A new iteration of the protein design model RoseTTAFold Diffusion, developed in the lab of 2024 Nobel prize-winner David Baker, has created blueprints for functional enzymes from scratch.¹

‘Enzymes are the workhorses of life,’ says Carlos Acevedo-Rocha, a senior researcher at the Technical University of Denmark, who was not involved with developing the new model. ‘They catalyse complex chemical reactions to supply cells with energy, nutrients and materials,’ he explains, and this make them extremely valuable in biotechnology.

Enzymes, like all proteins, are chains of amino acids but more complex and dynamic, making designing them a challenge.

Rohith Krishna, a postdoctoral fellow at the University of Washington who helped design the new model called RF Diffusion 2, says there are two main hurdles in enzyme design. First, enzymes’ targets are rarely proteins, but small molecules and the original models struggled with non-protein interactions. Second, is the positioning of the protein side chains and specifically their atoms in the catalytic or active site of the enzyme.

In protein–protein interactions the amino acid backbone sequence is key. ‘But for enzyme design, exactly where the side chains are in the active site is super important,’ says Krishna.

To overcome this the team expanded the model to include side chain atoms. ‘Every atom in the side chain had its own set of features and it could interact with all the other things in the network,’ explains Krishna. They also had to relax other parameters of the model. ‘For previous models, when you specify that you want some sort of residue in the protein structure, you have to say where in the sequence it occurs,’ says Krishna. ‘We found that when designing enzymes, especially for ones where there are no natural cognates, we didn’t know where they should go.’

They instead trained the model to decide for itself the best sequence order and saw an increase in the diversity of designs. ‘There’s so many different solutions,’ says Krishna, ‘and [the model] gets to explore all of them.’

To test the model, the team attempted to create zinc-based enzymes that break ester bonds. They first used quantum chemistry calculations from naturally occurring zinc metallohydrolases to determine where the important atoms in the active site are.² Trained on this data, RF Diffusion 2 was able to design proteins with a scaffold where the atoms were positioned in the active site to break ester linkages.

The proposed designs were then synthesised and tested in the lab where they achieved enzymatic activities close to those found in nature. ‘They’re not the best enzymes that have ever existed in nature, but they’re in that range of natural activity,’ says Krishna.

The new enzyme sequences also showed little similarity to known proteins, confirming the model’s designs were unique.

‘This development is very exciting,’ says Steffen Lindner-Mehlich, a biochemist at the Charité University Hospital in Berlin. He says it could provide groups like his own with novel enzymes to ‘design tailored synthetic pathways’ for a variety of new applications.

Acevedo-Rocha agrees the work is promising but points out, ‘the success rate for highly efficient enzymes is still in the ~1% range’. This means many designs need to be synthesised and screened, something that is too costly for many labs. Other properties need to be assessed too. ‘For example, how suitable they might be for a given industrial environment like a bioreactor.’ However, he is encouraged by the speed of developments such as the release of RF Diffusion 3.

RF Diffusion 3 is now freely available and, according to Krishna, handles more non-protein molecules and catalytic sites and ‘it’s 10 times faster, but also more precise at placing those atoms’.