NVIDIA’s RFD4-Proteina takes a different approach to protein design by creating sequences and all-atom structures together. The model uses reward-guided inference-time search to improve both the design process and the resources needed to run it.
That matters because protein design requires more than finding a plausible sequence. A useful design also needs a structure that fits its intended target, follows geometric rules, and reaches a strong confidence score. RFD4-Proteina brings those parts into one system instead of separating them into a sequence design step and a later structure handoff.
A single model for sequences and structures
RFD4-Proteina combines explicit backbone coordinates with a latent representation. This removes the separate inverse-folding handoff used by prior approaches, allowing the model to co-design protein sequences and all-atom structures in one process.
The model works on a residue-level latent representation rather than an atom-level representation. That choice helps prevent memory use from growing too quickly as designs become larger, which is one reason RFD4-Proteina uses less memory than RFD3.
The system also changes how it handles diffusion. Engineering the diffusion process cuts sampling from about 400 steps in Proteina-Complexa to 60. Fewer steps do not tell the whole story, since the model also concentrates computation on the parts of the process that matter most through its reasoning capabilities.
Together, these changes make the design process faster without treating structure, sequence, and evaluation as separate jobs.
Search helps refine each design
RFD4-Proteina uses inference-time search to test and improve candidate designs. The search calls CUDA-accelerated oracles, including OpenFold3 and TMol, which score designs against geometry, interaction, and confidence metrics.
OpenFold3 helps evaluate structural results, while TMol provides physics methods within the search process. Because these oracles are CUDA-accelerated, the search can examine and refine designs without adding the same delay expected from slower evaluation steps.
This approach puts more work into judging the designs that need attention instead of spending the same amount of computation on every part of generation. The result is a system that combines model reasoning with targeted evaluation during inference.
The model and its kernels were co-designed with the cuEquivariance and BioNeMo Inference Runtime, or BIR, teams. That work focuses on speed and scalability, linking the model’s design with the software needed to run its calculations.
What the benchmark shows
At 200 residues, RFD4-Proteina delivers 3.8 times faster generation than RFD3 and uses 84% less peak reserved memory in the reported generation benchmark. Those figures show the effect of combining residue-level processing, shorter diffusion sampling, targeted computation, and accelerated search.
The comparison also gives the performance numbers a clear baseline. RFD4-Proteina is not only faster than RFD3; it reaches that speed while using far less peak memory at the same 200-residue size.
That memory difference can shape which protein design tasks are practical to run. Lower peak use gives the system more room to handle its sequence and structure calculations, while faster generation allows more designs to move through the search process in the same amount of time.
From conditions to finished designs
Both the open model and the NIM microservice accept annotated CIFs or condition specifications. This gives users two ways to describe the starting information or requirements for a design.
Those conditions can support tasks such as PD-L1 binder design with hotspot constraints. In that setting, the design process can use specified interaction requirements while the search evaluates geometry, interaction, and confidence.
The same workflow connects the input conditions to co-designed sequences and structures, then uses OpenFold3 and TMol to score and refine the results. That keeps the design target and the evaluation process linked rather than passing a finished sequence into a separate stage.
RFD4-Proteina brings together several changes that address the main costs of protein generation: representation size, diffusion length, evaluation speed, and memory use. Its residue-level latent representation avoids atom-level memory growth, diffusion drops from about 400 steps in Proteina-Complexa to 60, and CUDA-accelerated oracles make inference-time search faster.
As of Oct 21, 2026, the model’s reported 200-residue benchmark points to a clear goal for protein design systems: create richer sequence and structure outputs while spending less time and memory to produce them.
Based on
