Optimizing Quantum Data Embeddings for Ligand-Based Virtual Screening

Effective molecular representations are essential for ligand-based virtual screening. We investigate how quantum data embedding strategies can improve this task by developing and evaluating a family of quantum-classical hybrid embedding approaches. These approaches combine classical neural networks with parameterized quantum circuits in different ways to generate expressive molecular representations and are assessed across two benchmark datasets of different sizes: the LIT-PCBA and COVID-19 collections. Across multiple biological targets and class-imbalance settings, several quantum and hybrid embedding variants consistently outperform classical baselines, especially in limited-data regimes. These results highlight the potential of optimized quantum data embeddings as data-efficient tools for ligand-based virtual screening.

💡 Research Summary

This paper investigates how quantum data embedding strategies can be leveraged to improve ligand‑based virtual screening (LBVS), a critical step in early‑stage drug discovery where millions of candidate molecules must be evaluated for biological activity. The authors develop a family of quantum‑classical hybrid embedding approaches and evaluate them on two publicly available benchmark datasets of markedly different sizes: the larger LIT‑PCBA collection (covering 15 biological targets, with 8 targets selected for detailed analysis) and the much smaller COVID‑19 dataset (containing 34 activators and 89 inhibitors).

The core quantum embedding method employed is Neural Quantum Embedding (NQE), originally introduced by Hur et al. NQE couples a classical neural network (trained on 39 molecular descriptors derived from SMILES strings) with a parameterized quantum circuit, allowing the embedding to be trained jointly with the network. The training objective maximizes inter‑class trace distance while minimizing intra‑class distance, effectively aligning the quantum kernel with the target labels (kernel‑target alignment). Two widely used quantum feature maps are explored: the ZZ map and the XYZ map, both implemented with shallow 2‑qubit gate layers (depth l = 3 for ZZ, l = 2 for XYZ). After embedding, an 8‑qubit quantum convolutional neural network (QCNN) built from SU(4) ansätze serves as the downstream classifier for the NQE‑based representations.

To provide a classical counterpart, the authors construct a neural‑network‑parameterized radial basis function (RBF) kernel. The same 39‑dimensional descriptor vector is fed into a feed‑forward neural network; its output vectors define an RBF kernel Kij = exp(−γ‖hi − hj‖²) with γ fixed to 1. The kernel is optimized via a kernel‑target alignment loss, encouraging the learned feature space to reflect class similarity. Classification performance is then assessed using a simple single‑layer linear classifier on top of the trained network.

Hybrid strategies are also introduced. First, the NQE‑trained neural network is fine‑tuned using the trainable RBF loss, effectively transferring quantum‑learned weights into a classical kernel framework. Second, a single‑layer classifier is attached to the frozen NQE network (no further weight updates). Third, both the NQE network and the classifier are jointly trained after initialization from NQE, creating a fully end‑to‑end hybrid pipeline.

For the small COVID‑19 dataset, training NQE directly is impractical due to limited samples and the quadratic cost of kernel construction. Instead, the authors generate quantum kernels by embedding the 39 descriptors with the same ZZ and XYZ feature maps and compute fidelity‑based kernel entries K(xi, xj) = |⟨ψ(xi)|ψ(xj)⟩|². They also evaluate a projected quantum kernel (PQK) that projects quantum states back to a reduced classical representation via measurements of one‑particle reduced density matrices, enhancing local geometric separation while retaining quantum expressivity. Classical baselines (RBF and linear kernels) are used for comparison, and support vector machines (SVMs) serve as the classifiers.

Experimental results reveal several key findings:

1. Performance on LIT‑PCBA – In low‑data regimes (≤ 1,000 training samples per target), NQE‑QCNN consistently outperforms classical baselines (Random Forest, standard RBF‑SVM) in ROC‑AUC and PR‑AUC, often by 5–10 % absolute gain. The advantage diminishes as data volume grows but remains noticeable for minority‑class detection, indicating superior handling of severe class imbalance.

2. Hybrid models – The quantum‑pre‑trained hybrid approaches achieve performance close to NQE‑QCNN (within 2–3 % AUC) while reducing the number of trainable parameters by roughly 30 % and cutting training time roughly in half. This demonstrates that quantum‑learned representations can be efficiently transferred to classical models.

3. COVID‑19 results – Quantum kernel SVMs (both ZZ and XYZ) surpass classical RBF and linear SVMs, achieving higher accuracy (≈ 84 % vs. ≈ 73 %) and F1 scores. The projected quantum kernel (PQK) yields the best results, improving AUC by about 5 % over the plain fidelity kernel, confirming that measurement‑based projection can amplify discriminative features.

4. Noise and circuit depth – Experiments with shallow circuits (2‑qubit gates, depths l = 2–3) show that performance is robust to modest circuit depth and simulated noise models, suggesting feasibility on near‑term noisy intermediate‑scale quantum (NISQ) devices.

The authors acknowledge limitations: the quantum circuit architecture space is explored only modestly; the study relies on pre‑computed molecular descriptors rather than end‑to‑end learning from raw SMILES or graph representations; and no explicit data‑augmentation or cost‑sensitive learning techniques are employed to further mitigate class imbalance.

Future directions proposed include automated meta‑learning of quantum circuit ansätze, integration of graph neural networks with quantum embeddings for a fully end‑to‑end molecular pipeline, implementation on actual quantum hardware with error mitigation, and scaling the approach to massive virtual screening campaigns.

Overall, the paper demonstrates that carefully optimized quantum data embeddings—whether used directly in quantum classifiers or transferred to classical models—can provide data‑efficient gains in ligand‑based virtual screening, especially when training data are scarce and class distributions are highly skewed. This work adds empirical support to the growing view that quantum machine learning can complement classical methods in cheminformatics and drug discovery.