AI BriefingNVIDIA and University of Washington Release RFD4-Proteina on Hugging Face Spaces
·2026.10.21 21:39
Key point
The unified flow-matching model outperforms Proteina-Complexa by 5.8x in design yield while integrating sequence and all-atom structure generation.
Details
- NVIDIA and the University of Washington's Institute for Protein Design released RFD4-Proteina, an open BioNeMo model available on Hugging Face Spaces.
- In benchmarks across 24 protein binding targets, the model with high thinking settings produced 76.40 structurally distinct designs meeting computational criteria per GPU-hour.
- This performance significantly exceeds the 13.24 designs per GPU-hour achieved by the second-strongest method, Proteina-Complexa.
- The model integrates sequence and all-atom structure generation into a single flow-matching architecture to reduce information loss during handoffs.
Technical Architecture
- Combines a flow-matching model with an autoencoder that encodes amino acids, nucleotides, and chemical identities into latent vectors.
- Supports generation in complex molecular environments including other proteins, small molecules, nucleic acids, and modified amino acids.
- Offers a 'Thinking setting' (low, medium, high) to balance design refinement quality against computational cost.
Availability and Workflow
- Free access to a dedicated Space on NVIDIA RTX PRO 6000 GPUs is sponsored from October 21, 2026, to November 4, 2026.
- Post-sponsorship, the dedicated Space becomes paid, but a shared-capacity ZeroGPU Space remains free within Hugging Face account limits.
- Supports programmatic generation via command line, Python scripts, and the Diffusers library, requiring Linux x86_64, Python 3.12, and CUDA 13 compatible drivers.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.