Design Proteins with RFD4-Proteina Using Reward-Guided Inference-Time Search

NVIDIA has published RFD4-Proteina, an open model in its BioNeMo line developed with the Institute for Protein Design at the University of Washington's School of Medicine, along with a tutorial showing how to design protein binders against PD-L1, an oncology target. The model co-designs protein sequences and all-atom structures and adds reward-guided inference-time search so compute is directed at the most promising candidates. Two access paths are offered: the open model on GitHub, intended for training, inference, method development and direct control over sampling code, and a NIM microservice that exposes inference over an HTTP interface.
The design problem is combinatorial. An eight-residue sequence already allows more than 25 billion amino acid combinations, and at 100 residues the space exceeds the number of atoms in the observable universe, so exhaustive search is impossible and the method depends on knowing where to look. RFD4-Proteina builds on Proteina-Complexa's partially latent approach, combining explicit backbone coordinates with a latent representation plus learned sequence and atomistic variables. Its autoencoder decoder emits the sequence and all-atom structure together, which removes the separate inverse-folding handoff to a model such as ProteinMPNN that earlier pipelines required.
The speed and memory gains come from three changes. The model operates on a residue-level latent representation rather than atom-level, avoiding memory blow-up. Engineering of the diffusion process cuts sampling from roughly 400 steps in Proteina-Complexa to 60. And the oracles used in search, including OpenFold3 and physics-based methods such as TMol, are CUDA-accelerated. The model and kernels were co-designed with the cuEquivariance and BioNeMo Inference Runtime teams. In the reported generation benchmark at 200 residues, RFD4-Proteina delivers 3.8x faster generation and 84% lower peak reserved memory than RFD3. A separate benchmark shows the BioIR backend running 1.37 to 1.73 times faster than the default compiled backend across six tested Hopper GPU configurations; on H100 PCIe, total suite time falls from 161 to 111 seconds, about 31% less, with batch size 1 and summed warm end-to-end timings over 15 sequence lengths from 128 to 1,024 tokens.
Search is the centerpiece. During inference-time reasoning search, the model denoises generation trajectories to obtain structures, scores those designs, keeps and branches promising partial states, and continues until generation criteria are met. Rewards can measure geometry, interactions or independent prediction confidence, so extra search optimizes against a stated objective. The tutorial's example fixes a cropped PD-L1 target from PDB 5O45, holds its sequence and backbone fixed while allowing side chains to repack, steers binder generation toward hotspot atoms on Tyr56, Met115 and Tyr123, and asks for a 100-150 residue binder with free sequence and coordinates. Inputs are prepared with AtomWorks, the molecular data layer for fixed atoms, designable residues, hotspots and context; both the open model and NIM accept annotated CIFs or a condition specification paired with its source structure.
On the search preset, the demo uses two branches, a beam width of three, and scoring at steps 20, 40, 50 and 60. It rewards OpenFold3 pLDDT for the designed chain with a floor of 50 and saves qualifying rollouts at reward 80. The tutorial is careful to note this measures confidence in the binder, not evidence of binding; interface and refolding metrics must be inspected separately. The NIM path exposes the same knobs in a request payload, including step checkpoints, branch count, beam width, weighted reward metrics and clip values, and the OpenFold3 reward callback is mounted into the container as a plugin enabled with an environment flag.
The open model requires a supported Linux GPU host and Python 3.12, with setup via a script that caches a Hugging Face bundle containing the paired autoencoder; the repository was private at publication, so access needs an authenticated Hugging Face account. Self-hosting the NIM needs x86_64 or arm64 Linux, Docker, the NVIDIA driver and Container Toolkit, and an NGC API key; a hosted endpoint on build.nvidia.com is offered but described as not recommended for production runs. Generation is driven from a short Python script calling from_pretrained on nvidia/RFD4-Proteina, with a seed, an optional num_replicates for multiple candidates, and a revision argument for reproducibility. Rating candidates relies on metrics such as ipTM, ipSAE and RMSD from predictors like OpenFold3 and RoseTTAFold3, plus interaction energies from TMol.
NVIDIA also ships agent skills for running campaigns. The rfd4-proteina-campaign skill coordinates generation, search and evaluation while recording inputs, model versions and compute use, and rfd4-proteina-package produces a results archive with structures, metrics and a filtering report. The suggested loop is to define objectives and constraints, generate and rank diverse candidates, test them experimentally, then feed outcomes back into the next round. The accompanying manuscript is said to show state-of-the-art quality and throughput across a broad range of molecular tasks.
This suggests the practical shift is less about a single better model than about search becoming a first-class part of inference: teams can spend extra compute at generation time against a chosen objective instead of accepting a fixed number of samples. A likely trade-off is that reward-guided search multiplies cost per design, and the tutorial itself advises including refolding and scoring in compute comparisons and inspecting outputs, since retained design counts vary. Another caveat is that confidence scores from a refolding oracle are not binding evidence, so wet-lab validation remains the gate.
Why it matters: Computational biology and ML engineering teams get a drop-in set of model, container and agent tooling for structure-based binder design, with large memory and latency reductions that make iterative campaigns on a single GPU more feasible (an inference from the stated benchmarks). Product teams building on BioNeMo can wrap the NIM's HTTP interface rather than manage sampling code, while groups needing custom search behavior will likely prefer the open model. Expect the value to hinge on how well oracle rewards correlate with experimental success.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Protein design is a search problem. Choose amino acids for an eight-residue protein sequence and more than 25 billion combinations are possible. Choose amino...