# RFD4-Proteina Explained: Joint Protein Design and Inference-Time Search NVIDIA and the University of Washington are combining sequence and structure generation, with extra computation used to explore more promising designs. By Paras Madan | 2026-10-21 NVIDIA detailed RFD4-Proteina on October 21, 2026, a biomolecular design model co-developed with the University of Washington's Institute for Protein Design. It generates protein sequences and all-atom structures together, then can use inference-time search to explore promising candidates. Public code, model links and an NVIDIA-hosted development endpoint are available. The interesting engineering change is the connection between generation and evaluation. A model can produce many plausible candidates, but researchers still need a way to decide which ones deserve more computation and, eventually, experimental work. ## What RFD4-Proteina changes [NVIDIA's technical explanation](https://developer.nvidia.com/blog/design-proteins-with-rfd4-proteina-using-reward-guided-inference-time-search/) describes a model that combines explicit backbone coordinates with a learned latent representation. Its decoder produces the sequence and detailed structure together, avoiding a separate inverse-folding step that assigns a sequence after a backbone has been generated. Optional search scores intermediate designs, keeps promising states and branches from them. The scores can reflect geometry, interactions or prediction confidence. This lets the system spend additional computation exploring selected paths rather than treating every generated candidate equally. There are two decisions here that should remain visible to the researcher: what the model proposes and what the scoring system rewards. A search can become very effective at improving its chosen score without establishing that the score captures the entire scientific objective. ## What does reasoning mean in this model? The [research project](https://rfd4-proteina.org/) distinguishes recurrent refinement inside the model from external search guided by scoring functions. Its shared chemical and conformational representation covers more than standard amino acids, including modified amino acids, sugars and nucleic acids. Here, reasoning describes additional computational refinement and search. It should not be confused with a language model explaining a scientific argument in prose. The project reports computational benchmarks and experimental validation across several design tasks. Those results are evidence for the evaluated settings. They do not mean that an arbitrary generated molecule will behave as intended, or that a computationally attractive candidate is ready for therapeutic use. A useful way to read this release is as a proposal for allocating research computation. The practical question becomes whether the extra search produces better candidates per unit of total effort, including evaluation and laboratory work. ## How to read the speed and memory results NVIDIA reports a 200-residue generation benchmark of 1.35 seconds and 1.36 GB peak reserved memory for RFD4-Proteina, compared with 5.12 seconds and 8.50 GB for RFD3. That corresponds to approximately 3.8 times faster generation and 84% lower peak reserved memory in the reported test. These are generation results, not an end-to-end budget for a research campaign. A workflow that invokes external scoring models or evaluates many branches will have additional costs. For evaluation, teams should separate candidate-generation time, scoring time and the number of candidates that pass their own criteria. Otherwise, a faster generator can look more economical while simply moving the bottleneck into the next stage. ## Code, model access and API availability The [public repository](https://github.com/RosettaCommons/RFD4-Proteina) provides Python and command-line interfaces, released model-bundle setup and deployment documentation. Its principal environment is Linux x86_64 with Python 3.12 and a compatible NVIDIA GPU. CUDA 12 and CPU-only container options are also documented; macOS and Windows are not supported. The repository links browser-based Hugging Face Studios. Both require sign-in and screen designs with SecureDNA. These are useful access options, but a shared service should not be mistaken for dedicated production capacity. The code carries a BSD-3-Clause license. Model and service terms should be checked separately rather than assuming the code license covers every distribution route. [NVIDIA's hosted endpoint](https://build.nvidia.com/ipd/rfd4-proteina) is labeled free for development, with possible rate limits and throttling. It identifies NVIDIA API trial terms and the NVIDIA Community Model License. NVIDIA's technical guide explicitly says the hosted endpoint is not recommended for production runs. ## What researchers should compare next A sensible comparison would keep the scientific objective, evaluation criteria and total computational budget fixed. It should report failed candidates as well as successful ones, and separate computational scores from experimental outcomes. That creates a more demanding test than comparing generation throughput alone. RFD4-Proteina's useful promise is that a combined model and search process can make candidate selection more productive. The evidence that matters next is how consistently that advantage survives outside the release's selected evaluations. I'm Paras. I work at the intersection of AI, engineering and distribution. I help technical teams explain research and AI products clearly, without losing the details that make the work useful. [Book a 30-minute call](https://cal.com/paras-madan/30min).