Google Deepmind has launched a new version of AlphaFold, an innovative machine learning model that predicts the shape and behavior of proteins. AlphaFold 3 is not only more accurate, but also predicts interactions with other biomolecules, making it a much more versatile research tool. And the company is making a limited version of this model available online for free.
Since the first AlphaFold debuted in 2018, this model has remained the leading method for predicting protein structure from the sequence of the amino acids that make up the protein.
This sounds like a rather narrow challenge, but understanding at the molecular level how proteins play an almost infinite variety of roles in the body is the basis of nearly all biology. In recent years, computational modeling techniques such as AlphaFold and RoseTTaFold have replaced expensive laboratory-based methods and accelerated the work of thousands of researchers across many disciplines.
But as DeepMind founder Demis Hassabis said at a press conference about the new system, the technology is still in its infancy, and each model is “just a step forward.” The company teased its release late last year, but this is its official debut.
We'll leave you to our scientific blog to explain in more detail how the new model improves the results, but here we'll explain how various improvements and modeling techniques have made AlphaFold 3 not only more accurate; Suffice it to say that it is now more broadly applicable.
One of the limitations of protein modeling is that even though we know what shape an amino acid sequence will take, we don't necessarily know how it will combine with other molecules. And if they wanted to actually do something with these molecules, which they almost always do, they had to find it through more laborious modeling and testing.
“Biology is a dynamic system, and we need to understand how biological properties emerge through interactions between different molecules within cells. You can think of it as a big first step,” Hassabis said. “Of course, we can model not only proteins that interact with other proteins, but also proteins that interact with other biomolecules, such as important DNA or RNA strands.”
AlphaFold 3 allows you to simulate multiple molecules at once. For example, a DNA strand, some DNA-binding molecules, and maybe some ions for added spice. This is what you get with one such specific combination. A DNA ribbon runs down the center, and proteins shine on the sides. I think these are ions in the middle like little eggs.
Of course, this in itself is not a scientific discovery. But just understanding whether an experimental protein actually binds, binds in this way, or transforms into this shape is typically a task that takes at least several days, and sometimes weeks or months. did.
It's difficult to overstate the excitement in the field in recent years, but researchers have been largely held back by the lack of interaction modeling (which newer versions offer in its form) and the difficulty of deploying models. I'm here.
This second problem is probably the larger of the two. New modeling techniques are “open” in a sense, but like any AI model, they are not necessarily easy to deploy and operate. That's why Google Deepmind offers AlphaFold Server, a free, fully hosted web application that makes this model available for non-commercial use.
It's free and very easy to use. I did it in a separate window during the call while they were explaining (this is how I got the image above). All you need is a Google account, enter as many sequences and categories as you can handle (some examples are provided), and submit. It takes minutes and displays a colored live 3D molecule representing the model's confidence in the conformation at that location. As you can see in the image above, the tip of the ribbon and the parts exposed to rogue atoms become brighter or redder, indicating less confidence.
I asked if there is any real difference between the publicly available model and the model used internally. “We have made available most of the features of the new model,” Hassabis said, without providing further details.
This is clearly Google pushing all of its weight, but it also keeps some of the best parts for itself, which of course is Google's prerogative. Creating such a free hosted tool requires committing significant resources to this task. Don't get me wrong, this is a gold mine, an expensive shareware version (for Google) to convince researchers around the world that AlphaFold 3 is what it's supposed to be. At least they had arrows in their quivers.
Image credit: Google Deepmind
But that's okay. That's because the technology will likely print money through Alphabet's subsidiary (Google's…cousin?) Isomorphic Labs. Isomorphic Labs uses computational tools like AlphaFold for drug design. Well, everyone uses computational tools these days, but as Hassabis pointed out, Isomorphic was the first to crack his Deepmind's latest model and call it “some unique stuff about drug discovery.” ” combined. The company already has partnerships with Eli Lilly and Novartis.
However, AlphaFold is not the last resort in biology. As countless researchers agree, it's just a very useful tool. And this allows us to do what his Max Jaderberg at Isomorphic calls “rational drug design.”
“If you think about how this is impacting isomorphic laboratories every day, scientists and drug designers can formulate and test hypotheses at the atomic level and generate highly accurate structural predictions within seconds. “Scientists then reason out what interactions should occur and how to proceed with their design to make good drugs,” he said. “This compares to what could take months or even years to do this experimentally.”
While many will praise the achievements and widespread availability of free hosted tools like AlphaFold Server, some may rightly point out that this is not an actual victory for open science.
As with many proprietary AI models, AlphaFold's training process and other information important to its reproduction (which, as you may recall, is a fundamental part of the scientific method) are mostly It's being hidden. The paper published in Nature describes in some detail how it was made, but many important details and data are missing, leaving science wanting to use the most powerful molecular biology tools on the planet. The watchful eye of Alphabet, Google, and DeepMind (no one knows which one will actually take the lead).
Proponents of open science have long argued that while these advances are impressive, sharing this kind of thing openly is always good in the long run. After all, that's how science has evolved, and indeed how some of the world's most important software has evolved as well.
Providing AlphaFold Server for free to academic or non-commercial applications is a very generous gesture in many ways. But Google's generosity rarely comes with strings attached. Still, I have no doubt that many researchers will take advantage of this honeymoon period to use this model as long as humanly possible before the other shoes drop.