Computer-based design for custom proteins

“Imagine a future in which you can design enzymes - natures catalysts – for your specific application at the push of button”, says Gustav Oberdorfer who headed the ERC Starting Grant project ‘HelixMold’ at the Institute of Biochemistry of the Graz University of Technology (TU Graz). Devising a broadly applicable method for the computational design of non-naturally occurring proteins with specific characteristics – this was the general objective of the project. Now, at the end of the project, Oberdorfer and his team have achieved their goal, laying the foundations for significantly more specialised, faster and more precise design of proteins for various applications, including biocatalysts for the degradation of polymers such as cellulose, and the production of active ingredients for pharmaceuticals.

From fragments to parametrisation

“In the past, we would take fragments of known proteins from a protein database and then reconstruct them using computer-aided random sampling experiments – called Monte Carlo simulations – in order to find the best configuration,” Gustav Oberdorfer explains. “There are often close similarities between the small components that make up different proteins, but it isn’t entirely clear how they fold to create a complete protein. So building something new from old parts was a challenge.”

Oberdorfer’s strategy, for which he was awarded an ERC Starting Grant in 2018, was centred on parametric design. The underlying idea was put forward by Francis Crick, who in 1953 made calculations aimed at directly determining the position of atoms in a given protein. This approach gave the researchers a far greater degree of control and enabled them to computer-generate thousands of very similar, but also slightly different starting structures.

Suitable structure for active centre

Using the method devised under the HelixMold project, the research team is able to create tens of thousands of slightly varying basic structures for a protein on a computer. In the simulation, the researchers then determine which of the generated structures can host a catalytic centre geometrically. These active centres enable specific catalytic reactions. Once a suitable structure has been identified, it is secured in the centre. The team then moulds the rest of the protein around the centre using specially developed simulation software. This is done until the interaction energy between individual amino acids in the protein reaches an energetic minimum. 

Besides the regular elements in a protein structure, unstructured, so-called loop regions, are another important protein component. They are flexible and play a key role in protein function, in particular for connections with other structural elements. In the past, however, loops often ended up behaving differently than originally anticipated in the design. For this reason, as part of the project Florian Wieser developed an AI that was trained using thousands of experimentally determined loop structures, in order to discover what plausible loops look like. The AI was deployed during the design process for quality control purposes, to determine whether a particular loop design would be effective or not.

AI revolution leads to shift in focus

Ultimately, artificial intelligence had a far bigger part to play in the project than was originally assumed. The HelixMold team capitalised on the rapid advances in machine learning, which enabled them to complete their calculations even more efficiently. The basic idea remained the same, but the results were achieved more quickly.

“AI methods such as AlphaFold and RosettaFoldDiffusion have revolutionised protein-structure prediction and structure generation, so in the end, parametric design was not as crucial as it had been at the start of the project,” says Gustav Oberdorfer. “At the end of the day, we achieved what we set out to do with HelixMold. It’s a question of understanding a chemical reaction, finding a structure for an active centre on the computer, and then moulding the protein around it. So now, the focus of research is shifting away from identifying and adapting natural proteins, and towards the targeted design of new molecules for specific applications. The principles and knowledge that we generated in HelixMold will be a vital building block for future research. And we’re already drawing up plans for our next steps.”