# How does protein folding work?

> Protein folding is how a chain of amino acids takes on the 3D shape it needs to work. The sequence decides the shape and the burial of water-avoiding side chains pays for it, yet one changed residue can flip a protein into a different fold. Learn how folding works, how it can be so fast, what chaperones do and how misfolding causes disease.

## What is protein folding?

Protein folding is the process by which a newly made chain of amino acids takes on the one three-dimensional shape that lets it work. The chain comes off the ribosome as a floppy string; in as little as microseconds it packs into a compact structure with a pocket, a binding surface or a pore. The amino acid sequence decides which shape that is, and water does most of the pushing.

That rule is true and still surprising in practice. Patrick Alexander, Philip Bryan and colleagues made two 56-residue proteins, GA98 and GB98, that differ at a single position, leucine 45 in one and tyrosine 45 in the other. GA98 folds into a bundle of three helices. GB98 folds into four β-strands and one helix, a completely different fold with a different function. At the same time, many unrelated sequences share one fold. The sequence determines the fold, but not residue by residue.

What makes sense of both cases is a balance: a fold is the arrangement whose favorable contacts outweigh the cost of ordering the chain, usually by a small margin. That margin also explains how folding can be fast, why cells need chaperones, how misfolding causes disease and why a structure predicted by [ESMFold](/app/esmfold) answers only part of the question.

## How the amino acid sequence determines protein folding

The evidence that the sequence carries the information came from Christian Anfinsen's work on ribonuclease A, a 124-residue enzyme held together by four [disulfide bonds](/guides/disulfide-bonds). Anfinsen's group [denatured it](/guides/protein-denaturation) completely in urea and broke the disulfides, then let the chain reoxidize without the denaturant. It refolded and recovered its enzyme activity with no help from anything but the buffer.

From this Anfinsen proposed his thermodynamic hypothesis: under normal conditions, the native structure is the arrangement of lowest free energy, so it is set by the sequence.

The disulfides show how much information that is. Eight cysteines can pair into four bonds in $7 \times 5 \times 3 \times 1 = 105$ different ways, so we calculate that a chain pairing them at random would land on the native set less than 1% of the time. The folded chain finds it because the fold brings the right cysteines together first.

How does a string of letters encode a shape? Mostly through the pattern of water-avoiding and water-loving residues along the chain, its [hydropathy](/guides/kyte-doolittle-hydropathy). Michael Hecht's group wrote genes that fixed only that pattern for a four-helix bundle, placing nonpolar residues where the helices would touch and polar residues on the outside, while letting the exact amino acids vary widely. Most of the resulting proteins, each with a different sequence, folded into compact helical structures.

The binary pattern explains the many-sequences-one-fold half of the puzzle. GA98 and GB98 explain the other half. In GA98, leucine 45 packs against the helical core; tyrosine cannot pack there as well, and in the β fold the same tyrosine finds an aromatic partner, phenylalanine 52. One substitution weakens one fold and strengthens the other, and the chain switches. A sequence encodes its fold through a balance of contacts, and a balance can be tipped.

## Why proteins fold

A protein folds because the folded state has lower free energy than the unfolded one, and the main reason is the hydrophobic effect. Nonpolar side chains such as leucine, isoleucine and phenylalanine force surrounding water into ordered cages; burying them in the core frees that water. Hydrogen bonds in [helices and sheets](/use-cases/protein-secondary-structure-prediction), tight van der Waals packing and, in secreted proteins, disulfides add to the bill.

Nick Pace's group measured how the bill splits. Across 22 proteins of 36 to 534 residues, hydrophobic interactions contributed 60 ± 4% of the stabilizing free energy and hydrogen bonds 40 ± 4%. Burying a single CH<sub>2</sub> group added 1.1 ± 0.5 kcal/mol on average, from 148 mutants in 13 proteins.

The catch is that these large favorable terms are nearly cancelled by the cost of ordering the chain, which Pace's group put at about 2.4 kcal/mol per residue. In the 36-residue villin headpiece subdomain, hydrophobic contacts alone contribute about 40 kcal/mol, yet the folded protein is only 2.7 kcal/mol more stable than the unfolded chain at 25 °C. That thin margin is why a point mutation, a few degrees of heat or a change in pH can unfold a protein, and why a stability prediction from [ThermoMPNN](/app/thermompnn) is worth running before making a variant.

## How fast proteins fold

Small single-domain proteins fold in microseconds to seconds; large or complex ones take longer. That speed is the point of Levinthal's paradox. Suppose each residue of a 100-residue chain can take one of three [backbone shapes](/guides/ramachandran-plot). We calculate how long a random search through every conformation would take at a billion conformations per second:

$$
\begin{aligned}
3^{100} &\approx 5 \times 10^{47}\ \text{conformations} \\
t &= \frac{5 \times 10^{47}}{10^{9}\ \text{s}^{-1}} \\
&\approx 5 \times 10^{38}\ \text{s} \\
&\approx 10^{31}\ \text{years}
\end{aligned}
$$

The universe is 13.8 billion years old. A random search cannot be how proteins fold.

Robert Zwanzig, Attila Szabo and Biman Bagchi showed that the paradox disappears with a small bias. In their model, each residue is either in its native shape or a wrong one, and a wrong shape costs a small energy. A bias of a few kT per residue, where kT is the thermal energy, cuts the search time from cosmological to biological. We recalculated their model for a 100-residue chain to see how steep that drop is (Table 1).

  <caption><strong>Table 1. Mean time for a 100-residue chain to reach its native state in the Zwanzig, Szabo and Bagchi model.</strong> ProteinIQ calculation in which each residue has one native and two non-native shapes, flips toward native at 10<sup>9</sup> per second, and pays the listed energy for each non-native residue. Times are exact mean first-passage times from a fully non-native chain, computed as a birth-death process on the number of native residues. 1 kT is 0.62 kcal/mol at 37 °C. At zero bias the time is shorter than the naive 10<sup>31</sup> years because all 100 residues change shape in parallel. The model has no cooperativity or physical speed limit, so it shows the search problem, not any real protein. Model: Zwanzig, Szabo and Bagchi (1992).</caption>
  <thead>
    <tr>
      <th scope="col">Bias per residue (kT)</th>
      <th scope="col">Bias per residue (kcal/mol, 37 °C)</th>
      <th scope="col">Mean folding time</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>0 (random search)</td>
      <td>0</td>
      <td>8 × 10<sup>28</sup> years</td>
    </tr>
    <tr>
      <td>0.5</td>
      <td>0.31</td>
      <td>8 × 10<sup>15</sup> years</td>
    </tr>
    <tr>
      <td>1</td>
      <td>0.62</td>
      <td>390,000 years</td>
    </tr>
    <tr>
      <td>1.5</td>
      <td>0.92</td>
      <td>2.8 days</td>
    </tr>
    <tr>
      <td>2</td>
      <td>1.23</td>
      <td>1.0 second</td>
    </tr>
    <tr>
      <td>2.5</td>
      <td>1.54</td>
      <td>260 microseconds</td>
    </tr>
    <tr>
      <td>3</td>
      <td>1.85</td>
      <td>1.5 microseconds</td>
    </tr>
  </tbody>

A bias of 2 kT, about 1.2 kcal/mol per residue, brings the time from 10<sup>28</sup> years to about a second. That is close to what burying one CH<sub>2</sub> group is worth, 1.1 kcal/mol in Pace's data. Each extra half kT buys between 2 and 13 orders of magnitude, because the chain keeps the local moves that work instead of starting over. Energy landscape theory describes the same idea as a funnel: a rugged surface sloping toward the native state, so that many routes lead down it.

Real proteins stop well short of the table's bottom rows. Jan Kubelka, James Hofrichter and William Eaton estimated a speed limit of about $N/100$ microseconds for a single domain of $N$ residues, or about 1 microsecond for 100 residues.

Length is not what sets most folding rates, though. In a set of simple single-domain proteins, Kevin Plaxco, Kim Simons and David Baker found no significant link between length and folding rate; proteins whose contacts were mostly between residues close in sequence folded faster than those built on long-range contacts. Kubelka's group likewise expected α-helical proteins to fold faster than β or mixed ones.

Larger proteins do not fold in one leap. Walter Englander and Leland Mayne describe them as built from foldons, cooperative units of about 20 residues that form and unfold as blocks, with each one guiding the next. The small size of each unit is itself an answer to Levinthal: no single search has to cover the whole chain.

## How proteins fold inside a cell

A test tube is a forgiving place to fold. The cytoplasm of [*E. coli*](/guides/e-coli-statistics) holds 0.3 to 0.4 g/mL of protein and other macromolecules, Steven Zimmerman and Sam Trach estimated. A half-folded chain there has exposed hydrophobic patches, and the nearest sticky surface may belong to another half-folded chain. The cell answers with two things: folding that starts during synthesis, and chaperones.

Folding begins before the chain is finished. Chaperones meet the nascent chain as it emerges from the ribosome and guide it along a productive route. Their job is not to choose the shape. The sequence still does that. Chaperones such as trigger factor and the Hsp70 protein DnaK hold unfinished chains so they do not aggregate, and chaperonins such as GroEL close a client inside a cage where it can fold alone.

How much of a proteome needs that help? Michael Kerner and colleagues in Ulrich Hartl's group mapped GroEL clients across *E. coli*. About 250 proteins interact with GroEL, but most can fold with the upstream chaperones instead. Only about 85 depend on it, including 13 essential proteins, and those occupy more than 75% of GroEL's capacity. They get stuck in kinetically trapped intermediates and reach the native state only inside the cage. For those proteins the sequence encodes a stable fold the chain cannot easily reach on its own. Chaperones fix the route, not the destination.

## What happens when protein folding goes wrong

A misfolded protein can lose its function, [get degraded](/guides/how-long-do-proteins-last-in-the-body) or join other copies in [protein aggregation](/guides/protein-aggregation), where exposed hydrophobic surfaces stick together. Fabrizio Chiti and Christopher Dobson describe an inherent tendency of proteins to convert into amyloid, fibrillar aggregates linked to Alzheimer's and Parkinson's disease, type 2 diabetes and the systemic amyloidoses. Chaperones and degradation machinery resist it, but that network loses capacity with age, one reason these diseases rise late in life.

Prions push Anfinsen's rule furthest. The normal prion protein, PrP<sup>C</sup>, and the disease form, PrP<sup>Sc</sup>, have the same sequence; PrP<sup>Sc</sup> carries much more β-sheet and acts as a template that refolds normal copies into more of itself. Stanley Prusiner's work placed the strain differences of prion diseases in this tertiary structure, not in a genome. One sequence, two stable shapes, and the wrong one spreads.

## What CFTR ΔF508 shows about folding

Cystic fibrosis shows how the parts of folding can give different answers about the same protein. CFTR is a chloride channel in the plasma membrane of epithelial cells, and the most common disease mutation, ΔF508, deletes a single phenylalanine.

Judged by its location, the mutant has failed. Seng Cheng and colleagues found that ΔF508 CFTR never matured: it stayed incompletely glycosylated in the endoplasmic reticulum and was degraded there. They argued that mutations with this phenotype account for at least 70% of cystic fibrosis chromosomes, so most of the disease comes from a protein that is missing from the cell surface.

Judged by its sequence, the mutant still encodes a working channel. Glenn Denning and colleagues noticed that chloride channel activity appeared when ΔF508 was expressed in frog eggs and insect cells, which are kept cooler than mammalian cells. When they lowered the incubation temperature of mammalian cells, processing moved back toward that of the normal protein, and cAMP-regulated chloride channels appeared in the membrane.

Put together, the deletion does not erase the fold; it makes the fold hard to reach at 37 °C, and the cell's quality control destroys the protein before it gets there. The defect is in the folding process, not in the information. That distinction decides what a treatment has to do: help the protein fold and escape the endoplasmic reticulum, rather than replace a broken function.

## Where the one-sequence, one-fold rule stops

Anfinsen's rule describes compact globular domains well. Many proteins do not fit it. Intrinsically disordered regions stay flexible under normal conditions and work as shifting ensembles, often folding only when they meet a partner. David Jones's group predicted long disordered segments, more than 30 residues, in 2.0% of archaeal, 4.2% of bacterial and 33.0% of eukaryotic proteins, concentrated in nuclear proteins that regulate transcription and signaling.

Fold switchers like GA98 and GB98 and the prion protein show that a sequence can hold more than one stable answer. A single structure file shows one state, or an average. [BioEmu2](/app/bioemu-2) samples the range of conformations a sequence visits at equilibrium, which is closer to how a flexible protein behaves than one static model.

## AlphaFold solved prediction, not folding

AlphaFold largely solved one part of the problem: predicting the folded structure of many single chains from their sequence. It did not explain the folding process. People often say "protein folding" when they mean [protein structure prediction](/use-cases/protein-structure-prediction), but the two questions differ (Table 2). Structure prediction asks where the chain ends up; folding asks how it gets there and whether it will get there in a given cell.

Carlos Outeiral, Daniel Nissley and Charlotte Deane tested whether the intermediate structures that AlphaFold2, RoseTTAFold and six other predictors pass through resemble real folding pathways. They did not: the predicted trajectories were uncorrelated with experimental intermediates and folding rates, and predicted pathway features worse than a trivial classifier using chain length.

A confident prediction says what the folded protein probably looks like. Read [pLDDT](/guides/plddt) as confidence in the predicted coordinates, not as a measure of how stable the protein is or whether it will fold in *E. coli*.

  <caption><strong>Table 2. Questions about protein folding and where to start on each.</strong> What decides the answer and the first check worth running.</caption>
  <thead>
    <tr>
      <th scope="col">Question</th>
      <th scope="col">What decides it</th>
      <th scope="col">First check</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>What shape does this sequence adopt?</td>
      <td>The sequence's lowest free energy fold</td>
      <td>Predict with [ESMFold](/app/esmfold) or [AlphaFold2](/app/alphafold-2) and read the confidence per residue</td>
    </tr>
    <tr>
      <td>Is part of it disordered?</td>
      <td>Whether a region has a hydrophobic core to bury</td>
      <td>Low-confidence stretches in a prediction, then an ensemble from [BioEmu2](/app/bioemu-2)</td>
    </tr>
    <tr>
      <td>Will a mutation destabilize it?</td>
      <td>The few kcal/mol margin between folded and unfolded</td>
      <td>Predicted stability change from [ThermoMPNN](/app/thermompnn)</td>
    </tr>
    <tr>
      <td>Will it aggregate at high concentration?</td>
      <td>Exposed hydrophobic patches on the folded or partly folded protein</td>
      <td>Surface aggregation map from [Aggrescan3D](/app/aggrescan3d)</td>
    </tr>
    <tr>
      <td>Will it fold in my expression host?</td>
      <td>Kinetics, chaperones, disulfides and compartment</td>
      <td>An expression test; no prediction answers this reliably</td>
    </tr>
  </tbody>

The last row is the one predictions most often get credit for wrongly. A well-predicted structure and a protein that folds well in a cell are separate facts. CFTR ΔF508 still encodes a working channel, yet at body temperature most of it never reaches the membrane.

## Why a small margin explains folding

The parts of protein folding fit together through one number: the few kcal/mol by which the folded state wins. The sequence decides which contacts are possible and the hydrophobic effect pays for them, but after the cost of ordering the chain, little is left over. A bias that small per residue is enough to turn Levinthal's impossible search into microseconds, and small enough that one substitution tips GA98 into the fold of GB98.

Inside a cell, the same thin margin leaves time for a half-folded chain to meet a sticky neighbor, which is why chaperones guard the route and why quality control can destroy a protein that would have folded at a lower temperature. Predictors now read the destination from the sequence well. The route, and whether a given cell lets the chain finish it, still has to be measured.
