AlphaFold and the Fifty-Year Protein Problem

A protein's job depends almost entirely on its shape. Hemoglobin carries oxygen because of the specific pocket it folds into. An enzyme speeds up a reaction because its folded shape holds two molecules in exactly the right position to react. Knowing a protein's amino acid sequence — the order of building blocks specified by its gene — tells you almost nothing about that shape on its own. For roughly fifty years, figuring out how a sequence folds into a 3D structure was one of biology's most stubborn open problems. In 2020, an AI system called AlphaFold effectively solved it. This reading covers how AlphaFold's approach actually works, and then — because knowing about a tool is different from being able to use one — whether a student can actually put it to work on a real protein.

Why folding was ever a "fifty-year problem"

In 1961, biochemist Christian Anfinsen showed that a small enzyme, ribonuclease, could be fully unfolded with chemical treatment and then would refold on its own, with no outside help, back into its original working shape once the treatment was removed. That result — work he won a share of the 1972 Nobel Prize in Chemistry for — established what's still called Anfinsen's dogma: a protein's amino acid sequence contains all the information needed to determine its final 3D structure. In principle, then, structure should be computable directly from sequence. In practice, nobody could do it. A modest-sized protein has an astronomical number of geometrically possible ways to fold — far more than could ever be checked one by one, a problem often called Levinthal's paradox, since a real protein finds its correct shape in milliseconds despite that search space, while a computer trying every possibility would not finish before the universe grew old. The only reliable way to actually determine a protein's structure was to measure it directly, using slow, expensive, physical methods — X-ray crystallography, NMR spectroscopy, and later cryo-electron microscopy — each of which could take months or years of work per protein, and each of which requires a stable, purified, often difficult-to-produce sample. By 2020, roughly 170,000 protein structures had been solved this way and deposited in the Protein Data Bank (PDB), the field's shared archive — a huge body of work, but a small fraction of the hundreds of millions of distinct proteins known to exist across life on Earth. Genome sequencing had gotten fast and cheap. Knowing what the proteins those genomes encode actually look like had not kept pace at all.

How AlphaFold's approach actually works

AlphaFold, built by Google DeepMind and led by John Jumper and Demis Hassabis, does not search through possible foldings the way Levinthal's paradox makes sound necessary. It predicts structure directly, using patterns learned from the roughly 170,000 structures already solved and deposited in the PDB. Two ideas make that prediction possible. First, AlphaFold starts not from one sequence but from a multiple sequence alignment (MSA): versions of the same protein family found across many different species, lined up against each other. Positions that mutate together across species — a change at one site consistently paired with a change at another, even though the two sites are far apart in the raw sequence — are a strong clue that those two positions sit physically close together in the folded 3D structure, because evolution tends to preserve a working shape even as it tolerates sequence drift. Second, AlphaFold uses attention, the same core mechanism behind large language models, in a module called the Evoformer, to let the sequence alignment and a running map of residue-to-residue distances repeatedly inform and refine each other across many internal passes, before a final structure module converts that refined internal representation directly into 3D atomic coordinates. The whole system is trained end-to-end against the experimentally solved PDB structures, learning what shapes real protein sequences actually fold into rather than being told the rules of folding directly. Critically, AlphaFold also outputs a per-position confidence score, called pLDDT, alongside every predicted structure — it does not just guess a shape, it also tells you how much to trust each part of that guess.

What AlphaFold actually does today

At the 2020 CASP14 competition — the field's blind, biennial test where research groups predict structures nobody has solved yet and an independent panel scores the results against the real answer once it's determined — AlphaFold's predictions were close enough to experimentally measured structures that CASP organizers described the protein-folding problem as effectively solved for single protein chains. DeepMind and the European Bioinformatics Institute then built the AlphaFold Protein Structure Database, a free, public library that has grown to more than 200 million predicted structures — not a slight expansion of the roughly 170,000 structures sixty years of crystallography had produced, but nearly every protein sequenced across life on Earth. Researchers have since used it to speed up work on neglected tropical diseases, design new enzymes, identify drug targets, and study antibiotic resistance, among many other applications. A newer version, AlphaFold 3 (2024), extended the approach beyond single proteins to predict how proteins interact with DNA, RNA, and small-molecule drugs together. In October 2024, Hassabis and Jumper were awarded half of the Nobel Prize in Chemistry for AlphaFold, with the other half going to David Baker for related work using similar computational methods to design entirely new proteins that don't exist in nature.

Can a student actually use it to study a protein?

Yes, and in two genuinely different ways. The easiest is to look up a protein that's already been predicted: the AlphaFold Protein Structure Database (alphafold.ebi.ac.uk) is free and searchable by protein or gene name, and returns an interactive 3D model you can rotate in a browser, color-coded by that same pLDDT confidence score — dark blue for very high confidence, orange and red for regions the model isn't sure about. Looking up the CFTR protein (the gene involved in cystic fibrosis) or the beta-globin chain of hemoglobin (relevant to sickle cell disease) is a genuinely productive way to connect a structure to a real disease students have likely already heard of. The second, more involved option is predicting a structure that doesn't exist yet: a free tool called ColabFold runs in Google Colab, letting anyone paste in their own amino acid sequence and get a predicted structure back, using a graphics processor Google provides for free, with no coding experience required beyond running a notebook cell.

Using either tool well means reading the confidence score, not just the pretty picture. A high pLDDT score means AlphaFold has seen enough related sequences and solved structures to make a genuinely trustworthy prediction. A low score in a specific region is not always a failure of the model — it's frequently the model correctly reporting that a stretch of protein is intrinsically disordered, meaning it doesn't settle into one fixed shape in real life either. AlphaFold's predictions are also weakest for proteins with few known relatives to build a useful sequence alignment from, and the model predicts a single resting structure rather than the range of shapes many real proteins shift between while doing their job. A student using AlphaFold to study a protein is doing real science — the same database professional researchers use — but reading the confidence score honestly, rather than trusting every colorful ribbon diagram equally, is the actual skill being practiced.

Anfinsen's dogma

The principle, established experimentally in 1961, that a protein's amino acid sequence contains all the information needed to determine its final folded 3D structure.

Levinthal's paradox

The observation that a protein has far too many geometrically possible folded shapes to find its correct one by random search, yet real proteins fold correctly in milliseconds.

Multiple sequence alignment (MSA)

Versions of the same protein found across many species, lined up against each other so that positions which mutate together across species — a clue to physical closeness in the folded structure — can be identified.

pLDDT

AlphaFold's per-position confidence score for a predicted structure; high values indicate a trustworthy prediction, and low values often correctly flag a genuinely disordered or uncertain region rather than a modeling failure.

Protein Data Bank (PDB)

The shared international archive of protein structures determined experimentally by methods such as X-ray crystallography, NMR spectroscopy, and cryo-electron microscopy; the training data AlphaFold learned from.

Check your understanding

Sources: Anfinsen CB, "Principles that Govern the Folding of Protein Chains," Nobel Lecture, 1972; Jumper J, Evans R, Pritzel A, et al., "Highly accurate protein structure prediction with AlphaFold," Nature, 596:583–589, 2021; The Nobel Prize in Chemistry 2024, press release and popular science background, NobelPrize.org; EMBL-EBI and DeepMind, AlphaFold Protein Structure Database (alphafold.ebi.ac.uk); Mirdita M, et al., "ColabFold: making protein folding accessible to all," Nature Methods, 2022. DRAFT — verify current 2018 Virginia Science Standards of Learning biology codes with the current Curriculum Framework before publishing.