How do proteins fold so fast and so reliably?

The question

A protein is born as a long chain of amino acids, like a string of beads. To do its job — to digest food, carry oxygen, copy DNA — it must fold into a precise three-dimensional shape. Most proteins do this on their own, in a fraction of a second, and get it right almost every time.

In 1969 Cyrus Levinthal pointed out why this is astonishing. A chain of a hundred amino acids can take an astronomical number of shapes. If it tried them one by one, even at billions per second, finding the right one would take longer than the age of the universe. Yet proteins fold in microseconds to seconds. This is Levinthal's paradox.

The question is marked as partly answered. Since 2020, artificial intelligence has been able to predict the final shape of most proteins with remarkable accuracy, and the work behind it received the 2024 Nobel Prize in Chemistry. But a prediction of where the chain ends up is not an explanation of how it gets there.

What we know

In the 1950s and 60s Christian Anfinsen showed that the amino acid sequence alone contains everything needed to determine the final shape: an unfolded protein refolds into the same shape in a test tube, without help.
Proteins don't search at random. The modern picture is an energy landscape shaped like a funnel: from any starting point, the chain can slide downhill towards its stable shape along many routes, forming partial structures that help further folding.
In 2021 AlphaFold 2 predicted structures with near-experimental accuracy for most proteins, and its database now holds predicted shapes for hundreds of millions of proteins.
Supercomputers built for the purpose can simulate the folding of small proteins atom by atom. In 2011 a group watched a dozen small proteins fold in simulation, matching experiments.
In living cells, many proteins need help: molecular machines called chaperones protect them from clumping while they fold.
Not all proteins fold. Roughly a third of human proteins have long regions with no fixed shape, and these "disordered" parts are essential to how they work.
When folding goes wrong, misfolded proteins can clump together. This is linked to Alzheimer's, Parkinson's and prion diseases.

Main hypotheses

The funnel is the whole answer

On this view, Levinthal's paradox is solved in principle: evolution has selected sequences whose landscapes are smooth funnels, so folding is fast and reliable. What remains are details for each protein.

Pathways matter

Others think that the real routes, the order in which parts of the chain form and the intermediate states along the way, are as important as the landscape. Two proteins with the same final shape can fold very differently, and the route can decide whether a protein folds or clumps.

Learned but not understood

AI systems may have captured the rules of folding in a form that works but can't be read by humans. Some researchers hope to extract physical principles from them; others doubt that they have learned physics rather than patterns in known structures.

Where understanding ends

We can predict where folding ends but not, in general, how it gets there. For most proteins we can't say which parts form first, what the intermediate states look like, or how fast the protein will fold. Simulations remain limited to small, fast proteins; larger ones fold too slowly to simulate in full.

The AI breakthrough also has clear limits. It predicts one static shape, while proteins move, switch between shapes and work through those changes. It is much less reliable at predicting how a single mutation changes a protein, whether a protein will misfold, or what disordered regions do. These are exactly the questions that matter for diseases.

And there is a deep puzzle linked to P vs NP. Finding the lowest-energy shape is, in simplified mathematical models, among the hardest problems computers know. Nature solves it in milliseconds. The likely answer is that evolution only kept the rare sequences that are easy to fold, but this means that what a protein "solves" isn't the general problem at all, and we still don't fully understand what makes a sequence foldable.

What would settle it

Experiments or simulations that follow the full folding path of large proteins, atom by atom, and predict it in advance.
Models that predict not only the final shape but folding speed, intermediate states and the effects of mutations, and that explain why they work.
A clear understanding of how misfolding begins in the cell, tested by preventing it.