RELEASE · MODELS · #2349
NVIDIA BioNeMo releases RFD4-Proteina: reward-guided protein design model and NIM microservice
RFD4-Proteina, an open-model co-developed by NVIDIA BioNeMo and the Institute for Protein Design (UW), co-designs protein sequences and all-atom structures and uses reward-guided inference-time search to focus compute on promising candidates. The method uses a residue-level latent representation and engineered diffusion to cut sampling from ~400 to 60 steps, yielding ~3.8× faster generation and ~84% lower peak memory than RFD3 at 200 residues; the project provides an open GitHub model for development and a NIM microservice with an HTTP interface for rapid inference and automated campaigns (examples include PD-L1 binder design with hotspot constraints).
KEY POINTS
- RFD4-Proteina, an open-model co-developed by NVIDIA BioNeMo and the Institute for Protein Design (UW), co-designs protein sequences and all-atom structures and uses reward-guided inference-time search to focus compute on promising candidates.
- The method uses a residue-level latent representation and engineered diffusion to cut sampling from ~400 to 60 steps, yielding ~3.8× faster generation and ~84% lower peak memory than RFD3 at 200 residues; the project provides an open GitHub model for development and a NIM microservice with an HTTP interface for rapid inference and automated campaigns (examples include PD-L1 binder design with hotspot constraints).
- This release accelerates and memory‑optimizes AI-driven protein design while integrating reward-guided search and CUDA‑accelerated scoring oracles, enabling faster iteration and deployment for research and drug‑discovery workflows.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
NVIDIA BioNeMo and the University of Washington Institute for Protein Design released RFD4-Proteina and an accompanying NIM microservice. The model co-designs amino-acid sequences and all-atom structures, uses reward-guided inference-time search, and employs a residue-level latent representation plus engineered diffusion to cut sampling from ~400 steps to ~60. Reported benchmarking vs RFD3 at 200 residues shows ~3.8× faster generation and ~84% lower peak memory. The project provides an open model on GitHub, a NIM microservice with an HTTP interface for rapid inference and automated campaigns, and Hugging Face Spaces (including a sponsored free-access window).
WHY NOW
RFD4-Proteina reduces sampling steps, runtime, and peak memory relative to RFD3, and integrates reward‑guided search plus CUDA‑accelerated scoring oracles (OpenFold3, TMol). The combination of algorithmic efficiency, oracle integration, an HTTP microservice, and browser-accessible Spaces lowers the friction for iterative computational protein design and experimental workflow prototyping for computational biologists, ML engineers, and drug‑discovery researchers.
WHO IS AFFECTED
Directly affected: computational biologists, ML engineers, and drug‑discovery researchers who run or develop protein‑design models or pipelines; users of prior RFD models (e.g., RFD3) who may migrate for performance/memory improvements; users of Hugging Face Spaces who can try the model in‑browser (including during the sponsored free‑access window).
CONFIRMED
All statements below are explicit in the supplied sources: 1) RFD4‑Proteina is an NVIDIA BioNeMo open model co‑developed with the Institute for Protein Design at the University of Washington. 2) The model co‑designs protein sequences and all‑atom structures and removes the separate inverse‑folding handoff. 3) It uses reward‑guided inference‑time search to focus compute on promising candidates. 4) The architecture pairs a flow‑matching model with an autoencoder using an anchor‑atom coordinate plus a learned latent vector per residue (supports standard and some modified residues and other molecular contexts). 5) Residue‑level latent representation and engineered diffusion reduce sampling from about 400 steps to 60. 6) Reported performance vs RFD3 at 200 residues: ~3.8× faster generation and ~84% less peak memory. 7) Inference‑time search integrates CUDA‑accelerated oracles including OpenFold3 and TMol. 8) The project provides an open model on GitHub and a NIM microservice with an HTTP interface; skills named rfd4‑proteina‑campaign and rfd4‑proteina‑package coordinate generation, search, evaluation, and archiving. 9) The release includes example workflows such as PD‑L1 binder design with hotspot constraints. 10) Hugging Face offers RFD4‑Proteina Spaces (dedicated and ZeroGPU); NVIDIA sponsored free access to the dedicated Space from Oct 21 to Nov 4, 2026. (All items are drawn directly from the provided excerpts.)
UNCERTAIN
Missing or uncertain (not resolved by supplied excerpts): 1) Model licensing, redistribution terms, and any usage restrictions are not stated in the excerpts. 2) Details, scope, and quantitative results of the referenced "experimental validation results across design tasks" are not provided. 3) Real‑world wet‑lab validation rates, success in biological assays, or clinical relevance beyond provided examples (PD‑L1 binder example) are not described. 4) Scalability and performance outside the reported 200‑residue benchmark (e.g., for much larger constructs) are not reported. 5) Hardware, cost, and throughput characteristics for NIM microservice deployments at production scale are not specified. 6) Security, access control, and dual‑use or biosecurity mitigations are not described. 7) Any changes to access/pricing after the sponsored free‑access period are not stated. No sources in the excerpts conflict with each other; the uncertainty items reflect missing evidence.
WHAT TO WATCH
Concrete observable signals to monitor next: 1) GitHub: commits, releases, license file, and training/inference code updates in the RFD4‑Proteina repository. 2) NIM microservice: published API docs, docker/container images, and example campaign templates (rfd4‑proteina‑campaign/package). 3) Hugging Face: activity or changes to the RFD4‑Proteina Spaces, usage statistics, and any end‑of‑free‑access notices. 4) Benchmarks/publications: detailed benchmark papers or posts from NVIDIA or Institute for Protein Design describing experimental validation and broader residue‑length performance. 5) Integration status: documentation or releases confirming OpenFold3 and TMol integration or CUDA‑oracle availability and hardware requirements. 6) Community signals: forks, third‑party tutorials, and independent benchmark results posted by academic or industry groups.
WHY IT MATTERS
This release accelerates and memory‑optimizes AI-driven protein design while integrating reward-guided search and CUDA‑accelerated scoring oracles, enabling faster iteration and deployment for research and drug‑discovery workflows.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
RFD4-Proteina was co-developed by NVIDIA BioNeMo and the Institute for Protein Design at the University of Washington.
SUPPORTEDChecked 2026-10-21 · 2 supporting
The NVIDIA Developer excerpt states the model is an NVIDIA BioNeMo open model co-developed with the Institute for Protein Design at UW.
SUPPORTS Hugging FaceMEDIA · 2026-10-21The Hugging Face excerpt repeats that RFD4-Proteina is co-developed by NVIDIA and the University of Washington's Institute for Protein Design.
RFD4-Proteina co-designs protein sequences and all-atom 3D structures and uses reward-guided inference-time search to focus compute on promising candidates.
SUPPORTEDChecked 2026-10-21 · 2 supporting
NVIDIA Developer states the model co-designs sequences and all-atom structures and explicitly describes reward-guided inference-time search to focus compute on promising candidates.
SUPPORTS Hugging FaceMEDIA · 2026-10-21Hugging Face describes the model bringing sequence and all-atom structure generation into one model (co-design); it does not emphasize reward-guided search in the excerpt.
Residue-level latent representation and engineered diffusion reduce sampling from about 400 steps to 60, yielding roughly 3.8× faster generation and ~84% lower peak memory than RFD3 at 200 residues.
SUPPORTEDChecked 2026-10-21 · 1 supporting
The model pairs a flow‑matching model with an autoencoder that represents each residue or nucleotide as an anchor‑atom coordinate (for amino acids, the alpha carbon) plus a learned latent vector; the decoder expands these into all‑atom structures and sequences.
SUPPORTEDChecked 2026-10-21 · 1 supporting
SOURCES & TIMELINE
2RFD4-Proteina co-designs protein sequences and all-atom structures using reward-guided inference-time search to focus compute on promising candidates. The model combines explicit backbone coordinates with latent representation, removing the separate inverse-folding handoff required by prior approaches. Residue-level latent representation and engineered diffusion reduce sampling from about 400 steps to 60, deliverin…
Protein design starts with setting a goal: a target to bind, a molecule to recognize, or an arrangement of atoms to preserve. Turning that goal into a design has usually meant splitting up the workflow across models that each handle one modality or one stage and reassembling the pieces afterward. Information is lost at every handoff. RFD4-Proteina is a BioNeMo open model for programmable protein design, co-developed…