Google Deepmind has unveiled a new version of AlphaFold, its transformative 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 the model available online for free.
Since the launch of the first AlphaFold in 2018, the model has remained the primary method for predicting the structure of proteins from the sequence of amino acids that make them up.
Although it seems like a rather narrow task, it is fundamental to almost all biology to understand proteins – which perform an almost infinite variety of tasks in our bodies – at the molecular level. In recent years, computer modeling techniques like AlphaFold and RoseTTaFold have replaced expensive laboratory-based methods, accelerating the work of thousands of researchers in so many fields.
But the technology is still a work in progress, with each model “just a step along the way,” as Deepmind founder Demis Hassabis said during a press call about the new system. The company announced this release late last year, but this marks its official debut.
I’ll let the science blogs explain exactly how the new model improves the results, but suffice it to say here that a variety of improvements and modeling techniques have made AlphaFold 3 not only more accurate, but more widely applicable.
One of the limitations of protein modeling is that even if you know what shape an amino acid sequence will take, that doesn’t mean you necessarily know what other molecules it will bind to and how. And if you actually wanted to do things with these molecules, which most of them did, you had to find out through more laborious modeling and testing.
“Biology is a dynamic system, and it is necessary to understand how the properties of biology emerged through the interaction between the different molecules of the cell. And you can think of AlphaFold 3 as our first big step toward that,” Hassabis said. “It is capable of modeling proteins interacting, of course, with other proteins, but also with other biomolecules, notably strands of DNA and RNA.”
AlphaFold 3 allows you to simulate multiple molecules at once – for example, a strand of DNA, some DNA-binding molecules, and maybe a few ions to spice things up. Here’s what you get for such a specific combination, with the DNA ribbons running up the middle, the proteins glowing on the side, and I think it’s the ions nestled in the middle like little eggs:
Of course, this does not constitute a scientific discovery in itself. But even understanding that an experimental protein would bind, or this way, or contort itself into this shape, was usually the work of several days, or even weeks, or even months.
While it is difficult to exaggerate the enthusiasm generated by this area in recent years, researchers have been largely paralyzed by the lack of interaction modeling (of which the new version offers one form) and by the difficulty of deploying the model.
This second problem is perhaps the more important of the two, because even if the new modeling techniques were “open” in some sense, like other AI models, they are not necessarily simple to deploy and operate. That’s why Google Deepmind offers AlphaFold Server, a free, fully hosted web application making the model available for non-commercial use.
It’s free and pretty simple to use – I did it in another window on the call while they explained it (that’s how I got the image above). You just need a Google account, then you feed it as many sequences and categories as it can handle (a few examples are provided) and submit it; Within a few minutes your work should be complete and you will receive a living 3D molecule colored to represent the model’s confidence in the conformation at that position. As you can see in the one above, the ends of the ribbons and the parts most exposed to unwanted atoms are lighter or red to indicate less trust.
I asked if there was a real difference between the publicly available model and the one used internally; Hassabis said, “We have made the majority of the capabilities of the new model available,” but did not provide further details.
It’s clearly Google putting all its weight behind it, while to some extent keeping the best for itself, which is of course its prerogative. Creating a free, hosted tool like this means devoting considerable resources to this task. Make no mistake, this is a money pit, an expensive (for Google) shareware release intended to convince researchers around the world that AlphaFold 3 should be, at the same time. at least one arrow in their quiver.
Image credits: Google Deep Mind
That’s okay, though, because the technology will likely print money through Alphabet subsidiary (making it Google’s…cousin?) Isomorphic Labs, which puts computing tools like AlphaFold to work drug design. Well, everyone uses computational tools these days – but Isomorphic got first access to Deepmind’s latest models, combining them with “more exclusive things related to drug discovery,” as Hassabis noted . The company already has partnerships with Eli Lilly and Novartis.
AlphaFold, however, is not the be-all and end-all of biology – just a very useful tool, as countless researchers will agree. And that allows them to do what Isomorphic’s Max Jaderberg called “rational drug design.”
“If we think on a day-to-day basis about the impact that this has on isomorphic laboratories: it allows our scientists, our drug designers, to create and test hypotheses at the atomic level, and then produce in seconds very precise structure predictions… to help scientists think about what interactions to make and how to advance those designs to create a good drug,” he said. “That’s compared to the months or even years it would take to do it experimentally.”
While many will celebrate the success and wide availability of a free hosted tool like AlphaFold Server, others may rightly point out that this is not really a victory for open science.
Like many proprietary AI models, AlphaFold’s training process and other information crucial to its reproduction – a fundamental part of the scientific method, remember – is largely and increasingly hidden. Although the paper published in Nature examines the creation methods in some detail, it is missing many important details and data, meaning that scientists wanting to use the most powerful molecular biology tool on the planet will have to do so under cover of law. the watchful eye of Alphabet, Google and Deepmind (who knows who really holds the reins).
Open science advocates have argued for years that while these advances are remarkable, in the long run it is always better to share this stuff openly. After all, that’s how science advances, and it’s also how some of the world’s most important software evolved.
Making AlphaFold Server free for any academic or non-commercial application is in many ways a very generous act. But Google’s generosity is rarely unconditional. No doubt many researchers will nonetheless take advantage of this honeymoon period to use the model as much as humanly possible before the other shoe drops.
techcrunch
