Scientists' attempts to design enzymes from scratch using artificial intelligence (AI) have often yielded limited results, with the enzymes typically ceasing to function after the first step of the reaction. Designing enzymes with complex active sites that mediate multi-step reactions remains a formidable challenge.

Using serine hydrolases as a model system, David Baker's team at the University of Washington combined the generative capabilities of RFdiffusion with an integrated generative approach for assessing active site pre-organization, starting from minimal active site descriptions to design enzymes.

Experimental characterization revealed catalytic efficiencies (kcat/Km) as high as 2.2×10^5 M^−1 s^−1, with crystal structures closely matching the design models (Cα RMSDs < 1 Å).

By selecting structural compatibility along the reaction coordinate, new catalysts with five different folds, distinct from those of natural serine hydrolases, were identified in low-throughput screening.

The enzymes they designed accelerated a four-step chemical reaction crucial to many biological and industrial processes, including plastic recycling.

Huimin Zhao, a synthetic biologist at the University of Illinois Urbana-Champaign, said: "This is a milestone in enzyme engineering. It shows that we can now design enzymes with natural activity, making them practically valuable."

The study, titled "Computational design of serine hydrolases," was published in Science on February 13, 2025.

One hypothesis regarding the defects in designing serine hydrolases has been the potential geometric incompatibility between the complex hydrolase active sites and the fixed protein scaffold libraries previously used.

"It's like going to a thrift store to buy a suit that may not fit you perfectly. This is what happens when we try to design enzymes in this way," said Anna Lauko, a protein designer at the University of Washington in Seattle and one of the paper's authors.

Baker's team investigated whether increasing scaffold diversity could help identify backbones that more accurately reconstruct the desired active sites and conducted preliminary design activities, searching for the location of serine hydrolase active sites in a large scaffold library based on the nuclear transport factor 2 (NTF2) fold.

As in previous studies, experimental characterization of the resulting designs showed that, although the experimental structures were very close to the designed structures, serine was activated but there was no catalytic conversion of ester substrates. They suspected that the inability to install key catalytic features (e.g., the backbone oxygen anion hole contacts common to all serine hydrolases) into the NTF2 limited the functionality of these designs.

The researchers inferred that advancements in deep learning for protein design could enable the de novo design of proteins, directly constructing serine hydrolase active sites and assessing the design compatibility of the entire multi-step catalytic cycle.

The latest paper using RFdiffusion to build functional sites indicates that the method can improve computational and experimental success rates across a range of design tasks; here, Baker's team aimed to use the same approach to generate serine hydrolases starting from the geometric description of the active site.

In simple terms, the researchers first used a method called RFdiffusion to generate new enzyme structures from scratch. Then, they created a deep neural network called PLACER, which improves structural designs by simulating the positions of atoms in the enzyme and the molecules it binds to at each step of the reaction.

The researchers said that this AI acts like a "filter": it checks whether the enzyme's active site (the part that interacts with molecules) is compatible and correctly aligned for each step of the reaction. Huimin Zhao described this as a "very innovative application."

Experimental Results

The experiments showed that the team's designs had significant catalytic efficiency, active site complexity, and atomic precision, representing a major advancement in computational enzyme design.

The serine catalytic triad oxyanion hole mechanism involves a complex mechanism that is difficult to construct (e.g., compared to Kemp elimination enzymes, which only require a general base in a hydrophobic environment), necessitating the chemical activation of serine and proceeding through a complex multi-step mechanism spanning the chemically stable AEI.

The designed serine hydrolases here had efficiencies as high as 2.2×10⁵ M⁻¹ s⁻¹, a significant functional advancement for computationally designed enzymes.

For example, previously designed esterase OE1 had a kcat/Km of 210 M⁻¹ s⁻¹, and after four rounds of directed evolution and screening of over 12,000 clones, despite using a more active Nδ-methylhistidine nucleophile, its efficiency reached 3,190 M⁻¹ s⁻¹.

In terms of mechanism, the closest de novo design study involved mutating a cysteine-based catalytic triad to a peptide-based helical bucket, proceeding through a more active thioester intermediate, with a kcat/Km of 3.7 M⁻¹ s⁻¹ and kcat of 0.0005 s⁻¹, which were 60,000 times lower and 400 times slower than the most efficient design (momi120-103) and highest turnover design (momi120) described here, respectively.

The ability to accelerate the hydrolysis of chemically stable acyl-enzyme intermediates has been a challenge in enzyme design for decades. To approximate the increase in deacylation rate, the researchers compared the uncatalyzed rate of ethyl acetate hydrolysis [(2.5–5.0)×10⁻¹⁰ s⁻¹] with the lower limit of the momi deacylation rate constant (kcat, 0.076 s⁻¹, pH 7.0, 25°C), estimating a rate enhancement of over 10⁸.

Overall, the design of serine hydrolases spanned a fivefold improvement beyond what is found in natural esterases, with significantly higher activity than previously designed esterases and accelerated deacylation, representing a key advancement in enzyme design.

Discussion and Evaluation

Here, the researchers used a de novo construction method combining RFdiffusion with PLACER integrated analysis to ensure the accuracy and pre-organization of the designs, enabling them to test these hypotheses directly, which should complement more traditional methods optimized through structural inspection, computational analysis, and directed evolution.

Previous designs based on the catalytic triad failed to achieve multiple turnovers; in some cases, such as the team's preliminary designs based on NTF2, the backbone amide oxygen anion hole could not be realized due to scaffold limitations, and in other designs based on natural scaffolds, the geometry of histidine was difficult to control, which may have limited the activation of the leaving group and water.

The de novo generation method of using RFdiffusion to build the backbone outward from a specific active site overcomes these limitations, as it can generate almost any required catalytic geometry.

The researchers further demonstrated that the deep neural network PLACER can rapidly generate a collection of reaction intermediates to predict pre-organization and provide insights that would otherwise require labor-intensive structural studies.

For example, PLACER revealed widespread off-target conformational changes in acyl-enzyme intermediates, providing feedback on design flaws that would be overlooked when considering only a single state in the catalytic cycle.

The value of this approach lies in the significantly increased experimental success rate after filtering with PLACER, indicating that this ensemble generation will benefit the development of enzyme design.

Although the designs described here do use known mechanisms, the geometries sampled and the folds supporting them are different from those found in natural proteins, and the insights PLACER provides on these geometries suggest that the method should be valuable for assessing catalytic geometries without natural precedents.

The researchers stated that the ability to precisely position multiple catalytic groups using RFdiffusion and to assess active site organization throughout the complex reaction cycle using PLACER should enable the design of various novel catalysts in the near future, such as PETase, amidases, and ligases.

References:

https://www.science.org/doi/10.1126/science.adu2454

https://www.nature.com/articles/d41586-025-00488-3