDDT of the designed chain, with a floor of 50, and rollouts that reach a reward of 80 are saved.

Rewards can measure geometry, interactions, or independent prediction confidence, and NVIDIA notes that physics-based scoring with TMol, a Rosetta-inspired all-atom scorer, can be used too. The idea is the same as search in language-model reasoning: spend extra compute at inference to get higher-scoring outputs. The trade-off is that you must include refolding and scoring in compute comparisons, and that the search optimizes whatever the reward says, which is only as good as the oracle.

What does the PD-L1 tutorial show?

The worked example designs protein binders against PD-L1, an oncology target. It takes the cropped PD-L1 structure from PDB entry 5O45, fixes the target's sequence and backbone, lets side chains repack, and directs binder generation toward hotspot atoms on Tyr56, Met115 and Tyr123. The new binder is 100 to 150 residues with free sequence and coordinates.

Inputs go through AtomWorks, the molecular data layer, which prepares fixed atoms, designable residues, hotspots and context. Both the open model and the NIM accept annotated CIF files or a condition specification plus its source structure. Running it takes a few commands:

• Open model: clone the RosettaCommons RFD4-Proteina repository, run its setup script on a supported Linux GPU host with Python 3.12, load nvidia/RFD4-Proteina through from_pretrained, and run a generation script. Add replicates for more candidates. The tutorial says setup caches the Hugging Face bundle, including the paired autoencoder, and that while the model repository is private you must authenticate with a Hugging Face account that has access.
• NIM: pull the nvcr.io/nim/ipd/rfd4-proteina container with an NGC API key, run it with the NVIDIA runtime, and post a JSON request to its generative endpoint. NVIDIA also lists a hosted API route through build.nvidia.com but says it is not recommended for production runs.
• Search: for the NIM, mount an OpenFold3 reward callback and enable the oracle plugin; for the open model, install the optional OpenFold3 runtime and weights, with Apptainer available.

After generation, NVIDIA advises filtering candidates on constraints, motif preservation, clashes, interfaces, refolding consistency and diversity, using metrics such as ipTM, ipSAE and RMSD from OpenFold3 or RoseTTAFold3 along with TMol interaction energies, then testing diverse candidates in the lab and feeding results back into the next round.

How do agents fit in?

The tutorial describes two skills, rfd4-proteina-campaign and rfd4-proteina-package, that let an AI agent coordinate generation, search, evaluation and archiving with a record of inputs, model versions and compute use. The suggested prompt asks an agent to design PD-L1 binders, compare plain generation with reward-guided search under an agreed compute budget, return up to ten diverse candidates with scores, and package the results. This builds on the BioNeMo Agent Toolkit we covered earlier, which connects tools like these to coding agents such as Claude, and the original toolkit launch. NVIDIA frames the human role clearly: scientists define the objectives and use experimental evidence to guide the next round.

What does the tutorial not show?

• No wet-lab results in the tutorial body. The post says it "introduces experimental validation results across design tasks," but the examples shown are computational scores. Check the accompanying manuscript for hit rates, binding affinities and how many designs were tested.
• Confidence is not binding. NVIDIA states that the pLDDT reward measures confidence in the binder, not evidence of binding, and that interface and refolding metrics should be inspected separately.
• State-of-the-art claims are NVIDIA's. The post says the manuscript shows the model is state-of-the-art across a breadth of molecular tasks in quality and throughput. We have not seen independent replication.
• Comparisons are to NVIDIA's chosen baselines and hardware. The speed numbers compare against RFD3 at one length and on Hopper GPUs.
• Access caveat. The tutorial mentions the model repository being private at the time of writing, so check current access terms before planning work, and read the license for the open model and for any NIM use.

Who should try it first?

Computational biology groups that already run structure-prediction pipelines are the natural early users, because they have the oracles and the lab partners that give search its meaning. A practical first step is to reproduce the PD-L1 example end to end, then swap in your own target and hotspot residues and compare plain generation against search at the same compute budget. Record the number of candidates that pass your filters per GPU-hour, not just the mean score, since search spends compute on scoring and refolding that a one-shot run does not. If the search path does not beat generation plus a larger sample on that measure for your target, you have learned something worth knowing before committing GPUs to it.

Why it matters

Protein design is one of the clearest places where generate-then-score loops pay off, because good scoring models exist and lab testing is expensive. A model that is faster, uses less memory, and lets you steer compute at inference makes it practical for smaller groups to run many design rounds. The pattern of open weights plus a hosted-style microservice plus agent skills mirrors how NVIDIA ships other science models, and it connects to wider AI-for-science work we have covered, including Claude designing scorpion-toxin protein binders and Google's autonomous wet-lab digital twin. The open question is the one that always follows a design benchmark: how many of the top-scoring designs bind in a real assay.

Related reading

• NVIDIA BioNeMo Inference Runtime with Boltz, OpenFold and Protenix
• NVIDIA BioNeMo Agent Toolkit with Claude for protein structure
• NVIDIA BioNeMo Agent Toolkit for drug discovery agents
• Claude and scorpion toxin protein binders
• Google Research VeriD digital twin for wet-lab agents
• NVIDIA Europe AI-in-science pledge

Details reflect NVIDIA's October 21, 2026 technical blog; benchmark figures are NVIDIA's and access terms may change.