QuEra’s Aquila Neutral Atom Processor is currently being used to transform molecular data in a new approach to toxicity prediction, taking advantage of the molecular toxicity dataset, a specific collection of molecular structures used in an experiment, to test whether quantum computers can generate more informative features than classical methods. Instead of directly determining whether a molecule is toxic or not, the quantum system’s measurements become new data points that are fed into standard machine learning classifiers. The researchers found that these quantum features often add information that was missed by classical quantum features. The experiment was run through Kipu Quantum’s Hub platform and processed 122 samples from the dataset. About two-thirds of them were labeled as non-toxic and one-third as toxic, and the researchers aimed to improve the accuracy of identifying potentially harmful compounds by re-expressing the molecular data in a richer format.
Quantum feature extraction of molecular toxicity
Subtle but important advances in machine learning are emerging from the intersection of quantum computing and molecular toxicity prediction. Researchers have demonstrated that quantum processors can reshape data in ways that increase the accuracy of classical machine learning models. The core of this research focuses on transforming molecular data into a format that can be more easily interpreted by algorithms designed to identify potentially harmful compounds, bypassing the limitations of traditional feature engineering. This experiment utilized the molecular toxicity dataset, a dataset used to test predictive models in this field. Building on their previous success with superconducting quantum chips, the research team investigated whether neutral atomic quantum hardware could achieve similar results. Unlike digital quantum computers that rely on gate sequences, neutral atom machines exploit the collective dynamics of atoms held in place by laser beams. We then compared these to traditional features using several machine learning algorithms, including XGBoost and Google Research’s TabFM.
The results revealed consistent trends. The quantum features matched or exceeded the classical baselines tested. While we saw only a small improvement in simple accuracy, balanced accuracy and F1 score, which are important metrics for such an imbalanced dataset, showed significant improvement, with quantum features increasing balanced accuracy to about 0.7 and F1 score to about 0.6.
Hardware providers such as QuEra, D-Wave, Pasqal, Atom Computing, Oratomic, and Planqc believe that today’s industrial applications need to be considered with immediate added value for their customers.
Implementing a Neutral-Atom processor using QuEra’s Aquila
The pursuit of quantum advantages in machine learning is diversifying beyond superconducting circuits, with neutral atom processors emerging as a promising alternative. Recent work by QuEra Computing researchers demonstrates a new application of the Aquila processor to transform molecular data to enhance toxicity prediction. This approach is not intended to replace classical machine learning, but rather to utilize molecular toxicity datasets to enhance machine learning with features extracted through quantum processing. The researchers assigned each molecule in the dataset a specific driving schedule, arranged 200 atoms on a grid, and controlled their interactions through laser excitation. This process was performed on Aquila via the Kipu Quantum Hub platform, producing a statistical portrait of each molecule’s atomic response. The researchers processed 122 samples, approximately two-thirds labeled non-toxic and one-third labeled toxic, to train and test the model.
In particular, the quantum features consistently matched or outperformed the performance of traditional baselines across a variety of classifiers, including XGBoost, CatBoost, and Google’s TabFM. The most powerful model achieved an accuracy close to 0.78 with classical features, but the addition of quantum features resulted in balanced accuracy and significantly improved F1 scores. Balance accuracy increased to approximately 0.7 and F1 score increased to approximately 0.6. The strongest performance was obtained from the quantum feature-based support vector machine, which reached an accuracy of approximately 0.78 and the best area under the curve.
Performance of quantum and conventional functions
QuEra Computing is actively researching how quantum processors can be used to power machine learning, with a particular focus on transforming data into more useful features. Their recent research focuses on leveraging the molecular toxicity dataset, a dataset used to evaluate predictive models of chemical hazards, to test new approaches to feature extraction using neutral atomic quantum hardware. The core of their method relies on encoding data representing molecules into the states of a quantum system. This system then evolves according to the laws of quantum physics, producing interactions that are difficult to reproduce classically. Importantly, this quantum evolution measurement is not used to directly determine toxicity. Instead, they become new features fed into traditional machine learning classifiers.
While we saw only a small improvement in simple accuracy, balanced accuracy and F1 score, important metrics for such an imbalanced dataset, showed significant improvements, with quantum features increasing balanced accuracy to around 0.7 and F1 score to around 0.6. The strongest performance was obtained from the quantum feature-based support vector machine, which reached an accuracy of approximately 0.78 and the best area under the curve.
Enhancing classification with quantum-derived data
The pursuit of more accurate molecular toxicity predictions is rapidly evolving, and researchers are now exploring the potential of quantum computing to improve the data used in machine learning models. Rather than directly imposing classification on quantum computers, the focus has shifted to using these systems as data transformers to extract new features that enhance classical algorithms. This approach was recently tested using QuEra’s Aquila neutral atom processor and provides a path to integrating quantum capabilities into existing machine learning workflows. The core of this technique involves encoding molecular data into the states of a multi-element quantum system. As the system evolves, driven by the collective dynamics of neutral atoms, an interaction of exponentially complex effects occurs. Measurements taken from this evolved state yield new features, a statistical portrait of the atom’s response to each molecule, which is then fed into standard machine learning classifiers.
The team processed 122 samples, labeling roughly two-thirds as non-toxic and one-third as toxic, producing about 1,000 quantum features per molecule. The results show that leveraging quantum-derived features can significantly improve performance, especially for metrics that are sensitive to imbalanced datasets. The strongest model with classical features achieved a balanced accuracy of about 0.55 and an F1 score close to 0.3, while the quantum features increased the balanced accuracy to about 0.7 and doubled the F1 score to about 0.6. Support vector machines proved to be the most powerful combination overall when driven by quantum features, reaching an accuracy of approximately 0.78. The researchers conclude that the neutral atom-based feature mapping used here achieves meaningful results, demonstrating the case that only quantum features can quantitatively improve performance, and suggesting that hardware providers need to consider today’s industrial applications with immediate added value for customers.
The difference becomes even bigger when you consider the balance between accuracy and F1 score. The classical baseline is around 0.55 for balanced accuracy and close to 0.3 for F1. This is a clear sign of a model that primarily predicts common classes, but the quantum feature increases the balanced accuracy to about 0.7, roughly doubling the F1 score to about 0.6.
