AlphaFold

AlphaFold is a deep learning system developed by DeepMind (Google) that predicts the three-dimensional structure of proteins from their amino acid sequences. Its 2020 performance at the Critical Assessment of Protein Structure Prediction (CASP14) competition — where it achieved accuracy comparable to experimental methods on most targets — was widely described as solving the protein folding problem, a grand challenge of structural biology that had remained open for fifty years. The description was simultaneously accurate and misleading in ways that illuminate how scientific revolutions are narrated.

The Problem AlphaFold Addressed

Proteins are chains of amino acids that fold into specific three-dimensional structures; these structures determine their biological functions. Predicting structure from sequence — the protein folding problem — matters because sequence is easily determined by genetic sequencing while structure determination requires laborious experimental techniques: X-ray crystallography, cryo-electron microscopy, nuclear magnetic resonance. The gap between known sequences and known structures had grown to over one hundred million to one by the time AlphaFold was deployed.

The problem had resisted solution for five decades despite sustained effort. The Levinthal paradox formalized the theoretical obstruction: a protein exploring all possible conformations randomly would require longer than the age of the universe to fold, yet proteins fold in milliseconds. This meant evolution had found an efficient pathway — but the pathway was not, for most of the twentieth century, computationally accessible. Thousands of researchers in hundreds of laboratories had made incremental progress using physics-based simulations, comparative modeling, and fragment assembly. AlphaFold bypassed this accumulated apparatus almost entirely.

How AlphaFold Works

AlphaFold does not simulate physics. It learns statistical patterns from the Protein Data Bank — a repository of experimentally determined protein structures — using a neural architecture (the "Evoformer") that processes multiple sequence alignments and pairwise distance geometries. The system predicts atomic coordinates directly, without running a physical simulation of the folding process.

This is the source of what the historical record will eventually have to reckon with: AlphaFold solves the prediction problem while leaving the mechanistic problem entirely open. It can tell you what structure a protein will adopt; it cannot tell you why it adopts that structure, what the folding pathway is, or what physical principles determine the relationship between sequence and fold. The fifty-year problem of predicting structure is solved. The deeper problem — understanding protein folding as a physical process — is as open as it was before.

Cultural Reception and the Mythology of Revolution

The announcement of AlphaFold's CASP14 performance was greeted with language that reveals more about the cultural moment than about the science. Phrases like "one of the most significant achievements in the history of science" (attributed to John Moult, CASP co-founder) were common in the scientific and popular press. The Nobel Prize in Chemistry 2024, awarded in part to Demis Hassabis and John Jumper for AlphaFold, cemented the narrative.

The historical parallel that clarifies the situation is the sequencing of the human genome, declared complete in 2000 with similarly apocalyptic fanfare. The genome sequence was essential; it was not a theory of gene regulation, development, or disease. Two decades later, we know that knowing the sequence is the beginning of the biological problem, not its solution. AlphaFold occupies an analogous position: it provides data at a scale and speed that makes new research possible, while the interpretive framework for understanding what the data means remains underdeveloped.

This is not a critique of AlphaFold. It is an observation about how cultures narrate scientific progress. The pattern — a spectacular tool is announced, the underlying hard problem is declared solved, a decade of work reveals that the hard problem was not solved but only made more precisely stateable — recurs with sufficient regularity that it should be recognized as a genre of scientific narrative, not an accurate description of scientific resolution.

The Questions AlphaFold Opens

The most productive consequence of AlphaFold is not the structures it has predicted but the questions its failure to address has clarified. The protein folding problem, properly stated, was always multiple problems: prediction, mechanism, and design. AlphaFold addresses prediction; it provides no lever on mechanism; it has enabled new approaches to design (by providing targets for inverse folding) but does not itself perform design.

The mechanistic problem — why does this sequence fold into this structure? — is now more sharply stated because we have the structures. Understanding the rules of protein folding, as opposed to the statistical regularities that AlphaFold exploits, remains an open problem in biophysics. Whether that problem is tractable through computational means at all, or requires new physical theory, is unknown.

What AlphaFold has demonstrated, beyond its direct scientific contributions, is that biological prediction problems can be solved at scale by systems that have no understanding of biology. This is a cultural fact about the relationship between machine learning and science — a relationship whose implications have not yet been assimilated. Every field that touches structure prediction is now asking whether its own grand challenge is an AlphaFold problem waiting to happen: tractable through pattern recognition without mechanistic understanding, solvable without being explained.

The honest answer is that we do not yet know. And the cultural rush to declare AlphaFold a solved science — rather than a powerful instrument in the service of science not yet done — tells us more about our impatience with problems that have no clean ending than it tells us about the protein folding problem itself.