Google Deepmind launches huge AlphaFold update and free proteomics-as-a-service web app

Google Deepmind has unveiled a new version of AlphaFold, their transformative machine learning model that can predict the shape and behavior of proteins. Not only is AlphaFold 3 more accurate, but it can predict interactions with other biomolecules, making it a more versatile research tool — and the company is making a limited version of the model available for free online.

Since the first AlphaFold debuted in 2018, the model has been the leading method for predicting protein structure based on the sequence of amino acids that make up the protein.

While this may sound like a rather narrow task, it is fundamental to almost all of biology in understanding proteins at the molecular level – proteins that perform an almost endless variety of tasks in our bodies. In recent years, computational modeling techniques such as AlphaFold and RoseTTaFold have replaced expensive laboratory-based methods, accelerating the work of thousands of researchers in multiple fields.

But the technology is still very much a work in progress, and as Deepmind founder Demis Hassabis said at a press conference about the new system, each model is “just a step forward.” The company teased the launch late last year, but this marks its official debut.

I’ll let the Science Blog go into detail about how the new model improves the results, but suffice it to say here that various improvements and modeling techniques make AlphaFold 3 not only more accurate, but also more broadly applicable.

One of the limitations of protein modeling is that even if you know the shape of an amino acid sequence, that doesn’t mean you necessarily know which other molecules it will bind to and how. If you want to actually do something with these molecules (which most molecules do), you need to find out through more laborious modeling and testing.

“Biology is a dynamic system, and you have to understand how the properties of biology work through interactive between different molecules within cells. You can think of AlphaFold 3 as our first step toward that goal. “It can model the interactions of proteins with other proteins and other biomolecules, important among them DNA and RNA strands,” said Hassabis. “

AlphaFold 3 allows the simulation of multiple molecules simultaneously – for example, a DNA strand, some DNA-binding molecules, and maybe some ions to spice things up. This is the result of such a specific combination, with the DNA band in the middle and the protein on the sides, I think of these ions nestled in the middle like little eggs:

Of course, this is not a scientific discovery in itself. But even figuring out whether an experimental protein binds, or binds this way, or twists into this shape usually takes at least days or weeks to months.

While it’s hard to overstate the level of excitement in the field over the past few years, researchers have been largely hampered by a lack of interactive modeling (a form of which the new version provides) and the difficulty of deploying models.

The second problem is probably the greater of the two, because while new modeling techniques are in some sense “open,” they, like other AI models, are not necessarily easy to deploy and operate. That’s why Google Deepmind offers AlphaFold Server, a free, fully managed web application that makes the model available for non-commercial use.

It’s free and very easy to use – I did it in another window on the call while they explained it (that’s how I got the picture above). All you need is a Google account, then you feed it as many sequences and categories as you like (some examples are provided) and submit; after a few minutes, your work is done and you get a live 3D molecule with the colors representing The model’s confidence in the conformation at that position. As you can see in the image above, the tips of the ribbon and those parts that are more exposed to rogue atoms are lighter or red, indicating lower confidence.

I’m asking if there are any real differences between the public model and the model used internally? Hassabis said “we have delivered most of the features of the new model” without elaborating.

Google is clearly exerting its influence while, to a certain extent, keeping the best bits for themselves, which of course is their prerogative. Making such a free hosting tool requires devoting a lot of resources to the task – no doubt it’s a money pit, an expensive (for Google) shareware version meant to be used by researchers around the world I believe AlphaFold 3 should be the most valuable. At least they have an arrow in their quiver.

Image Source: Google deep thinking

That’s okay, though, because the technology will likely be printed via Alphabet subsidiary (making it Google’s…cousin?) Isomorphic Labs, which is using computational tools like AlphaFold for drug design. Well, everyone is using computational tools these days—but Isomorphic cracked Deepmind’s latest model for the first time, combining it with “some more proprietary stuff related to drug discovery,” as Hassabis points out. The company already has partnerships with Eli Lilly and Company and Novartis.

However, AlphaFold is not the be-all and end-all of biology, just a very useful tool, as countless researchers would agree. It allows them to engage in what Isomorphic’s Max Jaderberg calls “rational drug design.”

“If we think about it every day what does this mean for isomorphism labs: It enables our scientists, drug designers, to create and test hypotheses at the atomic level and then produce highly accurate structure predictions in seconds…. ..to help scientists think about what interactions to have and how to advance those designs to create a good drug,” he said. “In contrast, experiments can take months or even years.”

While many will celebrate this achievement and the widespread availability of free hosting tools like AlphaFold Server, others may rightly point out that this is not a true victory for open science.

As with many proprietary AI models, AlphaFold’s training process and other information critical to its replication (which, you may recall, is a fundamental part of the scientific method) have been largely withheld, and increasingly many. While the paper published in Nature does detail how it was created, it lacks many important details and data, meaning scientists who want to use the most powerful molecular biology tool on the planet will have to work with Alphabet, Google and Deepmind is watching closely (who knows who exactly is in control).

Open science advocates have said for years that while these advances are notable, it is always better in the long run to share such things openly. After all, this is how science moves forward, and indeed how some of the most important software in the world develops.

Making AlphaFold Server freely available to any academic or non-commercial application is a very generous act in many ways. But Google’s generosity rarely comes without strings attached. No doubt many researchers will still take advantage of this honeymoon period to get as much use out of the model as possible before the other shoes drop.