Protein By DesAIgn - a blog post by ESR13 Martijn Bemelmans

In the Autumn of 2024, I had the pleasure of going on a secondment to Novo Nordisk in Copenhagen, Denmark. With most of my PhD work focusing on molecular dynamics simulations to search for (cryptic) small molecule binding pockets, it was truly an insightful shift of gears to get some exposure to larger modalities and generative artificial intelligence (AI) methods. The idea is straightforward: given a target protein, generate the backbone of a protein binding partner, then generate a multitude of sequences that should be able to adopt this backbone fold, and finally select designs based on your favorite scoring function. In practice, the process is naturally more involved, but it was still remarkable to see just how powerful this approach can be. This not only became clear during my stay at the company, but to finish off my secondment, I also got to attend the European RosettaCon.

The work presented by academic and industry groups alike at this conference was a true showcase of the value that generative AI can bring. First, several works showed that de novo protein design can generate enzymes that not only match the catalytic efficiency of their natural counterpart but even exceed it by orders of magnitude. Then, there were examples of methods that can design functional protein binders in one shot and pipelines to rapidly design antibodies. A final highlight for me was a deep dive into how AlphaFold2 generates structures. This work indicated that learning how to generate protein structures with reasonably accurate geometries is very fast, requiring only about 10% of the total data available for model training. However, it takes up to about 90% of training data before the model can assign an accurate confidence score that correlates with structure quality. In other words, the vast majority of training data is not required to achieve high-quality results but rather to get an accurate understanding of the quality of the results.
​
Altogether, my experience in Denmark has been instrumental in widening my perspective on the field of therapeutics discovery from simulation-based approaches and small molecules to generative AI-based protocols and protein design. It is interesting and encouraging to see the impact that computational work can have across the board in the pharmaceutical field. Whether the aim is to develop a new binder for a hidden pocket that only appears in rare protein conformations or to obtain protein-binding proteins, there are always actionable insights to be gained from computational investigation.