Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Google DeepMind has unveiled an artificial intelligence model of life’s building blocks and their interactions within cells, boosting efforts to unlock secrets of disease and find treatments for conditions such as cancer.
AlphaFold 3, the third generation of technology initially developed in 2018, gives the most sophisticated forecasts yet of how tiny biological structures look and mingle, according to a paper published in Nature on Wednesday.
The model, developed with the DeepMind drug discovery spin-off Isomorphic Labs, is the latest landmark in the quest to apply the predictive power of AI to improve understanding of life’s miniature mechanisms and how they go wrong.
“Biology is a dynamic system and you have to understand how properties of biology emerge through the interaction between different molecules in the cells,” said Sir Demis Hassabis, DeepMind’s chief executive and co-founder. “And you can think of AlphaFold 3 as our first big step towards that.”
The technological update expands its biological purview beyond the proteins it has previously analysed, offering a richer view of the biochemical networks that make organisms function. The model covers the genetic code DNA and RNA as well as ligands — molecules that bind to others and can be important markers of disease.
AlphaFold 3’s capabilities open up fresh opportunities for researchers to speedily identify potential new drug molecules, said Max Jaderberg, Isomorphic Labs’ chief AI officer. Isomorphic Labs has partnerships with pharmaceutical companies Eli Lilly and Novartis.
“That allows our scientists, our drug designers, to create and test hypotheses at the atomic level, and then within seconds produce highly accurate structure predictions with AlphaFold 3,” Jaderberg said. “This is compared to the months or even years it might take to do this experimentally.”
AlphaFold 3 demonstrates “significantly improved” predictive accuracy over many existing specialised tools including those based on its own predecessors, the paper says. It shows that developing the right AI deep learning frameworks can greatly reduce the amount of data needed to obtain “biologically relevant performance”, the research adds.
“We’re seeing really incredible improvements that we think are going to unlock a lot of new science,” said John Jumper, DeepMind’s AlphaFold team leader, who cited the potential of the technique to improve knowledge of plant biology and thus food security. “We’re already starting to see biologists and early testers use this to understand how the cell works — and start to think about how it might go wrong when in disease states.”
The molecules AlphaFold 3 suggests will still need to be validated experimentally and go through the normal process of clinical trials. DeepMind says it is making the majority of AlphaFold 3’s functionality available through a server that will be free to access for academic non-commercial users.
A study by Boston Consulting Group published this week suggests that drugs discovered by AI have a higher success rate in early stage trials than those discovered by other methods. Cautioning that the data was an early analysis of the technology’s effectiveness in drug discovery, researchers said AI could double the productivity of pharma research and development.
The server promises to change the way people do experiments, said Julien Bergeron, a structural biologist at King’s College London, who was not involved in the development of AlphaFold 3 but has been a test user of it.
“We can start testing hypotheses before we even go to the lab,” he said. “This will really be transformative.”
AlphaFold 3’s limitations include difficulties dealing fully with chiral — or mirror-image — molecules, as well as “hallucinations” of “spurious structural order” in areas that are in fact disordered. One remedy the model uses is to assign confidence measures to predictions, to reflect the likelihood of error.
With additional reporting by Ian Johnston
Read the full article here