Can quantum computers fold proteins?
No. As of August 2026, quantum hardware has "folded" only toy grid models of about 6–10 amino acids, puzzles a laptop solves exactly by checking every option. Meanwhile ordinary deep learning (AlphaFold) already predicts real protein shapes at scale. There is no known quantum speedup for predicting protein structure, and the ordinary computing baseline is one of the strongest anywhere.
Yes — classically, with no quantum computer involved Protein structure prediction is the clearest case in this section of a problem largely solved by ordinary computing while quantum hardware was still folding toys. AlphaFold 2 (2021) cracked shape prediction for single protein chains. AlphaFold 3 (2024) extended it to how biomolecules interact. Any claim that quantum computers are needed to "solve protein folding" ignores this. The practically useful version of the problem was solved without them.
Yes — for lattice toys of ~7–10 amino acids IBM researchers stored a coarse lattice-folding model in qubits. They found the lowest-energy shapes with a variational algorithm. It is a legitimate algorithm demonstration, and the qubits needed grow only polynomially with chain length. But it is clearly not folding a protein. The model puts amino acids on a grid with simplified forces, and problems this size are solved exactly by ordinary computers in seconds.
Yes — the annealing version, since 2012, at 6 amino acids The earliest hardware protein-folding demo mapped an HP lattice model onto a D-Wave annealer as an optimization problem. It matters historically as proof that biological toy problems can be put onto quantum hardware at all. The instance can be solved by listing every option. Fourteen years later, the problem sizes shown across all platforms have grown only a little.
No established quantum speedup, even on paper Unlike factoring (Shor) or chemistry (phase estimation), protein folding has no quantum algorithm with a proven, meaningful speedup. Careful cost studies of the lattice-folding algorithms find the quantum approaches do not beat ordinary solvers at any size studied. And a quadratic, Grover-type speedup on an NP-hard search (a kind of problem with no known fast general method) is exactly the kind that error-correction costs are expected to cancel — see <a href="/reality/optimization">the optimization page</a>.
Hasn't AlphaFold already solved this?
Why care? A protein is a long chain of building blocks called amino acids. The chain folds into a 3D shape, and the shape decides what the protein does in your body. Knowing the shape helps scientists design medicines. Think of a long string of beads that crumples into the same exact knot every time. Unlike beads, amino acids push and pull on each other and on the water around them. That is what makes the shape hard to predict.
Take the version people usually mean: given the chain of amino acids, predict the 3D shape. That is largely solved, with ordinary computers. AlphaFold learned how chains map to shapes from evolutionary data and known structures. It runs on GPUs, the graphics chips also used for AI. Its predictions are accurate enough to be standard tools in structural biology.
What AlphaFold does not do is simulate the folding process. That means the path a chain takes as it folds, how it misfolds, how proteins clump together, and how floppy (disordered) regions move. These stay truly hard. But they are hard for quantum computers too. Nobody has a quantum algorithm with any shown or proven edge on folding motion either. "Unsolved by ordinary computers" does not mean "quantum solves it".
What have quantum computers actually folded?
Lattice toys. Every hardware demonstration so far uses the same recipe:
- Simplify the protein to beads on a grid, called a lattice. The HP model is a common one. It labels each bead as water-avoiding (H) or water-loving (P).
- Score each shape with simple contact energies.
- Store the choice of turns along the chain in qubits.
- Search for the lowest-energy shape as an optimization problem, using annealing or a variational circuit.
The record sizes are about 6 amino acids on a D-Wave annealer (2012) and 7–10 amino acids in gate-based experiments (IBM, 2021). Real proteins run from about 50 to thousands of amino acids. The lattice model also throws away most of the real physics.
Count them, on a simple square grid. Each new bead can turn at most 3 ways: left, right or straight. A 10-bead chain has 8 such turns, so there are at most 3^8 = 6,561 shapes. A laptop checks those in a blink. A 100-bead chain would have up to 3^98 shapes, a number with 47 digits.
So every shown instance can be solved exactly on a laptop by listing all the shapes. These experiments are useful as tests of algorithms. As biology, they show nothing beyond what ordinary computers can do.
You can inspect variational circuits of exactly this scale in the lab, and see what today's QPUs can run.
Could quantum computers ever help with proteins?
Plausibly, but not by folding them. The believable niche is quantum chemistry on small, electron-heavy pieces of the problem. Examples are metal-containing active sites and reactions of covalent inhibitors, which are drugs that form a chemical bond with their target. Another is bond-level accuracy where ordinary force fields fail. That is the drug-chemistry story, with the same needs and caveats: error-corrected hardware first. It is covered in will quantum computers revolutionize drug discovery?
For the folding answer itself to change, three things would be needed:
- A quantum algorithm with a better-than-quadratic advantage for searching shapes or simulating folding motion.
- Error-corrected hardware to run it.
- Evidence that it beats not just brute force, but AlphaFold-class learned predictors and modern molecular dynamics too.
None of the three exists today. Set your expectations there.
Classifications follow the QPU137 editorial policy: every applied label carries a date, source, scale, and hardware, or it does not render. Found an error? Report it.