AI and the Nobel Prizes: An interview with Martin Engqvist

The 2024 Nobel prize for physics was awarded to John Hopfield and Geoffrey Hinton, known as the godfathers of Machine Learning and AI, “for foundational discoveries and inventions that enable machine learning with artificial neural networks.”

The chemistry prize went to three people who are pioneers in the use of AI in the study of proteins, the building blocks of life. Half went to David Baker of the University of Washington "for computational protein design," with the other half going to Google DeepMind's Demis Hassabis and John Jumper "for protein structure prediction." Hassabis and Jumper are best known for creating AI models that can predict three-dimensional protein structure, the most famous of which is AlphaFold, now in its third iteration.

As the head of enginzyme’s computational enzyme design group, Martin Engqvist leads* efforts to figure out how AI and machine learning can improve the performance of enginzyme’s biocatalysis solutions. Martin sat for an interview with James Connell, a former journalist with The New York Times, to discuss the Nobel Prizes and the roles AI and machine learning are playing in biotech and chemistry.

James Connell: Martin, tell me about your reaction when you saw the announcement for this years’ Nobel Prices in physics and chemistry.

Martin Engqvist: Well I was excited because I've been using these tools, more or less daily, for years now. I also read a lot of David Baker's work, and I use AlphaFold of course. It's very exciting for everyone in the biotech field.

Everyone is saying this is a huge moment for AI, but what do these Nobel Prizes say about advancement in the study of proteins, and more particularly, enzymes?

Biology is chemistry and both are governed by the laws of physics. Clean and easy, right? No, it is much messier than that.

Hassabis has said that if math is the language of physics, then machine learning might be the language of biology. What that means is that in physics you strive for something like Occam's Razor: You want to find elegant, clean, simple equations that describe laws of nature.

Of course biology is governed by the same laws, but it gets really messy, really fast, because biology is so complex. It is very hard to simulate a whole cell or simulate a brain. No one can do that. Yet. So that's where machine learning comes in: it lets you do supercharged pattern recognition and predict things. So you can predict outcomes. You can predict properties even if you cannot write down clean equations.

If it wasn’t obvious before this years’ prize it is now: machine learning and AI are the languages we need to use if we are going to better understand biology and get outcomes that are predictive and useful. A great example of useful AI predictions are those generated by AlphaFold. In minutes this model generates predictions of a protein's three-dimensional structure, something which is key to understanding how it works. Experimentally determining this structure in the laboratory can be as fast as a week or two, but may also take years. David Baker himself said that for him, the development of AlphaFold was proof that deep learning is extremely important and is the way to go. Since AlphaFold was made available -- I think it was in 2018 -- David and his team have published several tools that rely on deep learning to design proteins.

These are tools that you use in your work at enginzyme? What do they enable you to do in biocatalysis that was impossible before?

I use these tools to design enzymes (which are proteins) that help enginzyme produce advanced biocatalysts that do chemistry cleanly and efficiently, with less wasted energy and material. I use AlphaFold to predict structures for our enzymes, when no experimentally determined structures are available. With those structures in hand, I can determine which parts of the enzymes need to be edited in order to engineer improved biocatalysts. I can then use some of the tools from David Baker to make the required edits. For me, these tools have been absolutely game-changing as they let me make better enzymes in less time.

What are the major shortcomings when it comes to using AI in chemistry or the life sciences?

Deep learning models are amazing, but they are essentially black boxes — and that is a limitation. What I mean by that is that we don't understand how a deep learning model comes to its conclusions, so we cannot extract knowledge from them directly or explain why something works. But this has not been a deal-breaker for me so far.

Sure, it would be great if we could interpret exactly what happens in the model. But the protein engineer in me thinks this doesn’t much matter as long as the machine’s predictions are accurate. Still, as a scientist, I want to understand exactly how things happen and why.

There is a lot of concern that deep learning models are pushing us away from this explanatory science. In a way they are, but I think they also help us reach new places from which we can start asking new questions and actually understand things on a more fundamental level: Why does this protein fold in this way? This was unexpected. Let's try to figure it out.

I don't think we will move to a place where everything is solved through deep learning and we stop caring about how things work. We’ll always have scientists who won’t stop until they understand why and how, that’s just human nature.

*Martin has left enginzyme to take a job closer to his home in Gothenburg, Sweden. The interview was carried out Oct. 21.

Read more about this topic:

The Economist published a comprehensive overview of AI's big Nobel splash. It is lively and informative read.

In a 2003 interview, Martin Engqvist explained how he and a group of other scientists developed a machine learning model capable of predicting enzyme-substrate relationships across all proteins — a tool that would help narrow chemists’ efforts to find enzyme-substrate pairs that would be most likely to work. The model, known as ESP, demonstrated accuracy of more than 91% and can be applied across a wide range of different enzymes. Read their paper in Nature Communications.

Using Anthropic CEODario Amodei's "Machines of Loving Grace" essay as a starting point, Niko McCarty of Asimov Press discusses some of the bottlenecks that are holding back AI-enabled biology and medicine.

Madhumita Murgia of The Financial Times wrote up a fascinating post-Nobel interview with Demis Hassabis of DeepMind. He's excited about progress in biology, but also material design and climate modeling.

The official Nobel Prize website is very user-friendly and contains a wealth of interesting content.