Identifying which patients with kidney cancer will develop metastasis, meaning cancer that spreads to other organs, is one of the most important questions in oncology. If doctors could reliably predict this before it happens, they could intervene earlier with more aggressive treatment for high-risk patients while sparing lower-risk patients from unnecessary side effects.
This study explored whether quantum machine learning (QML), a field that combines principles of quantum physics with artificial intelligence, could improve the classification of kidney cancer biomarkers. Biomarkers here are patterns in gene and microRNA expression data that distinguish patients whose cancer spreads from those whose cancer remains localized. The specific cancer studied was clear cell renal cell carcinoma (ccRCC), the most common form of kidney cancer.
Quantum computing harnesses properties like superposition and entanglement to process information in fundamentally different ways than classical computers. While current quantum hardware is still early-stage and error-prone, researchers can simulate quantum circuits on classical computers to test whether quantum approaches offer genuine advantages for specific classification problems. This study used that simulation approach to compare quantum classifiers against their classical counterparts.
The central innovation of this work is a technique called Neural Quantum Embedding (NQE). Classical machine learning models work in standard mathematical spaces where all data points are represented as lists of numbers. Quantum algorithms work in a very different high-dimensional space called a Hilbert space. NQE bridges this gap by using a small neural network to transform the original data in a way that maximizes how well it can be separated in quantum space.
The NQE process works iteratively. The neural network transforms the input gene expression features and passes them through a quantum circuit built from gates that manipulate quantum states. The circuit measures whether pairs of data points from different classes, metastatic versus non-metastatic patients, are as far apart as possible in quantum space. This separation measure is used to update the neural network's parameters until the data is optimally arranged for quantum classification.
After embedding, two types of quantum classifiers were applied: QSVC (Quantum Support Vector Classifier) and QNN (Quantum Neural Network). QSVC uses a quantum kernel, a measure of similarity computed in quantum space, to draw boundaries between classes. QNN is a parameterized quantum circuit that learns classification rules through training. Both were compared against their classical counterparts, SVC and a standard neural network, using the same embedded data.
Three different datasets were analyzed: gene expression profiles from 424 non-metastatic and 78 metastatic ccRCC patients, and two microRNA datasets with different feature sets. Because metastatic cases were much rarer, SMOTE (Synthetic Minority Over-sampling Technique) was applied to create a more balanced training set, preventing the model from simply learning to always predict non-metastatic.
For the gene expression dataset using 9 and 10 principal components, NQE+QSVC outperformed NQE+SVC on the majority of performance metrics. The best quantum result, achieved with 10 principal components, reached an F1 score of 0.447 and a Matthews Correlation Coefficient (MCC) of 0.304. The MCC is particularly informative for imbalanced datasets because it accounts for all four categories of correct and incorrect predictions simultaneously.
For the microRNA datasets, the quantum advantage was even more pronounced. When using the miRNA profiles, NQE+QSVC outperformed the classical NQE+SVC on all 10 performance metrics simultaneously in some configurations. The smaller, more curated nature of the microRNA feature sets may have made them better suited to the quantum embedding approach, as quantum circuits work most naturally with compact, low-dimensional inputs.
Importantly, none of the models achieved near-perfect classification, which is expected given that the underlying biological data has inherent noise and that predicting metastasis from molecular snapshots is genuinely difficult. The study's contribution is demonstrating that quantum approaches can sometimes match or exceed carefully optimized classical approaches on the same data, which is a meaningful proof of concept for a field that is still maturing.
A key technical detail is that the quantum computations in this study were performed on classical computers simulating quantum circuits rather than on actual quantum hardware. This is standard practice in current quantum machine learning research because real quantum computers are still prone to errors and have limited numbers of quantum bits (qubits). Simulations allow researchers to test algorithms at small scales and verify they are theoretically sound before committing to expensive hardware runs.
The number of qubits used corresponds directly to the number of features fed into the quantum circuit. Using principal component analysis (PCA) to reduce the gene expression data from 48 original genes down to 4 to 12 principal components was essential for keeping the quantum circuit manageable in size. Larger quantum circuits require exponentially more classical computing resources to simulate.
Hyperparameter tuning was performed using grid search with five-fold cross-validation, meaning each configuration was tested on five different splits of the data to ensure that strong results were not just due to lucky data splits. This rigorous evaluation approach makes the comparative results more credible and reduces the risk of overfitting to the training data.
Quantum machine learning for cancer biomarker discovery is still at an early research stage. The systems in this study cannot yet be deployed clinically, and actual quantum hardware capable of running these circuits reliably at larger scales does not yet exist commercially. However, the field is advancing rapidly, and the algorithms validated on simulators today can be directly transferred to quantum hardware as it matures.
If quantum classifiers do offer consistent advantages over classical methods for genomic data classification, the eventual clinical payoff could be significant. Better identification of metastasis risk from a blood or biopsy sample at the time of diagnosis could guide decisions about surveillance intensity, adjuvant therapy, and clinical trial enrollment. For kidney cancer specifically, where surveillance strategies for non-metastatic patients remain debated, a reliable risk stratification tool would have immediate practical value.
This study also demonstrates an important methodological contribution: the NQE framework is general and could be applied to other cancer types and other biomarker discovery questions. As more teams adopt and refine quantum embedding approaches, the field will accumulate the evidence needed to determine definitively whether quantum computing provides a lasting advantage over classical AI for genomics.