Protein Folding — How a Sequence Becomes a Shape

A protein is a linear chain of amino acids that folds into a specific three-dimensional shape. That shape determines what the protein does — catalyzing reactions, transmitting signals, structuring cells, or defending against pathogens. The same sequence always produces the same structure, under the same conditions. The mapping from one-dimensional sequence to three-dimensional form is one of the most reliable and least intuitive processes in biology.

Understanding how it works required resolving a genuine paradox, revising assumptions about what happens inside living cells, and building computational models accurate enough to predict structures that have never been observed experimentally.

The sequence determines the structure

The foundation was laid by biochemist Christian B. Anfinsen in the 1950s and early 1960s, working with ribonuclease A, a digestive enzyme from bovine pancreas containing 124 amino acids held together by four disulfide bonds.

Anfinsen showed that denaturing the protein — unfolding it with urea and breaking its disulfide bonds with a reducing agent — produced an inactive random coil. When he removed the denaturants and allowed oxygen to reform the disulfide bonds, the protein spontaneously refolded into its native, fully active structure. Not one of the many possible arrangements of four disulfide bonds among ten cysteine residues, but the correct one.

The conclusion, published in a series of papers culminating in Anfinsen’s 1961 review with E. Schmitz, was that all the information needed to reach the native structure is encoded in the amino acid sequence itself. No external template or instruction is required. The protein folds because its native state is the thermodynamically most stable configuration — the one with the lowest free energy.

This became known as Anfinsen’s dogma: sequence determines structure, and structure determines function. It earned Anfinsen the Nobel Prize in Chemistry in 1972.

The dogma holds for many proteins under physiological conditions. It also has important exceptions that emerged as the field matured.

The paradox of speed

If the native state is simply the lowest-energy configuration, a protein could find it by sampling all possible conformations and settling into the best one. Cyrus Levinthal pointed out in 1969 that this is not physically possible.

A modest-sized protein with 100 residues has roughly three to four accessible backbone conformations per residue. Even assuming an optimistic three conformations per residue, the total conformational space is 3^100 — approximately 10^47 possibilities. Sampling each conformation for just 10^-13 seconds (a realistic lower bound for molecular motion) would take longer than the age of the universe.

Yet proteins fold in milliseconds to seconds. This gap between the time required for random search and the observed folding time is Levinthal’s paradox, introduced in his paper “How to Fold Graciously”.

The paradox is not a logical contradiction but a constraint: folding cannot be a random search. Something must guide the process.

The energy landscape and the folding funnel

The resolution came from reframing the problem. Instead of a single pathway through conformational space, proteins fold through an energy landscape — a multidimensional surface where each point represents a conformation and its associated free energy.

Ken A. Dill introduced the folding funnel concept in 1987. Rather than a flat landscape with one deep minimum (the native state), the landscape is shaped like a funnel: broad and relatively high in free energy at the top (unfolded states), narrowing and descending toward the native state at the bottom. The protein does not explore all conformations randomly. It follows the energetic gradient downhill, with the landscape itself biasing the search toward productive folds.

Jose Onuchic, Peter Wolynes, Zaida Luthey-Schulten, and colleagues formalized this in the mid-1990s with the principle of minimal frustration. Natural selection has shaped protein sequences so that interactions stabilizing the native structure are energetically favorable, while non-native interactions are minimized. A “frustrated” landscape would have many local minima — conformations that look stable but are not the native state — trapping the protein in misfolded intermediates. Evolution has reduced this frustration, smoothing the funnel so that folding is both fast and reliable.

This does not mean every protein folds perfectly on its own every time. It means the sequence has been selected to make folding the default outcome under physiological conditions. The energy landscape is biased, not flat.

What happens inside cells

Anfinsen’s experiments were done in test tubes, with purified proteins at low concentration. Inside a living cell, the environment is radically different. The cytoplasm is crowded — protein concentrations reach 300 to 400 grams per liter, creating a densely packed environment where molecules constantly collide. Misfolded proteins can aggregate, forming insoluble clumps that disrupt cellular function.

Cells solve this problem with molecular chaperones — proteins that assist other proteins in folding without being part of their final structure. The term was coined by Ron Laskey in 1978 (describing nucleoplasmin’s role in histone folding) and expanded by R. John Ellis in 1987 to cover the broader class of folding assistants.

Chaperones fall into functional families:

• Hsp70 binds exposed hydrophobic regions on unfolded or partially folded proteins, preventing premature aggregation. It cycles between ATP-bound (fast exchange) and ADP-bound (tight grip) states, releasing the client protein when conditions favor productive folding.

• Hsp60 (GroEL/GroES in bacteria) forms a double-ring structure with a central cavity. A misfolded protein enters the cavity, the chaperonin cap (GroES) closes, and ATP-driven conformational changes give the protein space and time to fold in isolation before release.

• Hsp90 specializes in activating signaling proteins — kinases, steroid receptors, and transcription factors — often working with co-chaperones in ATP-dependent cycles.

Estimates suggest chaperones assist more than half of all mammalian proteins at some point in their lifecycle. They do not violate Anfinsen’s dogma — the sequence still encodes the structure. Chaperones prevent off-pathway aggregation, giving the protein a better chance to find its native state on its own.

When folding fails

Misfolding is not a rare glitch. It is a central mechanism in several major disease classes.

Prion diseases provide the clearest example. Stanley Prusiner discovered prions in the 1980s — infectious agents composed entirely of misfolded protein, with no nucleic acid component. The normal cellular prion protein (PrP^C^) adopts an alpha-helical structure. The disease form (PrP^Sc^) is rich in beta-sheets and aggregates into fibrils. Crucially, PrP^Sc^ can template the conversion of normal PrP^C^ into the misfolded form, creating a self-propagating cascade. This mechanism underlies scrapie in sheep, Creutzfeldt-Jakob disease in humans, and variant Creutzfeldt-Jakob disease linked to bovine spongiform encephalopathy (mad cow disease). Prusiner received the Nobel Prize in Physiology or Medicine in 1997.

Amyloid diseases share a similar structural logic. Alzheimer’s disease involves accumulation of amyloid-beta peptides into extracellular plaques and hyperphosphorylated tau into intracellular tangles. Parkinson’s disease features alpha-synuclein aggregates known as Lewy bodies. In each case, a normally soluble protein adopts a beta-sheet-rich conformation that polymerizes into fibrils, disrupting cellular function through toxic gain-of-function or loss-of-function mechanisms.

These diseases reveal a vulnerability built into the folding problem: the same physical forces that drive proteins toward their native structure can, under the wrong conditions, drive them toward stable misfolded states instead.

Predicting structure from sequence

For decades, determining a protein’s structure experimentally required X-ray crystallography, NMR spectroscopy, or cryo-electron microscopy — techniques that are time-consuming, expensive, and not universally applicable. Computational prediction offered an alternative, benchmarked annually through the Critical Assessment of Structure Prediction (CASP) competition, which began in 1994.

Progress was steady but incremental. Methods that aligned the target sequence to proteins of known structure (homology modeling) worked well when a close homolog existed. They struggled with novel folds or distant relationships.

That changed at CASP14 in November 2020. DeepMind’s AlphaFold 2 achieved a median GDT (Global Distance Test) score of 92.4 across free-modeling targets — structures with no close homolog of known structure. For context, a GDT of 60 was considered acceptable experimental accuracy. AlphaFold 2 used an attention-based neural network that learned to predict inter-residue distances and orientations directly from the multiple sequence alignment and the protein sequence itself, refining its predictions through iterative self-attention.

The AlphaFold Protein Structure Database, launched in July 2021, provides predicted structures for nearly every protein in the UniProt database — over 200 million models spanning all domains of life. It does not replace experimental structure determination, but it provides a starting hypothesis for proteins that have never been crystallized or imaged.

AlphaFold has known limitations. It performs poorly on intrinsically disordered regions — segments that do not adopt a fixed structure even in the native state. It struggles with synthetic proteins that have no evolutionary homologs to draw from. And it predicts a single static structure, while many proteins are functional precisely because they shift between conformations.

The 2024 Nobel Prize in Chemistry was awarded to David Baker (for computational protein design using his Rosetta software), Demis Hassabis, and John Jumper (for AlphaFold), recognizing the convergence of computation and structural biology.

What the folding problem reveals

The protein folding problem — how a linear sequence reliably produces a functional three-dimensional structure — sits at the intersection of thermodynamics, evolution, and information theory.

Anfinsen’s dogma established that the sequence encodes the answer. Levinthal’s paradox showed that the answer cannot be found by brute force. The energy landscape theory explained how evolution shaped sequences so that folding is the default outcome, not a lucky accident. Molecular chaperones revealed that cells invest heavily in keeping the process on track. Misfolding diseases show what happens when the landscape’s biases fail or are overwhelmed. And computational prediction demonstrates that the mapping from sequence to structure is learnable — a pattern that can be extracted from evolutionary data and physical constraints.

The folding problem is not solved in every case. Multidomain proteins, membrane proteins, intrinsically disordered regions, and conformational switching remain active areas of research. But the core mechanism — a biased energy landscape guiding a chain toward its lowest-free-energy configuration — is one of the more complete explanations in molecular biology.