NVIDIA RFD4-Proteina Uses Reward-Guided Search to Jointly Generate Protein Sequences and Structures
RFD4-Proteina, jointly developed by NVIDIA and the Institute for Protein Design at the University of Washington School of Medicine, applies inference-time search to protein design: after the model generates candidate structures, it calls external structure predictors such as OpenFold3 for scoring, retaining and branching promising partial states for continued iteration. Simultaneously, the model merges sequence generation and all-atom structure generation into a single step, eliminating the previous handoff that required ProteinMPNN to perform inverse folding separately.
Protein design is fundamentally a combinatorial search problem. An eight-residue sequence has over 25 billion possible combinations, and a 100-residue combination count exceeds the number of atoms in the observable universe. Exhaustive search is impossible, so the key lies in knowing where to look. The answer provided by RFD4-Proteina, released by NVIDIA on October 21, 2026 (jointly developed with the Institute for Protein Design at the University of Washington School of Medicine), is to use inference-time search to focus computational resources.
The specific process of the search loop is as follows: the model denoises generation trajectories to obtain candidate structures, which are then scored by an external evaluator. Promising partial states are retained and branched for continued iteration until generation criteria are met. Evaluators include domain-specific tools such as the OpenFold3 structure predictor and the TMol physics scorer, integrated into the search loop via CUDA acceleration. Reward signals can be geometric metrics, interaction energies, or independent prediction confidence scores.
Architecturally, RFD4-Proteina implements a specific simplification: it jointly generates sequences and all-atom structures within the same autoencoder decoder, eliminating the handoff step previously required in schemes like Proteina-Complexa to call ProteinMPNN separately for inverse folding. The authors report that using residue-level latent representations (rather than atom-level) combined with compressing diffusion sampling from about 400 steps to 60 steps allows the model to generate proteins 3.8 times faster than RFD3 at a length of 200 residues, while reducing peak memory usage by 84%. These figures are vendor-reported benchmarks and have not been independently reproduced. The authors themselves note in their tutorial that OpenFold3's pLDDT score measures prediction confidence for the complex, not binding evidence itself, requiring separate inspection of interface and refolding metrics.