“The structure of more than 241 million proteins has been predicted, a catalog that would have taken hundreds of millions of years of experimental work to assemble.”
At some point, the history of human knowledge makes a leap that is impossible to anticipate. At the end of 2020, in a scientific competition almost unknown to the general public, an artificial intelligence algorithm called AlphaFold 2 solved in a matter of hours a puzzle that had defied biology for half a century: the protein folding problem.
This achievement was not a mere academic triumph. It became the foundation of a revolution that democratized access to structural knowledge, accelerated biomedical research on an unimaginable scale and was recognized, quite rightly, with the Nobel Prize in Chemistry in 2024. More than three million researchers around the world use its tools today, from university students to the largest pharmaceutical labs on the planet, rewriting what it means to do science.
The “Problem of the Century” in Biology:
To understand the magnitude of the achievement, we must first dive into the complexity of the problem. Proteins are the molecular machines that carry out practically all the functions of life: from transporting oxygen in the blood to defending the body against pathogens. How they work is dictated not only by the linear sequence of amino acids that make them up, but by the precise three-dimensional shape in which that chain folds in on itself.
For decades, deciphering that 3D structure was a Herculean task. Experimental methods such as X-ray crystallography or cryo-electron microscopy could consume years of work and hundreds of thousands of dollars for a single protein. The computational challenge was even more daunting. Given the astronomical number of possible ways an amino acid chain can fold, calculating the correct structure by chance, as Levinthal's paradox pointed out, would take longer than the age of the universe. Before AlphaFold, the best algorithms matched experimental accuracy in only 40% of the hardest cases.
From AlphaFold 1 to AlphaFold 3: how the model evolved
The Artificial Intelligence Assault: From AlphaFold 1 to AlphaFold 3:
AlphaFold's journey began, curiously, on a Go board. DeepMind's success with AlphaGo in 2016 showed that its AI techniques could master systems of overwhelming complexity, which inspired a small team to take on protein folding.
AlphaFold 1 (2018): The first impact. In its debut at the CASP13 competition, AlphaFold 1 surprised the scientific world. Its approach, which used a neural network to predict distances between amino acids and then assemble the structure, outperformed all other methods. However, it was a complex multistep system and its accuracy, though revolutionary, still did not reach the experimental level.
AlphaFold 2 (2020): The definitive solution. Everything changed in 2020. AlphaFold 2 was not a simple improvement; it was a radical conceptual redesign. It abandoned the fragmented process and adopted an “end-to-end” architecture based on the Transformer mechanism. Its core, the “Evoformer” module, simultaneously processed two streams of information: the alignment of evolutionarily related sequences (which reveals which amino acids mutate together) and the pairwise geometric relationships. This internal dialogue allowed the system to infer physical and evolutionary constraints in an integrated way, predicting structures with atomic accuracy comparable to experimental results in two-thirds of cases. The CASP organizers declared that the 50-year-old problem had been solved.
AlphaFold 3 (2024): Biology as a whole. If AlphaFold 2 deciphered the individual words (the proteins), AlphaFold 3 set out to understand the cell's complete grammar. This model dramatically expanded its reach to predict not just isolated proteins but entire molecular complexes: protein-protein interactions, protein binding to DNA and RNA, and docking with ligands (the small molecules that are often drugs). It swapped the structure module for a generative diffusion model, iteratively refining a cloud of atoms until it reached the stable structure, which makes it possible to visualize the molecular dance of life in unprecedented detail.
Evoformer: how AlphaFold 2 combines evolution and geometry
The Revolutionary Architecture: Evoformer and the Dialogue of Information:
The qualitative leap of AlphaFold 2 lies in its architecture. Unlike its predecessor, which predicted a distance map and then assembled the structure, AlphaFold 2 continuously maintains and refines two parallel representations:
The Multiple Sequence Alignment (MSA) representation: it encodes evolutionary information. If two amino acids far apart in the sequence tend to mutate in a coordinated way over the course of evolution, it is a strong signal that they are physically close in the 3D structure.
The Pair representation: it encodes direct geometric and spatial information between each pair of residues.
The Evoformer acts as the operations center. Through repeated blocks with specialized attention mechanisms (such as Triangular Self-Attention), it allows a constant exchange of information between these two streams. Evolutionary information helps refine the geometric relationships, and in turn the improved geometric relationships help reinterpret the evolutionary data. This iterative process culminates in the Structure Module, which finally translates this abstract, context-rich representation into precise three-dimensional atomic coordinates.
The open database and its impact on science and medicine
Democratization and Global Impact: The Database That Changed Everything.
The technical advance would have remained a laboratory curiosity if not for one momentous decision: releasing it to the world. In July 2021, DeepMind, in partnership with the European Bioinformatics Institute (EMBL-EBI), launched the AlphaFold Protein Structure Database. A year later, this database held the predicted structures for more than 200 million proteins, covering nearly all the sequences cataloged by science.
Access is open and free. This decision set off a wave of discoveries. More than 3 million researchers from 190 countries have used the database. University students, such as the Turkish brothers Alper and Taner Karagöl, taught themselves structural biology during the pandemic and published scientific papers. In countries with limited resources, scientists who could never have afforded structural experiments now have “atomic blueprints” to guide their research.
Tangible Impacts on Science and Medicine.
Concrete examples of AlphaFold's impact keep multiplying:
Heart health: it revealed for the first time the complete structure of apolipoprotein B100, a key protein in “bad cholesterol” (LDL) and atherosclerosis, opening new paths for the design of preventive drugs.
Conservation and food: It was used to understand a crucial immune protein in honeybees, guiding breeding programs for more resilient pollinators. It is also accelerating research to develop crops more resilient to climate.
Drug discovery: Companies such as Isomorphic Labs (founded after AlphaFold's success) are using these models for rational drug design, drastically shortening the timelines for identifying promising therapeutic candidates.
Accelerated basic research: An independent analysis showed that scientists who use AlphaFold submit 40% more novel protein structures to public databases, exploring uncharted territories of biology.
Current limits and the next frontier of digital biology
The Road Ahead: Limits and the Next Frontier.
Despite its transformative power, AlphaFold is not omnipotent. The scientific community is very aware of its limitations, which in themselves map out the next generation of tools.
A fundamental criticism is that it predicts a static structure, a “snapshot” of the protein in its most stable state. Yet proteins in the cell are dynamic: they move, change shape and adopt multiple conformations to fulfill their function. AlphaFold cannot easily model this molecular ballet, the effects of point mutations, or the interaction with very specific molecules such as some antibodies.
It also does not handle well certain chemical modifications that occur after the protein is synthesized (post-translational modifications), and it struggles with intrinsically disordered proteins, which lack a fixed shape. AlphaFold 3 addresses some of these points, such as interactions with other biomolecules, but predicting dynamics and multiple states remains an open field.
The future, inspired by AlphaFold, is already taking shape with models such as AlphaMissense (to predict the effect of genetic mutations) and AlphaProteo (to design new proteins from scratch). The vision is a “digital biology” in which researchers can generate and test complex hypotheses in silico before entering the lab, shifting the pace of discovery toward the digital.
The timeline: five years that transformed biology.
| Date | Milestone | Significance |
| December 2018 | AlphaFold 1 wins the CASP13 competition. | Shows for the first time that deep learning can outperform traditional methods. |
| November 2020 | AlphaFold 2 wins CASP14 with atomic accuracy. | The 50-year-old “protein folding problem” is declared solved. |
| July 2021 | Launch of the AlphaFold Protein Structure Database with EMBL-EBI. | Massive democratization: 350,000 structures are released, including the human proteome. |
| July 2022 | The database expands to more than 200 million structures. | Coverage of nearly every protein known to science. |
| May 2024 | Launch of AlphaFold 3 and the AlphaFold Server. | Ability to predict interactions among all the molecules of life. |
| October 2024 | The AlphaFold team receives the Nobel Prize in Chemistry. | Formal recognition of its historic contribution to scientific knowledge. |
| November 2025 | The database reaches 241 million structures (v6) and passes 3 million users. | Consolidation as a standard global research tool. |
The story of AlphaFold is more than that of a successful algorithm. It is the story of how a group of scientists, coming from artificial intelligence and not from traditional biology, reframed a fundamental problem and, by solving it, put into humanity's hands a new lens for observing the basic components of life. Its true legacy lies not only in the predicted structures, but in the millions of hypotheses they have inspired, in the diseases that are being rediscovered and in the global acceleration of scientific curiosity. As the film The Thinking Game documented, capturing the human drama behind the code, this was a triumph of disciplined imagination, a reminder that the boldest leaps often come from looking at old problems with entirely new eyes.
