Alphafold: Artificial intelligence predicts interactions between all molecules of life | The science
Demis Hassabis, the London-raised son of a Greek Cypriot and a Singaporean, is a chess prodigy. He started playing at the age of four, and by the age of 13 he was already a master. He studied computer science, earned a PhD in neuroscience, and founded Deepmind, which is now the artificial intelligence (AI) backbone of the company that owns Google. A few days ago, 47-year-old Hassabis recalled in an interview the day when he fully realized the irresistible power of this technology. Over coffee one morning in 2018, he played AlphaZero, the chess AI he created. He was able to win it without any problems. Within hours, the program, which had trained itself by playing hundreds of thousands of games, was on the verge of defeating him. At night he was “the best chess player who ever lived.” And all this in nine hours. “It was amazing to see it live. It was inevitable to ask himself, “What can this system do in science or any other complex problem?” he explains.
Game time is over. Since 2020, Alphafold, an artificial intelligence developed by Hassabis, has surpassed any human in the diabolical tasks of biology and determined the three-dimensional structure of 200 million proteins, virtually all known. Solving the shape of a single protein can take a graduate student several years of dedication, so AI would immediately save about a billion years of work.
Yesterday, the businessman held a press conference to introduce his new creation: Alphafold 3. For the first time, AI can predict the interactions between proteins and other molecules essential to life: DNA and RNA, small molecules – drugs – and antibodies. , tiny proteins of the immune system that specialize in fighting viruses, bacteria and even tumors. “Biology is a dynamic system, and its properties arise precisely as a result of the interactions between the various molecules that make up the cell. Alphafold 3 is our first big step in understanding them,” explained Hassabis. Details of this new artificial intelligence system were published today in Nature, a guide to the world’s best science. Google will also open a free server so scientists can use this new tool.
The most obvious implication of the new system is its potential for discovering new drugs, an area that will henceforth be explored, this time in the private sphere, by Laboratorios Isomorphic, a subsidiary of Alphabet (the owner of Google), headed by Hassabis. With Alphafold 3 and his own additional developments, a Google scientist hopes to halve the time it takes to discover a drug before testing in patients, from the current five years to two, he explained to the publication Financial Times. The company has already signed two cooperation contracts with multinational corporations Lilly and Novartis, which have invested tens of millions of dollars up front and promise several billion more when the results are in.
American chemist John Jumper, director of Deepmind, yesterday assured that the new system is far superior to its competitors. Alphafold 3 successfully predicts 76% of protein-small molecule interactions, compared with 52% for the next best tool, he noted. By binding proteins to DNA or interacting with antibodies, it doubles the capabilities of traditional methods. New AI provides a new level of prediction about what happens inside cells and what goes wrong when DNA damage occurs. “This has implications for understanding cancer and developing new treatments,” Jumper said, as well as understanding plant responses to pathogens and heat waves needed to ensure food security.
According to Max Jaderberg, head of artificial intelligence at Isomorphic Labs, the level of complexity to be explored with this new system is “absolutely enormous.” If we talk only about the small molecules that are most interesting in pharmacology, there are about 10 of them to the sixtieth power, many times more than there are stars in the entire Universe.
Facing this Goliath of artificial intelligence, American biochemist David Baker lives up to his name. The researcher is leading a public and fully open initiative to create artificial intelligence that can predict biochemical processes and invent new compounds with specific properties.
Two months ago, without any media pressure, Baker published The science its new AI, which reconstructs molecules atom by atom, predicts their binding to DNA and develops new compounds that do not exist in nature. “The creators of Alphafold 3 say it is more accurate than our system,” Baker explains in an email. “I think it will have a big impact, even though they can’t make new proteins,” he adds.
A researcher from the University of Washington (USA) highlights another important difference. “Deepmind does not publish the code (of its AI), which is different from normal practice in science,” he points out. Knowing the AI’s codebase allows the community to modify it and provide it with new capabilities, while the server only allows it to be used within the limits set by its creator.
Like other artificial intelligence systems such as ChatGPT, Alphafold has hallucinations– he invents some results – especially when describing things for which there is no information in the large databases with which he trains.
A human protein can be a veritable microscopic monster, with tens of thousands of amino acids—its basic building blocks—that fold in on themselves, forming hooks, rings, clamps and other moving parts that change position when the protein binds to another molecule. The new AI is particularly impressive at characterizing the “disordered zones” of proteins, regions without a fixed three-dimensional shape that are essential to understanding these interactions. “They are like the dark matter of proteins,” sums up Mafalda Diaz, a researcher at the Center for Genomic Regulation in Barcelona, comparing these regions to an unknown ingredient that makes up 25% of the universe. “The model was trained using static 3D structures, but since biology is dynamic, the output it suggests may be very different from reality. The creators of Alphafold were very open about these and other limitations,” emphasizes the Portuguese scientist.
Biologist Rafael Fernandez Leiro, who has dedicated his academic and professional life to the study of structural biology through crystallography and electron microscopy, gives an example to help understand the potential of the Alphafold 3 discovery. “Inside cells there is a very complex cocktail of proteins, nucleic acids, lipids, specialized proteins such as enzymes that allow DNA to be copied and read and, in turn, to produce other proteins. Until now we could only establish the structure of isolated proteins, but now we can study them associated with other proteins, with DNA, RNA, even explore what happens if we modify the set with a phosphate residue or phosphorylation (epigenetic modification), which changes the function of the entire set. This will be an amazing hypothesis generator. Of course, eventually the results will have to be confirmed using traditional methods, and here comes the difficult part, because if this system is correct almost 80% of the time, that means that it fails 20%, and this percentage is too high to spend time testing the drug on patients. But this will be of great importance in the first stages of searching for new drugs,” he clarifies.
Navarrean bioinformatician Iñigo Barrio, working at the Wellcome Sanger Institute (UK), highlights this AI’s new ability to explain how proteins join together or with other molecules to form complexes. A classic example is how hemoglobin binds to its ligand oxygen and transports it throughout the body. “The most relevant will be the ability to predict the binding of various ligands to proteins. This allows us to understand where and how the drug will bind to the target protein, how it will affect its biology and potential unwanted effects on other proteins,” he emphasizes.
In an interview he gave a few days ago for TED, Hassabis was asked what the next big problem he wants to solve with AI. The neuroscientist responded that when artificial general intelligence capable of solving many different problems at once is created, he would like to use it to understand nature at its most fundamental level, at the Planck level, where mind-blowing quantum phenomena occur. . “It’s sort of like a reality resolution scale,” he said.
“Nature should not publish such research”
Alfonso Valencia, director of the biological sciences division at the National Supercomputing Center, is critical of Deepmind’s new contribution, although he admits that “it certainly represents important progress.” “They show that a general model for predicting macromolecular complexes is possible, compare it to previous methods, mainly by David Baker, its only serious competitor, and show significant improvements, albeit based on a few cases, dozens in each category, which is significant. less reliable results,” he notes. “The obvious problem is that when they are offered on the server, users will tend to ignore the restrictions and assume the results are reliable in all cases. This problem is not new, and previous structure prediction servers have already suffered from incorrect interpretations. Now, with new methods that are more popular, powerful and visible, the problem will be worse. Although you can use this method as a web server, they do not publish the software. This is a mistake, and Nature should not publish studies with such results. They cannot be reproduced or independently verified. Whether or not to believe the results they present cannot be a matter of faith,” he says.
The expert continues: “Finally, we will see how the academic world can adapt to these new changes and how long it will take to have equivalent open and public methods. If we go by previous cases like Alphafold 2, I’d say: very little. Then we will have a more reliable and independent assessment of opportunities and limitations.”
You can follow ITEM V Facebook, X And instagramor register here to receive our weekly newsletter.