Protein Folding

A search that should take longer than the universe#

Take a modest protein — a hundred amino acids strung together in a line. Each link in that chain can rotate about two backbone bonds, and if you allow just three stable rotational settings per bond, the number of distinct shapes the chain can adopt is roughly

$$3^{2 \times 99} \approx 3^{198} \approx 10^{94}$$

That is a wildly generous count, so shrink it. Suppose the real number is only three settings per residue, not per bond: $3^{99} \approx 5 \times 10^{47}$ conformations. Now let the chain try them one after another as fast as physics allows — a single bond rotation takes about $10^{-13}$ seconds. Sampling all of them takes

$$5 \times 10^{47} \times 10^{-13}\ \text{s} \approx 5 \times 10^{34}\ \text{s} \approx 10^{27}\ \text{years}$$

The universe is about $1.4 \times 10^{10}$ years old. A random search would need roughly a hundred million billion universe-lifetimes to fold one small protein once.

Real proteins fold in microseconds to seconds. Cyril Levinthal pointed this out in 1969, and the contradiction is so violent that it functions as a proof: whatever proteins are doing, it is not sampling shapes at random and stopping when the right one turns up. Something must be steering the chain.

The resolution is that the search is not flat. Every partial step toward the native structure is, on average, rewarded with lower free energy, so the chain is not wandering a maze — it is rolling down a funnel. Levinthal's paradox is not a mystery about proteins; it is an argument about landscapes.

From a string of letters to a machine#

A protein starts as a linear polymer. Twenty standard amino acids share a common backbone — an amino group, a central carbon, a carboxyl group — and differ only in the side chain hanging off that central carbon. Condensation of one amino group with the next carboxyl group gives the peptide bond, and a few hundred of those give a polypeptide.

The order of side chains along that chain is the primary structure, and it is what DNA replication exists to preserve. A gene is a recipe for a sequence; the sequence is a recipe for a shape; the shape is what actually does the chemistry. Copying fidelity matters because a single substituted letter can change the shape and therefore the machine.

Four levels are conventionally distinguished:

• Primary — the covalent sequence of residues, N-terminus to C-terminus.
• Secondary — local, repeating backbone geometry stabilised by hydrogen bonds between backbone amide and carbonyl groups: the $\alpha$-helix (a right-handed coil, ~3.6 residues per turn, hydrogen bond from residue $i$ to $i+4$) and the $\beta$-sheet (extended strands hydrogen-bonded side by side, parallel or antiparallel), plus the turns and loops that connect them.
• Tertiary — the full three-dimensional arrangement of one chain, with distant parts of the sequence packed against each other.
• Quaternary — several folded chains assembled into one functional complex. Haemoglobin is four subunits; the cooperative oxygen binding that makes it useful is a quaternary property that no single subunit has.

Secondary structure is worth pausing on, because it is cheap. Helices and sheets are the two backbone geometries that satisfy essentially every backbone hydrogen bond while keeping the peptide planes and side chains out of each other's way. They form fast, locally, often before the chain has decided on its global topology — which gives the folding process a small set of prefabricated parts to assemble rather than a few hundred independent hinges.

Watching a chain bury its greasy residues#

The simplest model that captures the essential physics throws away almost everything. Put the chain on a two-dimensional square lattice, and sort the twenty amino acids into just two classes: H (hydrophobic) and P (polar). Score a conformation by counting how many H residues touch each other on the lattice without being neighbours in the chain, and call each such contact $-1$ of energy. That is the HP model, and it folds.

Three things to do here:

• Press Fold and watch the energy trace. The chain starts extended, thrashes while the annealing temperature is high, then compacts as the temperature drops. The violet line is the instantaneous energy and the green dashes are the best structure found so far. The collapse is not gradual and uniform — energy falls in bursts as whole segments of chain snap into place.
• Switch to blind and press Fold again. Now each conformation is drawn fresh at random rather than being a small modification of the last one. The energy trace rattles high and the green best-so-far line stalls early. Both searches are given exactly the same budget — three thousand conformations — and the funnelled one routinely buries three or four more contacts than the blind one ever finds. That gap between the two traces is Levinthal's paradox, reduced to something you can watch in ten seconds.
• Press New sequence a few times. Each random H/P pattern folds to a different compact shape, and some fold far better than others. Sequences with hydrophobic residues spaced to fall on the same side of a compact core reach deep energies; sequences with them scattered do not. The information for the fold is in the sequence and nowhere else — which is exactly what Christian Anfinsen showed in the 1960s when he denatured ribonuclease, removed the denaturant, and watched it spontaneously recover full activity with no cellular machinery present at all.

Why hydrophobic collapse is really about water#

The lattice model's scoring rule is a stand-in for the dominant force in real folding, and that force is regularly described in a way that is subtly wrong.

Folding happens when the free-energy change is negative:

$$\Delta G = \Delta H - T\Delta S < 0$$

Folding is bad for the chain's entropy — an extended polymer has enormous conformational freedom and a folded one has almost none, so $\Delta S_{\text{chain}}$ is large and negative, costing tens of kcal/mol. Something has to pay for that.

The usual story is that nonpolar side chains "attract" each other and huddle together. They do attract, weakly, by dispersion forces — but that is not the main term. The dominant contribution comes from the solvent.

An exposed nonpolar surface cannot hydrogen-bond with water. The water molecules touching it must therefore arrange themselves to keep their hydrogen bonds pointing at each other, forming an ordered, cage-like shell. Those molecules are not chemically trapped, but they are entropically constrained: far fewer orientations are available to them than in bulk water. When two nonpolar surfaces bury each other, both shells are dismantled and their water is released back into the bulk, where it can tumble freely again.

So the driving force is $\Delta S_{\text{water}} > 0$ — the entropy of the solvent, not of the protein, and not an enthalpic attraction between the side chains. Written out:

$$\Delta G_{\text{fold}} = \underbrace{\Delta H_{\text{bonds}}}_{\text{small, mixed}} - T\underbrace{(\Delta S_{\text{water}} + \Delta S_{\text{chain}})}_{\text{positive} \;+\; \text{negative}}$$

Two consequences fall straight out of that framing. First, the hydrophobic effect gets stronger with modest heating over a range where enthalpic forces weaken, because the term that matters is multiplied by $T$. Second, the whole balance is precarious: the two large entropy terms nearly cancel, and typical globular proteins are stable by only about 5–15 kcal/mol — the free energy of a couple of hydrogen bonds. A protein is not a rock. It is a structure held together by a rounding error between two enormous opposing numbers, which is why mild heat, a pH shift, or a single mutation can undo it.

Hydrogen bonds, salt bridges, disulfide cross-links and van der Waals packing all contribute, mostly by specifying which compact structure wins once collapse has happened. But the collapse itself is water pushing the greasy parts together.

The funnel, and what happens when it is bumpy#

The modern picture, developed by Joseph Bryngelson, Peter Wolynes, José Onuchic, Ken Dill and others through the late 1980s and 1990s, replaces "the folding pathway" with a funnel. Plot free energy vertically and conformational entropy horizontally: the top is wide, because there are astronomically many unfolded shapes at similar high energy; the bottom is a narrow point, the native state. Any chain anywhere on the rim is surrounded by downhill directions.

This is the same object as the loss landscape in gradient descent, and the same intuitions transfer. A protein does not know where the native state is any more than a gradient-descent iterate knows where the global minimum is; it only ever feels the local slope. What makes folding reliable is not cleverness but the shape of the surface — evolution has selected sequences whose landscapes are smooth enough that local downhill moves almost always help.

Almost always. Real landscapes are rugged, and ruggedness is where folding fails.

• Set ruggedness near 0 and drop the chain. It slides straight to the native state in a couple of hundred steps. This is the idealised funnel: no decisions, no dead ends, arbitrarily fast.
• Push ruggedness past about 0.5 and drop it again. Now the descent stalls. The chain lands in a local minimum partway down, thermal jitter is not enough to lift it back over the barrier, and it sits there — folded into something, but not into the right thing. The readout calls this a kinetically trapped state, and that is precisely the language biophysicists use.
• Hit Reset a few times at the same ruggedness. The bumps are redrawn, and the same setting sometimes folds and sometimes traps. Folding is a stochastic process; yield is a probability, not a guarantee.

The trade-off has a name: the principle of minimal frustration. A random sequence would have a landscape full of competing near-optimal contacts — a glassy surface riddled with traps. Biological sequences are not random; selection has smoothed their landscapes so that the interactions favoured locally are the same ones present in the native state. A folding protein rarely has to break a good contact to make a better one.

Cells hedge anyway. Molecular chaperones — the Hsp70 family, and the barrel-shaped GroEL/GroES chaperonin in bacteria — bind exposed hydrophobic patches on nascent or partly folded chains. They do not carry the answer; they cannot, because the answer is in the sequence. What they do is shield sticky intermediates from each other, and use ATP to unfold trapped states so the chain can fall down the funnel again. A chaperone converts a one-shot problem into a repeated one, which turns a low per-attempt success rate into a high overall yield.

When the fold goes wrong, and what AlphaFold changed#

Misfolding is not merely a failed protein; it is often a harmful one. Exposed hydrophobic surfaces are sticky, and partly folded chains can associate with each other rather than burying their own cores. Some of these aggregates are amorphous, but many converge on the same architecture regardless of the starting sequence: the amyloid fibril, a stack of $\beta$-strands running perpendicular to the fibril axis, hydrogen-bonded into an indefinitely extensible ribbon. Amyloid is remarkably stable — in many cases more stable than the native fold, which means the native state is only a kinetically protected minimum, not the global one.

Protein misfolding and aggregation are described as central features of several conditions studied in molecular medicine: amyloid-$\beta$ and tau deposits in Alzheimer's disease, $\alpha$-synuclein in Parkinson's, misfolded prion protein in the transmissible spongiform encephalopathies, and a destabilised mutant enzyme in the amyloidoses. Cystic fibrosis is a folding disease of a different kind — the common $\Delta$F508 mutation in CFTR produces a protein that fails cellular quality control and is degraded before reaching the membrane, so the defect is in trafficking a marginally stable fold rather than in the channel's chemistry. These are descriptions of mechanism, not of clinical practice; the point here is only that the same physics that makes folding work makes it fragile.

Predicting the fold from the sequence was, for fifty years, the field's defining unsolved problem. Anfinsen's experiment says the sequence contains the answer; nobody could extract it. Physics-based simulation ran into the timescale wall — folding takes milliseconds, molecular dynamics resolves femtoseconds, and twelve orders of magnitude is a lot of compute.

The breakthrough came from a different direction. DeepMind's AlphaFold2, at the CASP14 assessment in 2020, predicted structures at roughly experimental accuracy — median backbone error under an ångström for most targets, against a field that had been stuck for a decade. It is a deep neural network, and it does not simulate physics at all. It reads the evolutionary record: a multiple sequence alignment of homologous proteins across species carries covariation signals — pairs of positions that mutate together, because they touch in three dimensions and a change in one must be compensated in the other. AlphaFold learns to turn that statistical shadow of the structure, plus geometric reasoning over a learned representation of residue pairs, into coordinates.

What changed in practice: the AlphaFold Protein Structure Database now holds predicted structures for essentially every protein in UniProt, hundreds of millions of them, freely available. Structural biology went from a discipline where obtaining one structure was a multi-year project to one where a plausible structure is a starting assumption. John Jumper and Demis Hassabis shared the 2024 Nobel Prize in Chemistry for it, with David Baker for the inverse problem — designing sequences that fold to a specified shape.

Two caveats keep the achievement in proportion. AlphaFold predicts the structure, not the pathway — it says nothing about how the chain gets there, so Levinthal's question is answered by the funnel picture rather than by the network. And it predicts a single dominant conformation, while much of biology lives in the motion between conformations, in disordered regions that have no fixed fold, and in complexes and ligand-bound states where the answer depends on context. The sequence-to-structure map is largely solved. The sequence-to-behaviour map is not.