Quantum machine learning for genomics, imaging, and drug screening
In January 2025 a paper in Nature Biotechnology reported cancer-drug candidates found with help from a quantum computer. The team, led by Mohammad Ghazi Vakili and Alán Aspuru-Guzik with Insilico Medicine and others, combined a 16-qubit quantum circuit with a classical generative model to propose molecules targeting KRAS, a protein mutated in many cancers and long considered “undruggable”. Of 15 molecules synthesized, two showed promising activity in the lab. It is a real result. The question this chapter teaches you to ask is: how much did the quantum part contribute?
On the plate: what it would take to detect a quantum contribution.
The general obstacles
Everything from Chapter 124 applies to biological data. Loading a large dataset of sequences or images into quantum amplitudes costs as much as reading it (Chapter 120). Many proposed quantum speedups for linear algebra on such data were dequantized by sample-based classical algorithms (Chapter 121). Expressive quantum models suffer barren plateaus and kernel concentration (Chapters 123–124). And models that avoid these may be classically simulable, as Cerezo and colleagues argued in 2025. Genomics, imaging and screening data are classical, abundant and well served by GPUs. Quantum machine learning on such data has no established advantage. Speculative
Detecting a contribution
In the KRAS study the quantum circuit, a quantum circuit Born machine, served as a prior distribution feeding a classical long short-term memory generator, which proposed molecules filtered by classical scoring. To show the quantum prior mattered, one needs an ablation: the same pipeline with a classical prior of comparable capacity, run side by side, with the same synthesis and testing. Suppose the classical pipeline’s hit rate is 0.10 and the quantum-assisted rate 0.15, with 15 molecules synthesized per arm. The plate gives the statistical power to detect the difference. With 15 molecules, power is tiny: no study of that size can show the quantum part helped, whatever the truth. The paper’s own in-silico comparisons suggested a modest benefit from the quantum prior; independent ablations at scale have not been reported. Speculative for quantum contribution.
What would change the picture
Three things would make a quantum-ML claim in biology convincing: a careful ablation against strong classical baselines with matched capacity; enough experimental validation to have statistical power; and a reason, grounded in complexity theory, why the quantum model captures structure a classical one cannot, ideally with quantum data, such as outputs of quantum chemistry or quantum sensors, where provable advantages exist (Chapter 124). For a reader with a machine-learning background, these are the same standards you would apply to any new architecture: baselines, ablations, power, and a mechanism. Established as method.
What you now have
Why loading costs, dequantization, barren plateaus and simulability limit quantum ML on biological data, the 2025 KRAS study and its architecture, why ablations and statistical power decide whether a quantum component mattered, and the standards a convincing claim must meet.
· · ·
Three questions to carry with you.
And one to work out. Two of 15 synthesized candidates were active. What is the hit rate?
Sources: M. Ghazi Vakili et al., Nat. Biotechnol. (January 2025). E. Tang, Proc. 51st ACM STOC, 217 (2019). M. Cerezo et al., Nat. Commun. 16, 7907 (2025). S. Thanasilp et al., Nat. Commun. 15, 5200 (2024). Power computed with the normal approximation for two proportions at α = 0.05, two-sided.