Google DeepMind has lifted the curtain on AlphaFold 3, the latest iteration of their groundbreaking machine learning model that predicts the shape and behavior of proteins. Not only is AlphaFold 3 more accurate than its predecessors, but it also predicts interactions with other biomolecules, making it a far more versatile research tool. In a significant move, the company is offering a limited version of the model free for online use.
Since the debut of the first AlphaFold in 2018, the model has remained the leading method of predicting protein structure from the sequence of amino acids that make them up. While this may seem like a narrow task, it is foundational to nearly all biology, as understanding proteins — which perform a nearly endless variety of tasks in our bodies — at the molecular level is crucial.
In recent years, computational modeling techniques like AlphaFold and RoseTTaFold have taken over from expensive, lab-based methods, accelerating the work of thousands of researchers across numerous fields. However, as DeepMind founder Demis Hassabis acknowledged, the technology is still a work in progress, with each model being “just a step along the way.”
While the specifics of how AlphaFold 3 improves outcomes are better left to science blogs, the new model is not just more accurate but also more widely applicable. One of the limitations of protein modeling is that even if you know the shape a sequence of amino acids will take, you don’t necessarily know what other molecules it will bind to and how. AlphaFold 3 addresses this by allowing multiple molecules to be simulated at once, including proteins, DNA, RNA, and even ions.
This capability, which DeepMind describes as its “first big step towards” understanding the interactions between different molecules in the cell, could significantly accelerate research in fields ranging from drug design to molecular biology.
While the excitement in this field has been palpable in recent years, researchers have been hamstrung by the lack of interaction modeling (which AlphaFold 3 now offers) and the difficulty in deploying the model.
To address the latter issue, Google DeepMind is offering AlphaFold Server, a free, fully hosted web application making the model available for non-commercial use. Users with a Google account can feed the server as many sequences and categories as it can handle, and within minutes, they’ll receive a live 3D molecule colored to represent the model’s confidence in the conformation at each position.
While DeepMind has made “the majority of the new model’s capabilities available” through AlphaFold Server, the company is likely keeping some of the best features for internal use. This move is part of Google’s strategy to position AlphaFold 3 as an essential tool for researchers worldwide, with the potential to generate revenue through Alphabet subsidiary Isomorphic Labs, which is applying computational tools like AlphaFold to drug design.
Although AlphaFold 3 is a remarkable achievement and the free hosted tool is a generous offering, some open science advocates have raised concerns. As with many proprietary AI models, the training process and other crucial information necessary for replicating AlphaFold are largely withheld, meaning scientists who want to use this powerful molecular biology tool will have to do so under the watchful eye of Alphabet, Google, and DeepMind.
While the AlphaFold Server’s non-commercial availability is undoubtedly a boon for researchers, it remains to be seen whether Google’s generosity will come without strings attached. Nevertheless, many researchers are likely to take advantage of this opportunity to utilize the model extensively before any potential limitations are imposed.
