At the heart of every biological process, from the replication of cells to the functioning of the human brain, lies the intricate world of proteins. Understanding the structure of these molecules, folded into complex three-dimensional shapes, is crucial to advance our collective understanding of biology and facilitate the development of new therapeutics.
AlphaFold is an artificial intelligence system that revolutionized our understanding of protein structures advancing research and drug discovery. The project was led by Google DeepMind CEO Demis Hassabis and Director John Jumper.
Drug development is an extraordinarily complex and time-consuming business. The formidable goal of treating disease demands unwavering scientific rigor across a sprawling set of variables and unknowns, in addition to huge amounts of funding. Still, it can be difficult to imagine just how demanding the creation of new medicines is. So, to illustrate the impact of AlphaFold, Jumper often tells a particular story.
“I went to a talk by someone who had designed a very interesting drug in clinical trials at a major pharmaceutical company,” Jumper recounts. “They described this beautiful molecule they had made and all the chemistry they had done to solve different problems with their molecule. Someone asked them afterward, ‘How many different molecules did you make to come up with this one?’ And he said, ‘5,000 over a seven-year period.’”
The story illustrates a fundamental fact: Science is an iterative enterprise. Trial and error in the laboratory—a word befitting its Latin root to labor—is how progress has traditionally been made. But now, AI systems are poised to dramatically accelerate this painstaking process.
AlphaFold application
For decades, scientists puzzled over the protein-folding problem of whether a protein’s structure and shape could be determined based on that protein’s amino acid sequence. In 2020, AlphaFold made a fundamental breakthrough: predicting the structure of proteins with a level of accuracy that had never been seen before in competitions run over the years on the Critical Assessment of protein Structure Prediction (CASP). Together with the European Molecular Biology Laboratory’s European Bioinformatics Institute, Google DeepMind shared these predictions for 200 million proteins in a free database that’s been used by millions of researchers globally.
When Google DeepMind’s AlphaFold was recognized by CASP organizers as a solution to the 50-year-old protein-folding problem, it represented a watershed moment for computational and AI methods for biology. AlphaFold demonstrated that AI techniques were advanced enough to be applied to real-world problems, including scientific discovery. In 2021, Hassabis founded Isomorphic Labs as an offshoot of Google DeepMind to specifically focus on rational drug design and apply AI to redefine the drug discovery process from first principles.
The latest version, AlphaFold 3, developed by Google DeepMind and Isomorphic Labs, is reshaping the field of structural biology and providing invaluable insights for drug design. The third iteration of AlphaFold expands beyond proteins, enabling it to predict the structures of a wide range of biological molecules—DNA, RNA, ligands, and more—with unprecedented accuracy. More importantly, it can model how these different types of molecules interact, a key piece of the puzzle for understanding disease and developing treatments.
Hassabis is also the CEO of Isomorphic Labs, which is applying these advancements to the field of drug discovery. The preclinical phase of drug development has traditionally relied on traditional experimental methods—X-ray crystallography and cryo-electron microscopy (cryo-EM)—to determine the structures of target proteins and potential drug molecules. Using X-ray crystallography, scientists must coax proteins to form crystals, then bombard them with X-rays to deduce their structure. Cryo-EM, meanwhile, involves an exceptionally intricate process of freezing proteins before imaging them with powerful electron microscopes. Both techniques are resource heavy and often take months or years to yield results. But Isomorphic Labs combines AlphaFold 3 with its own suite of AI models to generate high-confidence structure predictions in a matter of minutes.
Bringing molecular structures to life
This otherwise-lengthy process is optimized thanks to AlphaFold 3’s sophisticated artificial neural network architecture, which has been fine-tuned to predict molecular structures with unprecedented accuracy. One of the key components of this architecture is the Pairformer, a deep learning system that employs multiple types of information—including evolutionary information, or data on how proteins have changed over millions of years—to improve its predictions.
To transform these structural predictions into visible forms, AlphaFold 3 uses an approach called diffusion modeling. Similar to techniques used in popular AI-image generation platforms, diffusion lets the system start from a rough structural “sketch” and gradually refine it into a detailed, increasingly accurate final image. To ground predictions, AlphaFold provides a confidence score, giving researchers an estimate of the model’s accuracy.
Advancing drug design
One area in which AlphaFold 3 shows particular promise is the rational design of drugs. AlphaFold 3 can efficiently predict how ligands, a structural component of many drugs, will bind to a target protein, thereby guiding researchers in designing drug candidate structures for improved efficacy and reduced side effects. This not only hastens the drug discovery process but also has the potential to ultimately impact the overall cost of development.
“AlphaFold was a watershed moment for AI in biology. It was proof just how powerful AI could be in understanding the complexities of the biological world,” says Max Jaderberg, chief AI officer at Isomorphic Labs. “It’s what inspired the founding of IsoLabs three years ago.” A compelling example of AlphaFold’s effectiveness is demonstrated by Isomorphic Labs in a case study on TIM-3, one of many immune checkpoint proteins that regulate the immune system and prevent excessive responses. In some cancers, these checkpoints are hijacked to suppress the immune system, allowing the cancer to evade detection. TIM-3 has been identified as a potential target for cancer immunotherapy, as blocking its activity could help reactivate the immune response against tumor cells.
A 2021 study experimentally determined the structures of TIM-3 bound to several small-molecule ligands, a laborious process that revealed a previously unknown binding pocket. When given just the sequence of TIM-3 and the chemical structures of the ligands, with no information about the binding pose or pocket, AlphaFold 3 was able to accurately predict the experimentally determined structures. Crucially, the model correctly identified this novel binding pocket and the precise orientations of the ligands within it, even though these structures were not in the training set of AlphaFold 3.
“The technology we already have at our fingertips also opens up new possibilities to tackle diseases that have been previously beyond our reach. Our team of scientists and researchers at IsoLabs has created a whole host of unique AI algorithms. These models work together, allowing us to understand the intricate world of proteins and how they interact with different molecules. This gives us new ways to design molecules, which is incredibly exciting, and we believe it will completely change how we discover and design new drugs,” says Jaderberg.
The search for therapeutic antibodies
AlphaFold 3 also has the ability to predict antibody-protein binding, which can help researchers understand immune response and design new antibodies.
Antibodies are examples of therapeutic proteins that can be used to treat diseases like cancer. But discovering new ones by traditional means is an arduous task. With trillions of possible structures, finding the ones that effectively stick to a target of interest is like searching for a needle in a haystack, experts say.
“Finding new therapeutic antibodies is really sort of like panning for gold,” explains Derek Lowe, a veteran pharmaceutical researcher. “You have to check against incredible numbers of possibilities to find things that bind, and then you try to make them better and better.”
Taking advantage of AlphaFold 3’s ability to accurately model protein interactions, scientists can computationally screen vast numbers of potential candidates and identify the most promising ones for experimental validation. This computational prescreening could drastically reduce the time and effort required to discover new therapeutic proteins, expediting the development of treatments for cancer, contagious diseases, and other conditions for which these drugs have shown promise.
Shedding light on neglected diseases
AlphaFold is already generating optimism among those investigating overlooked illnesses. The Drugs for Neglected Diseases initiative (DNDi) has been using AlphaFold for years as a means to hasten the discovery of new treatments for two deadly parasitic diseases, Chagas disease and leishmaniasis, which affect tens of millions of people worldwide.
“AlphaFold 2 has been an immensely useful tool for us, but to turn protein structures into drugs is a journey, and to make progress in that journey, you need to be able to think about how those proteins interact with other molecules,” says Dr. Charles Mowbray, DNDi’s discovery director. “AlphaFold 3 offers huge possibilities in this vital next step. It can predict not just the structures, but also the interactions, taking us a vital step further in the drug discovery process.”
A tool, not a replacement
With such powerful AI tools emerging, some wonder what this means for the role of scientists. But as Lowe observed, the scientific community is more excited than worried. “We’re very happy about this technology because we have so many problems to deal with,” Lowe says. “The drug discovery and development process is so long, with so many twists and turns, that we welcome anything that could speed elements of it up or give us some insights.” Ultimately, while AI and AlphaFold are undeniably important parts of the drug discovery process, they are just that—parts. Structure prediction is a single piece of the puzzle, albeit a crucial one, that can help accelerate the development of essential medications for those who need them most.
“Leading the AlphaFold team has been the opportunity of a lifetime, but what I’m most excited for will be seeing scientists use AlphaFold to understand a disease and transform patients’ lives,” Jumper says. “I can’t wait to see that day.”
