193
Part Five · Convergence

Protein folding: AlphaFold-style ML versus quantum approaches, honestly compared

For fifty years the protein-folding problem, predicting a protein’s 3D structure from its amino-acid sequence, was one of biology’s grand challenges. In 2020 DeepMind’s AlphaFold 2 solved it for most single-chain proteins, at accuracy rivalling experiment, and its database now holds predicted structures for over 200 million proteins. No quantum computer was involved. Meanwhile, quantum-computing papers have folded tiny lattice proteins of ten residues. This chapter compares the two honestly.

On the plate: fold a lattice protein.

Lattice models

The HP model, introduced by Ken Dill in 1985, reduces a protein to a chain of hydrophobic (H) and polar (P) beads on a square or cubic lattice. The chain cannot cross itself, and each pair of H beads that sit next to each other without being bonded lowers the energy by one unit, mimicking the hydrophobic effect (Chapter 156). Finding the lowest-energy fold is NP-hard, as Bonnie Berger and Tom Leighton proved in 1998 for the 3D model. The number of possible folds grows exponentially, about 2.64n2.64^n self-avoiding walks on the square lattice. Lattice models capture the logic of folding, burying hydrophobic residues, but nothing of real protein geometry. Established

Fold it

Choose a sequence: , and . The plate searches all self-avoiding walks exhaustively to find the true optimum, and shows a random fold for comparison. For ten residues there are a few thousand distinct folds, trivial for a laptop. In 2021 Anton Robert, Panagiotis Barkoutsos, Stefan Woerner and Ivano Tavernelli at IBM encoded lattice folding of a ten-residue peptide on 22 qubits and folded a seven-residue chain on 9 qubits of real hardware, using a variational algorithm. It was a careful demonstration of an encoding, not an advantage: classical exact search solves these instantly. Established as a demonstration.

What is actually left

AlphaFold 2 (John Jumper and colleagues, 2021) and AlphaFold 3 (2024), which also predicts complexes with DNA, RNA and small molecules, earned the 2024 Nobel Prize in Chemistry together with David Baker’s protein design. Open problems remain: dynamics and conformational change, intrinsically disordered proteins, the effects of mutations, and binding energies. Most of these are sampling or data problems, where classical computing and machine learning are advancing fast. A quantum computer could matter only where a step needs accurate electronic structure, such as a metal site or a reaction in an enzyme, not for folding itself. Quantum annealing and QAOA for lattice folding (Chapter 199) have shown no advantage. Established for AlphaFold’s success; Speculative for any quantum advantage in folding.

What you now have

The HP lattice model and its NP-hardness, exhaustive folding of a short chain, the 2021 quantum lattice-folding demonstration and its limits, AlphaFold’s achievement, and where open problems in structural biology actually lie.

· · ·

Three questions to carry with you.

What does the HP model reward?tap
Non-bonded contacts between hydrophobic beads.
Why is a 10-residue quantum fold not an advantage?tap
A laptop enumerates all its few thousand folds instantly.
What remains open after AlphaFold?tap
Dynamics, disorder, mutational effects and binding energies.

And one to work out. With about 2.638 n2.638^{\,n} self-avoiding walks of nn steps, roughly how many folds does a 20-residue chain (19 steps) have? Give log⁡10\log_{10}.

log₁₀ ≈

Sources: K. A. Dill, Biochemistry 24, 1501 (1985). B. Berger and T. Leighton, J. Comput. Biol. 5, 27 (1998). A. Robert et al., npj Quantum Inf. 7, 38 (2021). J. Jumper et al., Nature 596, 583 (2021). J. Abramson et al., Nature 630, 493 (2024). Checked by the plate: exhaustive enumeration of self-avoiding walks.

Next · Chapter 194
Simulating tunneling in enzymes on a quantum device

Watch a proton tunnel on a simulated quantum computer, then check it against the exact answer.

Turn the page