Design Proteins with RFD4-Proteina Using Reward-Guided Inference-Time Search

NVIDIA and UW Release RFD4-Proteina to Accelerate Protein Design

NVIDIA and the University of Washington’s Institute for Protein Design have co-developed RFD4-Proteina, a new open model designed to navigate the immense search space of protein engineering. The model co-designs protein sequences and their corresponding all-atom structures simultaneously, employing a reward-guided inference-time search to intelligently allocate computational resources to the most viable candidates. This approach is critical for complex tasks like designing binders for oncology targets such as PD-L1, moving the field closer to more efficient therapeutic discovery.

Technical Advancements and Performance

RFD4-Proteina introduces significant architectural and performance improvements over previous iterations. By combining explicit backbone coordinates with a latent representation, it eliminates the need for a separate inverse-folding handoff, streamlining the design process. This, along with an engineered diffusion process, allows the model to achieve substantial efficiency gains and integrate with external validation tools for more refined results.

• Performance Boost: Delivers 3.8x faster generation and uses 84% less peak memory than its predecessor, RFD3, for a 200-residue protein.
• Reduced Sampling: The diffusion process is cut from approximately 400 steps down to just 60, concentrating compute where it matters most.
• Integrated Oracles: The inference-time search loop incorporates CUDA-accelerated oracles, including OpenFold3 and TMol, to score and refine designs against geometry and interaction metrics in real time.
• Dual-Access Model: It is available as an open model on GitHub for research and customization, and as an NVIDIA NIM microservice for rapid, scalable deployment via an HTTP interface.

Ecosystem Integration and Market Impact

The dual-release strategy for RFD4-Proteina addresses the needs of the entire computational biology community, from academic researchers to enterprise drug discovery labs. The open model provides transparency and control for method development, while the NIM microservice offers a production-ready, containerized solution that abstracts away infrastructure complexity. This release is further embedded within NVIDIA's broader strategy, with optimizations from the BioNeMo Inference Runtime (BioIR) and hooks for automation via the BioNeMo Agent Toolkit, signaling a push towards creating a comprehensive, end-to-end platform for AI-driven scientific research.

By offering RFD4-Proteina as both a customizable open model and a scalable NIM microservice, NVIDIA is building an integrated ecosystem that captures both the research and enterprise segments of the high-value computational biology market, solidifying its platform as the foundational stack for AI in drug discovery.