Abstract
De novo enzyme design traditionally involves constructing a protein scaffold around a theoretical "theozyme"—an ideal active site geometry designed to stabilize a reaction's transition state. Historically, this process has been limited by the need to specify the exact sequence positions and backbone coordinates of catalytic residues. Here, we discuss a generative AI method, RFdiffusion2 (also referred to as RFam), which enables the direct generation of protein structures around active site atoms without these constraints. Using this approach, researchers designed zinc metallohydrolases starting from density functional theory (DFT) descriptions. The resulting enzymes achieved catalytic efficiencies ($k_{cat}/K_M$) up to 53,000 $M^{-1} s^{-1}$, representing a significant leap in performance for computationally designed catalysts without the need for experimental optimization.
Introduction: The Challenge of Enzyme Scaffolding
Nature’s enzymes are highly optimized catalysts, but replicating their efficiency in human-designed proteins has proven difficult. The primary bottleneck in de novo design is the "scaffolding problem": once a theoretical active site is modeled, how does one build a stable protein that holds those specific residues in the precise orientation required for catalysis?
Previous iterations of tools like RFdiffusion required designers to pre-define the backbone coordinates and sequence positions of catalytic residues. This created a massive sampling challenge, as the possible combinations of sidechain orientations (rotamers) and sequence placements are nearly infinite. Consequently, most previous de novo enzymes exhibited very low activity, often requiring extensive rounds of laboratory-based directed evolution to reach functional levels.
Methodology: RFdiffusion2 and Atomic Scaffolding
To address these limitations, the researchers developed RFdiffusion2, an evolution of the diffusion framework that introduces two critical capabilities:
- Atomic Substructure Scaffolding: Unlike previous models that focused primarily on protein backbones, RFdiffusion2 can scaffold arbitrary sets of heavy atoms. This allows the model to build a protein around the functional groups of the active site itself, rather than a pre-defined skeleton.
- Sequence-Position Agnosticism: The model does not require the user to specify where in the primary sequence a catalytic residue should fall. It treats the active site atoms as fixed targets and "grows" the protein around them, automatically determining the optimal sequence placement and orientation to support the site.
The Design Pipeline
The workflow integrated quantum chemistry with generative modeling:
- DFT Modeling: Defined the ideal geometry for a zinc-mediated hydrolysis reaction.
- RFdiffusion2 Generation: Produced thousands of unique protein backbones around the DFT-defined geometry.
- ProteinMPNN: Assigned amino acid sequences to the generated backbones.
- Validation: Designs were filtered using AlphaFold2 for fold reliability and tools like PLACER, Chai-1, and ChemNet to predict the "preorganization" of the active site—ensuring the residues remain in a ready-to-react state in solution.
Results and Experimental Performance
The researchers tested an initial set of 96 designs. Unlike previous efforts where success rates were low, 86 of these designs were soluble and correctly folded.
Catalytic Efficiency
- First Round: The most active design from the initial set achieved a $k_{cat}/K_M$ of 16,000 $M^{-1} s^{-1}$.
- Second Round: A subsequent round of 96 designs yielded three additional highly active enzymes, with the top performer reaching a $k_{cat}/K_M$ of 53,000 $M^{-1} s^{-1}$ and a $k_{cat}$ of 1.5 $s^{-1}$.
These values are orders of magnitude higher than previous de novo metallohydrolases, which typically range from 3 to 60 $M^{-1} s^{-1}$. Notably, this performance was achieved "zero-shot," meaning the enzymes were highly active straight from the computer without any experimental fine-tuning.
Structural Novelty and Preorganization
Computational analysis revealed that the successful designs were structurally distinct from any known proteins in the Protein Data Bank (PDB). These were not mere mimics of existing enzymes but entirely new structural solutions.
A key factor in their success was preorganization. In the most active designs, the catalytic residues and the substrate were held in a rigid, optimal pose even before the reaction began. In contrast, inactive designs showed high fluctuations in their active site geometries. This suggests that the ability of RFdiffusion2 to create rigid, well-supported environments is central to its success.
Discussion and Future Directions
While the designs represent a breakthrough, they also highlight areas for refinement. For instance, mutagenesis experiments showed that certain residues intended to act as general bases occasionally competed for metal binding, suggesting that future models must better account for the complex interplay between metals and surrounding residues.
Future work will focus on:
- Oxyanion Hole Incorporation: Adding hydrogen-bond donors to further stabilize transition states.
- Bimolecular Reactions: Expanding beyond hydrolysis to more complex multi-step reactions.
- Iterative Refinement: Using partial diffusion to "noise" and "denoise" specific regions of a protein to optimize binding affinity and catalytic rate.
Conclusion
The ability to generate highly active catalysts using quantum chemistry-calculated geometries and generative AI marks a new era in protein engineering. By moving away from rigid backbone specifications and toward functional atomic scaffolding, RFdiffusion2 enables the creation of potent, designer enzymes that rival the efficiency of natural catalysts.
Summary updated