Atom level enzyme active site scaffolding using RFdiffusion2

Woody Ahern, Jason Yim, Doug Tischer, Saman Salike, Seth M. Woodbury, Donghyo Kim, Indrek Kalvet, Yakov Kipnis, Brian Coventry, Han Altae-Tran, Magnus S. Bauer, Regina Barzilay, Tommi Jaakkola, Rohith Krishna, David BakerbioRxiv 2025 · DOI

Abstract

De novo enzyme design traditionally begins with "theozymes"—idealized arrangements of catalytic residues around a reaction's transition state. However, translating these atomic-level descriptions into stable protein scaffolds has been hampered by the need to pre-calculate backbone geometries and pre-specify residue positions in a sequence. We present RFdiffusion2, a generative AI method that enables enzyme design directly from sequence-agnostic, atomic-resolution descriptions of functional groups. By eliminating the requirement for inverse rotamer generation and fixed sequence indexing, RFdiffusion2 achieved a 100% success rate on a diverse 41-case in silico benchmark, significantly outperforming previous state-of-the-art methods (39%). Experimental validation across three diverse catalytic sites—including a zinc metallohydrolase with a $k_{cat}/K_M$ of $53,000 \, M^{-1}s^{-1}$—demonstrates that RFdiffusion2 can produce highly active catalysts with remarkably high hit rates, typically requiring the screening of fewer than 96 sequences.

Introduction

Nature employs enzymes to catalyze complex chemical transformations with unparalleled precision and efficiency. For decades, computational biologists have sought to replicate this capability through de novo design. The standard workflow involves defining a theoretical active site (a theozyme) and then searching for or building a protein structure capable of supporting that site.

Despite recent advances in deep learning for protein design, two primary bottlenecks have persisted. First, the Backbone Bottleneck: most models require a pre-defined protein backbone, forcing designers to use "inverse rotamer sampling" to guess where a backbone might exist to support specific side-chain atoms. Second, the Sequence Bottleneck: designers typically must specify the exact sequence positions of catalytic residues, which severely limits the structural diversity the model can explore. RFdiffusion2 addresses these challenges by shifting the generative process to the atomic level.

Methodology: Atomic-Resolution Generative Design

RFdiffusion2 evolves the RosettaFold Diffusion architecture to handle the complexities of enzyme active sites without the constraints of traditional scaffolding methods.

Atomic Motif Conditioning

Unlike its predecessors, which often scaffolded entire residues or required backbone coordinates, RFdiffusion2 can be conditioned on "atomic motifs." This allows the model to take the specific coordinates of functional group atoms as input. The model then simultaneously predicts the optimal side-chain rotamers, the corresponding backbone, and the surrounding protein fold that best stabilizes the site.

Sequence-Agnostic Scaffolding

One of the most significant shifts in RFdiffusion2 is the move to "unindexed" residues. By providing the model with the coordinates of active site atoms without assigning them to specific positions in the sequence (e.g., Residue 50 or 120), the model is free to determine the most evolutionarily and physically plausible sequence placement. This flexibility allows the model to discover novel folds that human experts might overlook.

Technical Innovations

  • Flow Matching: The authors transitioned from standard diffusion to a flow matching framework, providing a more stable training objective and allowing the model to learn effectively from randomly initialized weights.
  • Controllable Burial (RASA): To ensure the active site is appropriately positioned for substrate access, the model allows designers to specify the Relative Accessible Surface Area (RASA), controlling whether a site is buried in the core or exposed on the surface.
  • The ORI Token: A specialized pseudo-atom (the ORI token) was introduced to provide a "prior belief" regarding the center of the protein relative to the ligand, helping the model orient the active site within the scaffold's geometry.

Results and Discussion

In Silico Benchmarking

To evaluate the model, the authors introduced the Atomic Motif Enzyme (AME) benchmark, featuring 41 diverse active sites across five Enzyme Commission (EC) classes. This benchmark represents a significant increase in complexity over previous standards. RFdiffusion2 successfully generated valid scaffolds for all 41 cases (100%), whereas the original RFdiffusion model only succeeded in 16 of 41 cases (39%). Notably, the model's performance remained robust even as the complexity of the active site—measured by the number of disconnected sequence segments or "residue islands"—increased.

Experimental Validation

The true test of de novo design lies in laboratory activity. RFdiffusion2 designs were tested for four distinct reactions, yielding high hit rates and impressive kinetics:

  • Zinc Metallohydrolase: Designed for 4MU-Phenylacetate, these enzymes achieved a $k_{cat}/K_M$ of $53,000 \, M^{-1}s^{-1}$, representing an improvement of several orders of magnitude over previous de novo zinc-based designs.
  • Cysteine Hydrolase: These designs achieved a $k_{cat}/K_M$ of $250 \, M^{-1}s^{-1}$.
  • Retroaldolase: The team identified four active variants from a small library.

In all experimental cases, active catalysts were identified by screening fewer than 96 sequences, a testament to the model's precision and the high quality of the generated scaffolds.

Conclusion

RFdiffusion2 marks a paradigm shift in enzyme design by moving from residue-level approximations to atomic-level precision. By automating the placement of functional groups and the generation of supporting backbones, it removes the need for expert-driven "inverse rotamer" heuristics. While these designed enzymes do not yet match the hyper-efficiency of natural enzymes—likely due to the simplicity of current theozymes—RFdiffusion2 provides a powerful foundation for the future of synthetic biology and chemical engineering. Future iterations focusing on the co-design of second-shell interactions and electrostatic environments may soon close the gap between human-made and natural catalysts.

Summary updated

Abstract

Abstract De novo enzyme design starts from ideal active site descriptions consisting of constellations of catalytic residue functional groups around reaction transition state(s), and seeks to generate protein structures that can accurately hold the site in place. Highly active enzymes have been designed starting from such descriptions using the generative AI method RFdiffusion [1–3], but there are two current methodological limitations. First, the geometry of the active site can only be specified at the residue level, so for each catalytic residue functional group placed around the reaction transition state, the possible locations of the residue backbone must be enumerated by building side chain rotamers back from the functional group. Second, the location of the catalytic residues along the sequence must be specified in advance, which considerably limits the space of solutions which can be sampled. Here we describe a new deep generative method, Rosetta Fold diffusion 2 (RFdiffusion2), that solves both problems, enabling enzymes to be designed from sequence agnostic descriptions of functional group locations without inverse rotamer generation. We first evaluate RFdiffusion2 on an in silico enzyme design benchmark of 41 diverse active sites and find that it is able to successfully build proteins scaffolding all 41 sites, compared to 16/41 with prior state-of-the-art deep learning methods. Next, we design enzymes around three diverse catalytic sites and characterize the designs experimentally; in each case we identify active catalysts in testing less than 96 sequences. RFdiffusion2 demonstrates the potential of atomic resolution generative models for the design of de novo enzymes directly from their reaction mechanisms.
BibTeX
@misc{ahern2025atom,
  author = {Woody Ahern and Jason Yim and Doug Tischer and Saman Salike and Seth M. Woodbury and Donghyo Kim and Indrek Kalvet and Yakov Kipnis and Brian Coventry and Han Altae-Tran and Magnus S. Bauer and Regina Barzilay and Tommi Jaakkola and Rohith Krishna and David Baker},
  title = {{Atom level enzyme active site scaffolding using RFdiffusion2}},
  year = {2025},
  doi = {10.1101/2025.04.09.648075},
  howpublished = {bioRxiv},
}
Figure