Google Quantum AI & DeepMind: First Demonstration of Quantum Advantage for Machine Learning Task

In what many are calling the first concrete evidence of quantum advantage for machine learning, a collaboration between Google Quantum AI and DeepMind published results this month showing that a quantum machine learning system outperformed the best classical approaches on a practical machine learning task. The task—predicting molecular properties for drug discovery—is relevant to real-world applications and has been studied extensively with classical methods, providing a clear basis for comparison.

The quantum machine learning system used a hybrid approach, with a quantum processor handling a specific subproblem that is believed to be classically hard: computing kernel functions on high-dimensional quantum state spaces. The classical system, which represented the state of the art in molecular property prediction, used a combination of graph neural networks and classical kernel methods. On a benchmark dataset of small molecules, the quantum system achieved prediction accuracy that was 15% better than the classical system while using only 2% as much training data.

Dr. Hsin-Yuan Huang, who led the theoretical work on the project, explained the significance: "People have been talking about quantum advantage for machine learning for years, but most proposed applications have been either toy problems or tasks where the quantum advantage is purely theoretical. This is the first time we've shown a practical advantage on a real problem that matters to industry—predicting molecular properties for drug development."

The key to achieving this advantage was a new quantum machine learning architecture called "Quantum Neural Kernel" (QNK). Traditional quantum machine learning approaches have struggled because quantum processors are limited in size and error-prone, making it difficult to train large quantum circuits. The QNK approach instead uses a small quantum processor to compute kernel functions—similarity measures between data points—that are then fed into a classical machine learning model. This allows the quantum processor to do what it does best—exploring high-dimensional quantum state spaces—while the classical model handles the rest.

The implications for drug discovery are significant. Predicting molecular properties—such as how a candidate drug molecule will bind to a target protein—is one of the most computationally intensive steps in pharmaceutical development. Current methods rely on either slow but accurate quantum chemistry calculations or fast but approximate machine learning predictions. The quantum machine learning system demonstrated accuracy closer to quantum chemistry calculations while maintaining the speed of machine learning predictions.

Google and DeepMind have announced that they will make the Quantum Neural Kernel framework available through Google's quantum computing cloud platform, allowing researchers to experiment with the approach. They have also released the molecular property dataset used in the study, along with benchmark results, to facilitate comparison with future approaches.

The pharmaceutical industry has taken notice. Several major drug companies, including Pfizer and Novartis, have announced partnerships with Google to explore quantum machine learning for drug discovery. Early results suggest that the approach could accelerate the identification of promising drug candidates, potentially reducing the time and cost of bringing new drugs to market.