RFdiffusion2 Achieves 100% Success on 41 Benchmarks with Atom‑Level Protein Generation

Previous generative protein‑design model RFdiffusion relied on describing ideal active sites at the residue level and required pre‑defining catalytic residues in the sequence, which limited the searchable solution space and forced researchers to enumerate side‑chain rotamers.

In April 2025 the University of Washington Protein Design Institute released RFdiffusion2, a new diffusion model that generates protein backbones from simple chemical‑reaction descriptions. The model removes dependence on residue enumeration and sequence indexing by introducing flow matching and stochastic centering, allowing direct atom‑level backbone generation.

RFdiffusion2 improves on three fronts: (1) it discards residue‑level enumeration, (2) it supports unindexed atom‑level active‑site generation with automatic ligand‑conformation sampling, and (3) it adopts a new AME benchmark covering 41 distinct active sites (up from 16 in the previous benchmark).

Experimental evaluation on three catalytic sites showed that RFdiffusion2 successfully generated protein backbones satisfying all constraints in every one of the 41 benchmark cases, achieving a 100% success rate, whereas RFdiffusion1 succeeded in only about 39% of cases.

The tool is hosted on the HyperAI website. Users can run it with a single click: navigate to hyper.ai, open the "Tutorial" section, select "RFdiffusion2: Protein Design Tool", click "Run this tutorial", clone the repository, choose an NVIDIA GeForce RTX 4090 instance with a PyTorch image, and start the job. Resource allocation takes roughly two minutes for the initial clone.

In the demo, entering the benchmark name (e.g., active_site_indexed_atomic) and executing the run produces the final protein‑design file, denoised structures at each diffusion timestep, and the predicted trajectory file, as illustrated by the accompanying screenshots.

Tutorial and demo links are provided for direct access.