Three and a half years after Google DeepMind engineers surprised the world of biomedicine with the artificial intelligence (AI) program AlphaFold 2, which predicts the structure of any protein, today its successor, AlphaFold 3, was presented, which raises the potential of AI in biomedical research at a higher level.
The new version of AlphaFold is no longer limited to studying proteins but can decipher any molecule of any living being, including nucleic acids such as DNA and RNA and tiny molecules such as ligands, which are decisive in the functioning of cells. And it is not satisfied with describing what these molecules are like on an atomic scale, but also predicts how they interact with each other, and how these interactions modify them.
Structure of a coronavirus protein determined with AlphaFold 3 (image provided by Google DeepMind)
With this advance it is expected, on the one hand, to advance the scientific understanding of how living beings work and, on the other, to transform the way drugs are created, reported yesterday at a press conference, Demis Hassabis, co-founder of DeepMind before it was acquired by Google and current CEO of the company.
Coinciding with the launch of AlphaFold 3, Google DeepMind has made the AlphaFold Server available free of charge to scientists around the world, a tool that allows biomedical researchers to use the AI platform without the need for advanced computer science knowledge.
It will have two main applications: creating drugs and better understanding how living beings work.
“All the structural biology groups in the world are going to adopt this system. It is very easy to use. It is going to change the way we do science,” said Julien Bergeron, a researcher at King’s College London who collaborates with the Google DeepMind team, at the press conference.
Since AlphaFold2 was introduced in November 2020 and made available to the scientific community, it has been used by almost two million users. It has been used in projects ranging from the development of new antibiotics or a vaccine against malaria to carbon capture with bacteria to mitigate climate change. He was recognized with the Breakthrough of the Year (the most important scientific advance of the year) by the magazine Science and its creators have already received some of the most important scientific awards in the world such as -among others- the Albert Lasker (considered the prelude to the Nobel Prize), the Frontiers of Knowledge from the BBVA Foundation or, more recently, the Breakthrough Prize.
Structure of a fungal enzyme determined with AlphaFold 3 (image provided by Google DeepMind)
AlphaFold 2 made it possible to deduce for the first time the three-dimensional structure of proteins from their amino acid sequence. In this way, investigations that could require years for each individual protein, in addition to expensive equipment such as X-ray crystallography machines, could be solved in a matter of minutes with a computer.
But AlphaFold 2 resolved “the static image of proteins,” while “biological phenomena are dynamic,” Demis Hassabis explained yesterday. To understand biology, “you have to understand […] the interactions between different molecules in cells. AlphaFold 3 is our next step in this direction.”
AlphaFold 3 has been made freely available to scientists around the world
One of its most relevant novelties is that it can determine how proteins bind to DNA or RNA, something that AlphaFold 2 could not do. “When I give lectures, there are people who tell me ‘that’s very good, but I’m researching a protein that binds to DNA, can you tell me how it binds?’; and I couldn’t,” explained John Jumper, who has led the development of the new AI.
AlphaFold 3, which is presented today in the scientific journal Naturededuces the structure of biomolecules based on amino acid sequences just like AlphaFold 2. But the strategy it uses is different.
Demis Hassabis, CEO of Google DeepMind, at the company’s headquarters in London
Jose Sarmento Matos/Bloomberg
The new AI has been developed by Google DeepMind and the pharmaceutical company Isomorphic Labs
AlphaFold 2 was based on data on how proteins have evolved to get the shape they have. The new version uses less evolutionary data and completes it with a generative AI strategy, similar to that used to generate images with AI. In this case, the system starts from a cloud of atoms and, through successive iterations, approximates the position of each atom in space until reaching the final three-dimensional structure of the biomolecule.
Researchers from Isomorphic Labs have participated in the development of AlphaFold 3, a company born as a subsidiary of Google DeepMind that uses AI to design new drugs by computer. Proof of AlphaFold’s potential for the pharmaceutical industry is that the multinationals Eli Lilly and Novartis have already signed collaboration contracts with Isomorphic Labs, worth $45 million for the first and $37.5 million for the second, to apply AI to the pharmaceutical industry. drug design.
“With the new tool, you can design a compound that binds to the surface of a protein and predict how strong the binding will be, which is critical for drug design,” explained Demis Hassabis, CEO of Google DeepMind. The effectiveness of antibodies, for example, depends on this type of binding, so AlphaFold 3 is expected to accelerate the development of new antibody-based therapies.
Marc Masip
