How RFdiffusion2 Revolutionizes Protein Design with Sequence‑Independent Active Sites

Limitations of the original RFdiffusion model

RFdiffusion generated protein backbones by specifying ideal active sites at the residue level. This imposed two major constraints:

Active‑site geometry could only be defined per residue, requiring exhaustive enumeration of side‑chain rotamers to explore possible backbone conformations.

Catalytic residues had to be pre‑assigned in the sequence, drastically limiting the searchable design space.

RFdiffusion2: New deep generative approach

Released in April 2025 by the Protein Design Institute (University of Washington), RFdiffusion2 removes the need for residue enumeration and sequence indexing. Its key technical innovations are:

Flow Matching combined with Stochastic Centering enables direct atom‑level backbone generation, bypassing residue‑wise constraints.

Support for unindexed atom‑level active‑site generation , allowing the model to ingest ligand information and automatically sample ligand conformations during diffusion.

A new AME benchmark that expands the evaluation set to 41 diverse active sites (up from 16), providing a more comprehensive assessment of catalytic‑site design performance.

Experimental validation

The authors evaluated RFdiffusion2 on three distinct catalytic sites across the 41‑case AME benchmark. Results showed:

RFdiffusion2 successfully generated protein backbones satisfying all geometric and functional constraints in 100 % of cases .

The predecessor, RFdiffusion1, succeeded in only ~ 39 % of the same cases.

These findings demonstrate a substantial improvement in AI‑driven protein design, particularly for tasks requiring precise active‑site placement without prior residue specification.

Typical usage example

In a command‑line or notebook environment, the benchmark can be invoked as follows: python run_benchmark.py --task benchmarks This command generates the full set of designed protein structures, intermediate denoised conformations at each diffusion timestep, and the final structure‑prediction trajectory.