NVIDIA details RFD4-Proteina protein design search workflow

NVIDIA on Wednesday outlined RFD4-Proteina, an open protein-design model that co-designs sequences and all-atom structures and adds reward-guided inference-time search to focus compute on higher-scoring candidates. The company said the system is intended to help researchers generate binder designs and other molecular candidates more efficiently than exhaustive search, while keeping the evaluation loop tied to experimental testing.

What NVIDIA says the model does

In the developer blog published on Oct. 21, 2026, NVIDIA said RFD4-Proteina is co-developed with the Institute for Protein Design at the University of Washington School of Medicine and builds on Proteina-Complexa. The model combines explicit backbone coordinates with latent representations and produces sequence and all-atom structure together, which NVIDIA says removes a separate inverse-folding handoff to a model such as ProteinMPNN.

The blog describes the model as using reward-guided search during inference to denoise candidate trajectories, score them, retain promising partial states and branch them until generation criteria are met. NVIDIA said the reward can be based on geometry, interactions or independent prediction confidence, depending on the objective selected by the user.

Reported speed and memory gains

NVIDIA said RFD4-Proteina is faster and uses less memory than RFD3 because it works on a residue-level latent representation rather than atom-level and because the diffusion process was engineered to reduce sampling from about 400 steps in Proteina-Complexa to 60. In a reported 200-residue benchmark, NVIDIA said the system delivered 3.8 times faster generation and used 84% less peak reserved memory than RFD3. Those figures are developer-reported benchmark results and not universal performance guarantees.

The company also said the BioIR backend accelerates RFD4-Proteina by 1.37 to 1.73 times over the default compiled backend across six Hopper GPU configurations. On an H100 PCIe system, NVIDIA said total suite time fell from 161 seconds to 111 seconds, which it described as about a 31% reduction in runtime, using batch size 1 across 15 sequence lengths from 128 to 1,024 tokens.

How the workflow is presented

The blog gives a PD-L1 binder example in which users define what may change, what must remain fixed and which success criteria should be used. NVIDIA said both the open model and the NIM microservice accept annotated CIF files or condition specifications, and it showed a workflow that fixes the PD-L1 target while allowing a new 100- to 150-residue binder to be generated around hotspot atoms on Tyr56, Met115 and Tyr123.

For search-based generation, NVIDIA said the open model can be paired with external oracles such as OpenFold3, while the NIM service can also run reward-guided search through an HTTP interface. In the example, the company said search-time scoring used OpenFold3 pLDDT as a confidence metric for the designed complex. NVIDIA explicitly noted that this measures prediction confidence, not binding evidence, and said interface and refolding metrics should be inspected separately.

Validation and limits

Background material supplied with the release says the broader RFD4-Proteina project has wet-lab validation across several design tasks, including selective binders, glycan control and catalytic enzymes, but those results concern experimental designs rather than established clinical efficacy. The same material notes that the accompanying preprint is still listed as forthcoming, so the launch claims, developer benchmarks and project-reported experiments should be treated separately from peer-reviewed validation.

The release does not provide a single next catalyst with a verified date beyond the Oct. 21, 2026 publication of NVIDIA's blog and the same-day UW Medicine announcement referenced in the background material. No production results, market data or public commercial terms were supplied in the source material.