A Significant Biology Problem Has Finally Been Solved

It seems obvious that AlphaFold 2 is picking up on something implicit about how chemistry and physics work

Soon after its AlphaGo AI made waves in 2016 by defeating world champion Lee Sedol at the game, DeepMind created AlphaFold. However, according to Hassabis, the objective was always to create AI that could address significant scientific issues. Nearly all species for which amino acid sequences are known to have had their protein structures made openly accessible in a public database by DeepMind.

Hassabis was interviewed by Scientific American on creating AlphaFold, some of its most intriguing future uses, and the moral ramifications of highly developed AI.

Why did you choose to develop AlphaFold, and how did you get to the point where it could essentially fold every protein that is currently understood?

 We essentially began the project the day after we returned from the Seoul AlphaGo match, where we defeated Lee Sedol, the reigning world [Go] champion. What is the next significant project that DeepMind should undertake? I was chatting with Dave Silver, the project head of AlphaGo. We had just solved what was essentially the peak of gaming AI, so I felt like it was time to take on a truly challenging scientific problem. I was eager to use AI in real-world contexts. The goal of DeepMind has always been to create general-purpose algorithms that may be used to solve a wide variety of issues. Games were where we started since they were a particularly effective way to test and develop ideas for a variety of purposes. However, that was never the final objective. The ultimate objective was to create things like AlphaFold.

It was a massive undertaking that took about five or six years to complete before CASP14. At the CASP13 competition, we used AlphaFold 1, an early version. That was cutting-edge, you know, significantly better than anything that had been done before, and I believe it was one of the first instances when machine learning had been utilized as the main building block of a system to attempt to solve this problem.

Were you surprised by AlphaFold's popularity?

 Yes, that was genuinely unexpected. The hardest thing we've ever done, in my opinion, as well as the most intricate system we've ever created. The Nature article, which includes all the methodologies' descriptions, supplemental data, and technical specifics, is 60 pages long. Each of the 32 individual component algorithms is required. It required a lot of ingenuity because the architecture was so sophisticated. Because of this, it took so long. Having all these varied viewpoints from various fields and backgrounds was crucial. And mixing those things together—rather than just machine learning and engineering—is something I believe DeepMind does very effectively.

Can you describe how AlphaFold works in the most basic terms possible?

It's a fairly complex situation. And there are many things about which we are unsure. It seems obvious that AlphaFold 2 is picking up on something implicit about how chemistry and physics work. It has some idea of what might be plausible. It discovered this after viewing actual protein structures, those that are familiar to us. One of our inventions was to use an early version of AlphaFold 2 to forecast several structures as well as the degree of confidence in those predictions. This technique is known as self-distillation.

We used a method called multisequence alignment to include this information on chemical bond angles and evolutionary history. These introduce some limitations, which aid in reducing the range of potential protein configurations. The search space is too big to handle manually. But given that proteins fold in nanoseconds or milliseconds, it is clear that real-world physics finds a way to resolve this. We are actually attempting to reverse engineer that process by studying examples of the output. I believe AlphaFold has effectively encapsulated a profound aspect of the physics and chemistry of molecules.