178
Part Four · Quantum biology

Quantum computing for drug discovery and protein folding (overview; detail in Part 5)

In 2024 the Nobel Prize in Chemistry went to Demis Hassabis and John Jumper for AlphaFold, which predicts protein structures from sequence with near-experimental accuracy, and to David Baker for designing new proteins. Neither used a quantum computer. So where, if anywhere, would quantum computing help drug discovery? The honest answer is narrow but real: problems that are fundamentally about electrons in strongly correlated arrangements. This chapter maps the pipeline and marks where quantum methods might plug in; Part 5 goes into depth.

On the plate: the drug-discovery pipeline, stage by stage.

The pipeline

Bringing a drug to market typically takes 10–15 years and costs on the order of a billion US dollars or more, most of it in clinical trials, and roughly nine in ten candidates entering human trials fail. The early, computational stages are: choosing a target protein; finding its structure; screening millions of molecules for ones that bind (hit discovery); optimizing hits for potency, selectivity and drug-like properties (lead optimization); and predicting absorption, distribution, metabolism, excretion and toxicity (ADMET). Computation already saves time at every stage, through physics-based simulation and, increasingly, machine learning. Established

Where would quantum help?

Choose a stage: . For each, the plate shows the classical state of the art, the quantum proposal and its status. The pattern: where the bottleneck is sampling, data or approximate force fields, classical methods and machine learning dominate. Where it is the electronic structure of a small, strongly correlated region, such as a metal cofactor in an enzyme active site, or a reaction pathway that DFT cannot rank reliably (Chapter 159), quantum computers could eventually give reference energies, plugged into QM/MM (Chapter 160) and used to train cheaper models. Speculative for practical advantage in drug discovery.

Honest expectations

Binding affinity needs accuracy around 1 kcal/mol: that error changes a predicted dissociation constant by a factor of about 5. Electronic-structure error is only one contributor; conformational sampling, solvation and protein flexibility often matter more. Resource estimates (Chapter 143) put useful chemistry calculations at around a million physical qubits, and the most promising targets are small, hard active sites rather than whole drug–protein complexes. Alexander Dalzell and colleagues’ 2023 survey concludes that chemistry and materials remain the most credible application area, with end-to-end advantage not yet shown. Expect quantum computers, if they arrive, to be one specialized instrument among many in a pipeline dominated by classical computing and AI. Speculative

What you now have

The drug-discovery pipeline and its costs, AlphaFold’s impact, which stages are electronic-structure problems, where a quantum subroutine would sit, the accuracy target for binding, and a realistic view of timelines.

· · ·

Three questions to carry with you.

Did AlphaFold use quantum mechanics?tap
No; it is a deep-learning model trained on known structures.
Which stages are electronic-structure problems?tap
Reaction mechanisms, metal cofactors, covalent binding, and accurate interaction energies of small regions.
What else limits binding predictions?tap
Conformational sampling, solvation and protein flexibility.

And one to work out. At 298 K, RT=0.593RT = 0.593 kcal/mol. By what factor does a 1 kcal/mol error change the predicted dissociation constant, e1/RTe^{1/RT}?

factor ≈

Sources: J. Jumper et al., Nature 596, 583 (2021). Nobel Prize in Chemistry 2024. A. M. Dalzell et al., arXiv:2310.03011 (2023). S. McArdle et al., Rev. Mod. Phys. 92, 015003 (2020). Y. Cao et al., Chem. Rev. 119, 10856 (2019).

Next · Chapter 179
Origin of life and quantum arguments

Did quantum mechanics choose which way life's molecules twist?

Turn the page