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 13 he was already a master. He studied computer science, earned a doctorate in neuroscience and founded Deepmind, which is currently the mainstay of artificial intelligence, or AI, of the company that owns Google. A few days ago, Hassabis, 47, recalled in an interview the day when he became fully aware of the unstoppable power of this technology. One morning in 2018, while having coffee, he played AlphaZero, the chess AI he had created. He was able to win it without much trouble. In a few hours, the program, which taught itself by playing hundreds of thousands of games, was on the verge of defeating it. By night he was “the best chess player who ever lived.” All in just nine hours. “It was amazing to see it live. It was inevitable to wonder what this system can do in science or any other complex problem,” he explains.
Game time is over. Since 2020, Alphafold, the artificial intelligence devised by Hassabis, has surpassed any human in devilish biology problems and determined the three-dimensional structure of 200 million proteins, practically all those known. Solving the shape of a single protein can take several years of dedication from a PhD student, so AI would have saved about a billion years of work at once.
The businessman held a press conference yesterday to present his new creation: Alphafold 3. For the first time, an AI can predict the interaction between proteins and the rest of the essential molecules of life: DNA and RNA, small molecules—medicines—and antibodies. , the tiny proteins of the immune system specialized in fighting viruses, bacteria, even tumors. “Biology is a dynamic system and its properties emerge precisely from interactions between the different molecules that make up the cell. Alphafold 3 is our first big step in understanding them,” Hassabis explained. The details of this new AI system are published today in Nature, a reference for the best world science. The Google company will also open a free server so that scientists can use this new tool.
The most obvious derivative of the new system is its potential to discover new drugs, a field that from now on will be explored, this time in the private sphere, by Laboratorios Isomorphic, a subsidiary of Alphabet (owner of Google) led by Hassabis. With the help of Alphafold 3 and additional developments of his own, the Google savant hopes to halve the time it takes to discover a drug before starting trials in patients, from the current five years to just two, as he explained to Financial Times. The company has already signed two collaboration contracts with the multinationals Lilly and Novartis, which have contributed tens of millions of dollars upfront and promise several billion more when there are results.
The American chemist John Jumper, director of Deepmind, assured yesterday that the new system is far superior to its rivals. Alphafold 3 successfully predicts 76% of protein-small molecule interaction cases, compared to 52% for the next best tool, he noted. In the binding of proteins to DNA or their interaction with antibodies, it doubles the capacity of conventional methods. New AI enables a new level of prediction about what happens inside cells and what goes wrong when there is DNA damage. “This has implications for understanding cancer and developing new therapies,” Jumper highlighted, but also for understanding the response of plants to pathogens and heat waves, essential to guarantee food security.
The level of complexity to be explored by this new system is “absolutely enormous”, in the words of Max Jaderberg, head of AI at Isomorphic Laboratories. When it comes to small molecules alone, the most interesting in pharmacology, there are about 10 to the sixtieth power, many times more than stars in the entire universe.
Faced with this Goliath of AI, the American biochemist David Baker lives up to his name. The researcher leads a public and completely open initiative to create artificial intelligences capable of predicting biochemical processes and inventing new compounds with specific properties.
Two months ago, without any media impact, Baker published in Science its new AI, which reconstructs molecules atom by atom, predicts their binding to DNA, and designs new compounds that did 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 great impact, although they cannot design new proteins,” he adds.
The researcher from the University of Washington (United States) highlights another important difference. “Deepmind does not publish the code [de su IA], which differs from usual practice in science,” he highlights. Knowing the code base of an AI allows the community to modify it and give it new capabilities, while a server only allows it to be used within the limits set by its creator.
Like other AI systems like ChatGPT, Alphafold has hallucinations,—he invents some results—especially when he describes 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 back on themselves to form hooks, ringlets, clamps, and other moving parts that change position when a protein binds to another molecule. The new AI is especially impressive when describing the “disordered zones” of proteins, regions without a fixed three-dimensional shape, which are essential to understanding these interactions. “They are like the dark matter of proteins,” summarizes Mafalda Dias, a researcher at the Center for Genomic Regulation in Barcelona, comparing these regions with the unknown ingredient that makes up 25% of the universe. “The model has been trained with static three-dimensional structures, but since biology is dynamic, the result it proposes can be very different from reality. The creators of Alphafold have been very frank about these and other limitations,” highlights the Portuguese scientist.
Biologist Rafael Fernández Leiro, who has dedicated his academic and professional life to the study of structural biology through crystallography and electron microscopy, gives an example to understand the discovery potential of Alphafold 3. “Within the 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 produce other proteins. Until now we could only establish the structure of isolated proteins, but now we can study them bound to others, to DNA, RNA, even explore what happens if we modify the set with a phosphate residue, or phosphorylation [una modificación epigenética] that changes the function of the entire set. This is going to be an amazing hypothesis generator. Of course, in the end the results will have to be confirmed using conventional techniques and here comes the difficult part, because if this system is correct in almost 80% of the cases, it means that it fails in 20%, a percentage that is too high to spend the time. how much money it costs to bring a drug to trials with patients. But it will make a big difference in the first phases of searching for new drugs,” he details.
The Navarrese bioinformatician Íñigo Barrio, who works at the Welcome Sanger Institute (United Kingdom), highlights the new capacity of this AI to explain how proteins join together or to other molecules to form complexes. A classic example is how hemoglobin binds to its ligand, oxygen, to transport it throughout the body. “The most relevant thing will be the ability to predict the binding of various ligands to proteins. “It allows us to understand where and how a drug is going to bind to the target protein, how it is going to affect its biology and the potential unwanted effects on other proteins,” he highlights.
In the interview he gave a few days ago for TED, Hassabis was asked what is the next big problem he wants to address with AI. The neuroscientist replied that when General Artificial Intelligence has been built—capable of solving many different problems at the same time—he would like to use it to understand nature at its most fundamental level, at the Planck scale, where mind-blowing quantum phenomena occur. . “It’s kind of like the resolution scale of reality,” he noted.
“‘Nature’ should not publish studies like this”
Alfonso Valencia, director of life sciences at the National Supercomputing Center, is critical of Deepmind’s new contribution, although he recognizes that “it undoubtedly represents important progress.” “They show that a general prediction model for macromolecule complexes is possible, compare it with previous methods, mainly by David Baker, its only serious competitor, and show significant improvements, although based on a few cases, dozens in each category, which makes much less reliable results,” he points out. “The obvious problem is that when offered on a server, users will tend to ignore the limitations and take the results as reliable in all cases. This problem is not new and previous structure prediction servers already suffered from abusive interpretations. Now, With new methods that are more popular, powerful and visible, the problem will be worse. Although you can use the method as a web server, they do not make the software public. This is a mistake and ‘Nature’ should not publish studies with results that. They cannot be reproduced or independently validated. It cannot be a matter of faith whether or not to believe the results they present,” he says.
The expert continues: “Finally, we will see how the academic world can adapt to this new change and how long it will take to have equivalent open and public methods. If we base ourselves on previous cases, such as Alphafold 2, I would say very little. Then we will have a more reliable and independent assessment of capabilities and limitations”.