Google Deepmind has unveiled a new version of AlphaFold, its transformative machine learning model that predicts the shape and behavior of proteins. In addition to being more accurate, AlphaFold 3 also predicts interactions with other biomolecules, making it a far more versatile research tool—and the company makes a limited version of the model available for free online use.
Since launching the first AlphaFold in 2018, the model has been the leading method for predicting protein structure based on the sequence of amino acids that compose them.
Although this may seem like a rather narrow task, understanding proteins – which perform an almost endless variety of tasks in our bodies – at the molecular level is fundamental to almost all of 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 in as many areas.
But the technology is still a work in progress, and each model is “just a step along the way,” as Deepmind founder Demis Hassabis put it in a press conference about the new system. The company announced the release late last year, but this is its official debut.
I’ll leave it to the science blogs to go into exactly how the new model improves the results, but suffice it to say that a variety of improvements and modeling techniques have made AlphaFold 3 not only more accurate, but also more widely applicable.
One of the limitations of protein modeling is that even if you know what an amino acid sequence will look like, that doesn’t necessarily mean you know what other molecules it will bind to and how. And if you wanted to actually do things with these molecules, which is the case with most of them, you had to figure it out through more sophisticated modeling and testing.
“Biology is a dynamic system, and one must understand how properties of biology arose as a result Interactions between different molecules in the cell. And you can think of AlphaFold 3 as our first big step in that direction,” said Hassabis. “It is able to model proteins that naturally interact with other proteins, but also with other biomolecules, primarily DNA and RNA strands.”
AlphaFold 3 can be used to simulate multiple molecules at once – for example, a DNA strand, some DNA-binding molecules, and maybe a few ions to spice things up. Here’s what you get for such a specific combination: the DNA ribbons run down the middle, the proteins shine on the side, and I think these are the ions embedded there in the middle like little eggs:
Of course, this is not a scientific discovery in itself. But even figuring out whether an experimental protein would bind or shape into that shape at all, or in that way, was generally the work of at least days, or perhaps weeks to months.
While it is hard to overstate the excitement in this area in recent years, researchers have largely been hamstrung by the lack of interaction modeling (of which the new version provides a form) and the difficulties in implementing the model.
This second problem is perhaps the bigger of the two, because although the new modeling techniques were “open” in a sense, like other AI models, they are not necessarily easy to deploy and use. That’s why Google offers Deepmind AlphaFold Server, a free, fully hosted web application that makes the model available for non-commercial use.
It’s free and pretty easy to use – I did it in another window during the call while they explained it (which is how I got the image above). All you need is a Google account and then enter as many sequences and categories as it can handle – there are a few examples – and submit. In a few minutes your work should be done and you will end up with a living 3D molecule, the color of which represents the model’s confidence in the conformation at that position. As you can see in the image above, the tips of the bands and the parts more exposed to unwanted atoms are brighter or red to indicate less confidence.
I asked if there was a real difference between the publicly available model and the internally used model. Hassabis said: “We have made most of the capabilities of the new model available,” but did not elaborate.
It’s clearly because Google is throwing its weight around – while keeping the best bits to itself to some extent, which is of course their prerogative. To create a free, hosted tool like this requires significant resources to be dedicated to the task – make no mistake, this is a money pit, an expensive (for Google) shareware version to convince researchers around the world that AlphaFold 3 should have at least one arrow in their quiver.
Photo credit: Google Deepmind
That’s OK, though, because the technology will likely print money through Alphabet subsidiary Isomorphic Labs (making it Google’s…cousin?), which uses computer tools like AlphaFold for drug development. Everyone uses computing tools these days — but Isomorphic was the first to have access to Deepmind’s latest models and combine them with “some other proprietary things that have to do with drug development,” Hassabis noted. The company already has partnerships with Eli Lilly and Novartis.
However, AlphaFold is not the be-all and end-all of biology – just a very useful tool, as countless researchers agree. And it allows them to do what Max Jaderberg of Isomorphic called “rational drug design.”
“As we think about it day in and day out, what impact this has on the Isomorphic labs: It allows our scientists, our drug developers, to create and test hypotheses at the atomic level and then make highly accurate structural predictions in seconds… to help. “Scientists are considering what interactions need to take place and how to advance those designs to develop a good drug,” he said. “This is compared to the months or even years it could take to do this experimentally.”
While many celebrate the achievement and wide availability of a free, hosted tool like AlphaFold Server, others may rightly point out that this is not a real win for open science.
As with many proprietary AI models, AlphaFold’s training process and other information critical to replication – a fundamental part of the scientific method, as you will recall – are largely and increasingly withheld. While the paper, published in Nature, goes into detail about the methods of its creation, it lacks many important details and data, meaning scientists who want to use the world’s most powerful molecular biology tool must do so within Nature’s vigilance Eye of Alphabet, Google and Deepmind (who knows who actually holds the reins).
Advocates of open science have been saying for years that while these advances are remarkable, it is always better in the long run to share such things openly. This is ultimately how science advances, and this is also how some of the world’s most important software programs have evolved.
Making AlphaFold Server freely available for any academic or non-commercial use is a very generous act in many ways. But Google’s generosity rarely comes with strings attached. Undoubtedly, many researchers will still take advantage of this honeymoon period to exploit the model as much as possible before the other shoe drops.