Google unveils AlphaFold 3 powerful neural network predicting protein structure: It is available to everyone

Google DeepMind division has unveiled a new version of its artificial intelligence model, AlphaFold, which predicts the shape and behavior of proteins. AlphaFold 3 not only boasts higher accuracy but now also predicts the interaction of proteins with other biological molecules; furthermore, its limited version is now available for free as a web application.

Since the release of the first AlphaFold neural network in 2018, it has become the leading method for predicting protein structures based on sequences of amino acids from which they are composed. Understanding the structure and fundamentals of protein interactions underpins almost all biology. Classical protein modeling methods have significant limitations: even knowing the shape a sequence of amino acids will take, it is impossible to predict in advance with which other molecules it will bind and how. And if a practical goal needs to be achieved, meticulous work in modeling and testing is required - previously taking several days, and sometimes even weeks and months.

AlphaFold tackles this task by predicting the likely shape of a protein molecule given a sequence of amino acids, indicating with which other proteins it may interact. The distinctive feature of the new AlphaFold 3 lies in its ability to predict the interaction of proteins with other biological molecules, including DNA and RNA chains, as well as the ions necessary for this.

A significant challenge for AlphaFold, like other AI-based tools, is the complexity of deployment. Therefore, Google DeepMind has launched the free AlphaFold Server web application - available for non-commercial use. The platform is relatively easy to use: after logging in with a Google account, users can input several sequences and categories, after which it will output the result in the form of a three-dimensional molecule, colored to reflect the model's confidence in its correctness. When asked if there is a significant difference between the publicly available version of the model and the one used internally by the company, the head of Google DeepMind's division, Demis Hassabis, assured that "we have made most of the new model's features available," but did not provide further details.