Quantum Annealing Boosts Drug-Like Molecule Generation Beyond Training Data, D-Wave and Shionogi Study Shows
核心洞察
A peer-reviewed study in Scientific Reports shows an annealing quantum computer improved a generative AI model's output for drug discovery, producing chemically valid molecules 97% of the time versus 73% for a classical-only model.
The quantum-assisted model generated drug-like molecules (QED ≥ 0.7) at 66.79%, more than double the 31.61% rate in the training data and about 55% higher than the best classical model's 43.15%.
The research, a collaboration between D-Wave (搜索) and Shionogi & Co., Ltd., used a neural hash function to bridge classical and quantum networks, with the D-Wave Advantage2 system serving as the model's sampling engine.
A newly published, peer-reviewed study in Scientific Reports offers some of the clearest evidence yet that annealing quantum computers could meaningfully improve the outputs of generative AI models for drug discovery. The research, titled "Molecular design beyond training data with novel extended objective functionals of generative AI models driven by quantum annealing computer," was conducted in collaboration between D-Wave (搜索) and Shionogi & Co., Ltd. (formerly the pharmaceutical division of Japan Tobacco). Together, the teams used a D-Wave annealing quantum computer to improve how an AI model designs new drug candidate molecules, producing compounds that were more chemically valid and more "drug-like" than those generated by an equivalent classical-only AI model.
The molecules generated by the quantum-assisted model even exceeded the quality of the very data the model was trained on. These results matter because drug discovery is notoriously slow, expensive, and prone to failure. Bringing a single new drug to market typically takes 10 to 15 years with a median cost of $708 million, according to a 2025 RAND Corporation study. Much of that expense and time is spent synthesizing and testing compounds that ultimately do not pan out. Technologies like quantum AI that can narrow the field earlier—by generating a shorter list of candidates that are more likely to be valid, drug-like, and worth pursuing in the lab—attack the problem at its most expensive stage.
The Drug-Likeness Bottleneck in Generative AI
Generative AI has shown promise as a powerful tool for the early stages of drug discovery. Instead of screening existing compound libraries, a model can generate entirely new molecular structures. However, a persistent challenge remains: generating molecules that meet all four essential criteria required for useful drugs—valid chemistry, desired activity against target proteins, acceptable chemical properties for use in the body (low toxicity, stability, easy absorption), and the ability to be easily synthesized. Reliably hitting all four criteria simultaneously is still the field's central bottleneck.
The space of chemically plausible, drug-sized molecules is estimated at more than 10⁶⁰, while the number of compounds humans can currently synthesize and test is closer to 10¹⁰. Generative models trained on the relatively tiny, synthesizable slice of that space tend to overfit, producing outputs that are either invalid chemistry or poor drug candidates. Researchers call this the drug-likeness problem, and it is one of the main reasons AI-generated molecules still need heavy filtering before any candidates can be synthesized.
Combining Classical AI with Annealing Quantum Computing
To determine whether annealing quantum computing could help, the researchers built a generative model—a transformer-based model augmented with a Boltzmann machine—and tested two versions. One ran entirely on classical hardware, and one used the D-Wave (搜索) annealing quantum computer as part of its sampling process. Comparing the two head-to-head, on the same data and architecture, isolated the contribution of D-Wave's annealing quantum hardware.
A key invention that made this pairing possible was a neural hash function (NHF), which converts the model's internal representation into the binary code the quantum hardware operates on, while also acting as a regularizer that keeps training stable. The NHF replaces the Gumbel-Softmax binarization technique used in discrete variational autoencoders. This custom-built bridge allowed the classical network and the annealing quantum system to train together as a single, coherent model rather than two loosely connected systems.
That binary code feeds into a Boltzmann-machine prior, where the D-Wave (搜索) hardware performs its work. Rather than approximating that prior with classical Markov chain (or similar) Monte Carlo methods, the researchers sampled it directly on a D-Wave Advantage2 quantum annealing system, drawing samples closer to the model's true target distribution. In effect, the annealing quantum component serves as the model's sampling engine—a role classical computers can only approximate.
Measurable Improvements in Validity and Drug-Likeness
The study demonstrated that the model using annealing quantum computing produced chemically valid molecules 97% of the time, compared with 73% for the same architecture running on a classical Boltzmann machine, and 54% for a previously published benchmark model on the same dataset.
On drug-likeness—measured by the Quantitative Estimate of Drug-Likeness (QED) score, a standard metric used to measure a compound's favorability to become an oral drug—the results were equally striking. The study defines "drug-like" molecules as those scoring QED ≥ 0.7, and reports the share of unique generated molecules meeting that bar: 31.61% for the original training data, 43.15% for the best classical-only model, and 66.79% for the annealing quantum model. In other words, the annealing quantum model generated drug-like molecules at more than double the rate present in the training data and about 55% higher than the equivalent classical-only model.
The compounds generated via the quantum-annealing generative models exhibit higher quality in both validity and drug-likeness than those generated via the fully-classical models, and even exceed the training data in terms of drug-likeness features, without any restraints and conditions to deliberately induce such an optimization.
Implications Beyond Drug Discovery
The implications of this research for AI are not limited to drug discovery. Based on the results, annealing quantum computing could act as a general-purpose stochastic generator inside generative AI, sampling more effectively from the "in-between" spaces that a model has not directly seen in training. This is exactly the capability that limits generative AI across many domains, not just molecular design.
This is not a hypothetical claim about quantum computing's future value. It is consistent with what D-Wave (搜索)'s annealing quantum computers are already doing in customer applications, addressing real-world optimization problems across manufacturing, telecommunications, retail, logistics, and defense. In each of these applications, the underlying task is the same: efficiently searching or sampling a large combinatorial space to find good solutions faster than classical methods alone. The study with Shionogi extends that same core capability—efficient sampling over vast, complex spaces—from operational optimization problems into the training and sampling process of a generative AI model itself.
For an industry accustomed to hearing about the promises of quantum computing's future potential, this study is notable for measuring something concrete: validity rates, drug-likeness scores, and a direct comparison against classical hardware running the identical model. These results show promising evidence that annealing quantum computing can act as a practical enhancement to AI today, not just a long-term bet.