Meet the Protein Decoder

This article is part of our DNDi: Medicine Makers Without Borders series. Read the entire feature on our website or tune in to the accompanying podcast for more.

Ask any literary sci-fi fan to name the greatest fictional supercomputer and they might nominate Douglas Adams’ Deep Thought, most famous for calculating the answer to the question of life, the universe, and everything. (The answer: 42).

The first AI celebrities certainly seemed just as far-removed from the humdrum concerns of hard science. A nonfictional namesake of Deep Thought won the World Computer Chess Championship in 1989, and as far into the “future” as 2017, AI superstardom remained bogged down in board games. But when Google-owned DeepMind’s AlphaGo defeated 18-world-title-holder Lee Sedol in Seoul (and next year, trumped world Go champion Ke Jie in the ancient canal town of Wuzhen), something changed.

Specifically, DeepMind started applying its AI to “real world” scientific problems, like predicting protein structures. Today, just like DNDi, DeepMind participates in open science. Its enormous AlphaFold Protein Structure Database was published online without paywalls in the summer of 2021. So perhaps its collaboration with the DNDi’s Open Synthesis network should come as no surprise. In Ep 4 of our podcast, we spoke to DeepMind’s Head of AI for Science, Pushmeet Kohli, to learn more about the bonds these two incredible organizations are forming.

How did AlphaFold come into being?
 

DeepMind’s mission is solving intelligence to benefit science and humanity. In the past, we’ve showcased our AI performing feats like automatically learning to play Atari Games. Many listeners will have heard of AlphaGo’s triumphs over the world’s best Go players. But in the last few years, we have started to apply our breakthrough AI research in the natural sciences. This led to the science program that I manage, which covers structural biology, quantum chemistry, problems in physics, pure mathematics, and more.

We’re looking at all of the problems in these fields from a scientific perspective, so in some ways we might nowadays call DeepMind a science company, both in terms of our focus on science applications and our scientific approach to asking and answering questions like, “What is intelligence? How can you recreate it? How can you make a machine intelligent?”

And how did DeepMind’s collaboration with DNDi begin?
 

The basic operation of any disease involves certain processes within the body in which proteins are involved and interacting. As a medicine maker, you may wish to break or otherwise intervene in those interactions. But to do that, you’ll want to understand the structure of the proteins involved. This could help you, for example, build a small molecule that binds to a particular protein.

This is the essence of the collaboration between DNDi and DeepMind. We want to zoom in on the proteins that are involved in some of these extremely problematic diseases, and consider the support we can offer using our structure-predicting technology.

What marks AlphaFold as “cutting edge”?
 

Very early on, we knew that the results provided by AlphaFold would create a very broad impact. Protein structure predictions are not only useful for understanding diseases and small molecules; they are like the roots of a tree. If you solve these ‘root node problems’, you unlock so many other solutions to so many other problems in everything from drug discovery to plastic pollution in the oceans. We realized that we wanted to help make significant and dramatic improvements, and we wanted to make them in the most responsible way.

To that end, we openly published our AlphaFold Protein Structure Database in partnership with the European Molecular Biology Laboratory. Previously, scientists working in any given drug development program who needed to develop a small molecule but didn’t know its structure would have to go through years of research. Often, they would have to isolate the protein and then send it to special facilities only available in certain labs in certain countries. Even then, success would not be guaranteed.

AlphaFold sidesteps that problem and accelerates the process. With the means to accurately predict these structures and make them accessible everywhere, we can really do something to help researchers working on neglected diseases – especially those in developing countries, who might not have access to the facilities they would otherwise need.

DeepMind and AlphaFold have blown a lot of minds. How about yours?
 

One of the great and unique things about DeepMind is its belief in the multidisciplinary approach. It’s something I knew about even before I joined them. DeepMind also thinks in the long term, and does not shy away from ambitious goals. I feel one of the biggest lessons I’ve learned from AlphaFold is that with the right approach we can do amazing things that many people dismiss as either impossible, or lying decades beyond our reach.

Listen to the full conversation