In July, DeepMind announced that its AlphaFold model had worked out how most of the proteins in our bodies fold. Pushmeet Kohli tells New Scientist that there is more to come
15 December 2021
IT TOOK decades for scientists to unlock the structure of just 17 per cent of the proteins in the human body. But UK-based AI company DeepMind raised the bar to 98.5 per cent in July when it announced that its AlphaFold model could quickly and reliably calculate the way proteins fold. This could lead to targeted drugs that bind to specific parts of molecules.
We caught up with Pushmeet Kohli at DeepMind to see how work is progressing with mapping almost every one of the more than 100 million known proteins that have been sequenced from across the tree of life.
Were you surprised at the success of AlphaFold, considering that figuring out protein folding previously required vast supercomputers?
We went in with the thesis that machine learning and AI had a role to play. But a lot of the team were uncertain as to whether this problem was solvable. It came as a very pleasant surprise.
You plan to release many more protein structures. Why not leave the problem with scientists who now have access to AlphaFold?
We open-sourced the model and the code so anyone on the planet can find the structure of any protein that they want. We’re already seeing universities and labs across the world using our code. But the reason we’re expanding the database release is because there’s a lot of time and investment involved, and you don’t want different people finding the structure of the same protein again and again, right? It will be very useful if we actually just do it once and for all, for everyone.
Which are you working on first?
We’ve received feedback from the community as to which organisms and which types of proteins we should prioritise next. So we’re working along that road map, eventually moving into what we have committed to, which is releasing the structure of the entire protein universe.
Does that involve new work, or just applying AlphaFold at scale?
The team has been constantly improving the accuracy of the model. But we also want to expand what AlphaFold can do. So, we’d worked on single proteins, but complexes are important because when you look at the biological mechanisms at play, it’s very infrequent that there will be a single protein just interacting with some other sort of small molecule in isolation. So, composite structures – that’s what we have been expanding AlphaFold to do.
Will you ever reach a point where you have mapped everything, and AlphaFold can retire?
Proteins will change, life changes. As evolution operates, you will see different types of proteins coming into play. And so AlphaFold will have a life, not only in complexes, but also in thinking about how the structure is evolving.
And what about covid-19?
Very early on, we found the structure of all the SARS-CoV-2 proteins. Some had been experimentally validated, but many were very difficult to figure out by experimental methods. When scientists actually found the structures, it was interesting to see that ours were nicely consistent.
Now, with variants, again there is an element that these small mutations lead to changes in the structure, but AlphaFold is not currently sensitive to very small changes. So we want to make sure that future versions of AlphaFold are able to really be sensitive to mutations.
Pushmeet Kohli heads the Robust and Reliable AI and AI for Science teams at DeepMind
2021 in review
This year, we have taken on two massive existential threats, the ongoing coronavirus pandemic and the climate emergency, but there were more headline-grabbing stories that give us reasons for optimism going into 2022.
More on these topics: