QCS-ADME: Quantum Circuit Search for Drug Property Prediction with Imbalanced Data and Regression Adaptation

1. Problem Statement

The paper addresses the limitations of existing Quantum Circuit Search (QCS) frameworks when applied to biomedical drug property prediction, specifically ADME (Absorption, Distribution, Metabolism, and Excretion) tasks.

• The Gap: Current QCS methods (e.g., Elivagar) are primarily benchmarked on simple, balanced datasets (like MNIST). They fail to address two critical complexities inherent in pharmacological data:
  1. Class Imbalance: ADME classification tasks often have highly skewed positive-to-negative ratios (e.g., 3:1 to 6:1), causing standard metrics to favor majority classes and ignore minority classes.
  2. Regression Requirements: ADME involves continuous regression targets (e.g., drug clearance rates), whereas existing QCS scoring mechanisms rely on discrete class separability, making them unsuitable for predicting continuous values.
• The Challenge: Training-free QCS requires rapid evaluation of thousands of candidate circuits. Existing solutions for imbalance (like Quantum-SMOTE) are computationally too expensive for this search process. There is a lack of a training-free scoring mechanism that can simultaneously handle regression topology and class imbalance.

2. Methodology: QCS-ADME Framework

The authors propose QCS-ADME, a framework that adapts the standard QCS pipeline (Search Space $\to$ Strategy $\to$ Evaluation) with specific modifications for ADME tasks.

A. Workflow Overview

1. Data Preprocessing: Chemical structures (SMILES) are converted to 166-bit MACCS keys, truncated to 128 bits, and mapped to a 7-qubit quantum Hilbert space using Angle Embedding.
2. Circuit Generation: Candidate circuits are generated using a biased random sampling strategy that prioritizes hardware-efficient gates based on real-time device calibration (coherence times, gate fidelities).
3. Training-Free Scoring: Instead of full training, circuits are ranked using a composite score:
   $$\text{Score}(C) = \text{CNR}(C)^{\alpha} \times \text{RepCap}(C)$$
   • CNR (Clifford Noise Resilience): A proxy for hardware fidelity based on circuit structure.
   • RepCap (Representational Capacity): The core innovation, modified to handle imbalance and regression.

B. Key Methodological Innovations

The paper introduces two specific adaptations to the RepCap metric:

1. Handling Class Imbalance (Weighted Matrix $R_w$):

   • Problem: Standard RepCap minimizes the distance between the quantum kernel matrix ($R_c$) and an ideal reference matrix ($R_{ref}$). In imbalanced data, $R_{ref}$ is dominated by majority class pairs, leading the search to ignore minority classes.
   • Solution: The authors introduce a density-aware weighting matrix $R_w$.
     $$\text{RepCap}(C) = 1 - \frac{\|R_c - R_w \otimes R_{ref}\|_2^2}{2 \cdot n_c \cdot d_c^2}$$
   • $R_w(i, j) = w_i \cdot w_j$, where $w_i$ is the inverse frequency of the class for sample $i$. This amplifies the contribution of minority class pairs to the score, forcing the circuit to learn robust representations for underrepresented classes.

2. Handling Regression Tasks (Gaussian Similarity):

   • Problem: Regression requires preserving continuous relationships, not just discrete class separation.
   • Solution: The reference matrix $R_{ref}$ is redefined using a Gaussian Radial Basis Function (RBF) kernel based on target values ($y$):
     $$R_{ref}(i, j) = \exp\left(-\frac{\|y_i - y_j\|^2}{2\sigma^2}\right)$$
   • This formulation encourages the quantum circuit to map inputs with similar target values to quantum states that are close in Hilbert space (isomorphic geometry), effectively generalizing the metric from discrete matching to continuous similarity.

3. Key Contributions

1. Novel Training-Free Scoring for Regression: The first QCS method to generalize Representational Capacity from discrete classification to continuous regression using Gaussian-weighted similarity, enabling the search for circuits that preserve continuous topologies.
2. Imbalance-Aware Search: Introduction of the $R_w$ weighting matrix, which corrects the "majority class bias" in standard QCS metrics, ensuring minority classes are prioritized during the circuit selection process without computationally expensive data augmentation.
3. Comprehensive ADME Benchmarking: Application of QCS to 12 real-world ADME tasks (8 imbalanced classification, 4 regression) and a rigorous Sim-to-Real analysis comparing noiseless simulations against execution on the IBM Rensselaer quantum processor.

4. Experimental Results

A. Correlation with Test Performance

• Classification: The revised scoring method showed a moderate to strong Spearman correlation between the pre-training score and test performance (e.g., $R=0.412$ for HIA_Hou), significantly outperforming the baseline Elivagar method which showed negligible correlation ($R \approx 0.02$) on imbalanced tasks.
• Regression: The method demonstrated a consistent negative correlation ($R \in [-0.33, -0.48]$) between the score and test MSE, confirming that higher scores reliably predict lower regression errors.

B. Performance vs. Baselines

• vs. Other QCS Methods: The QCS-ADME selected circuits (Top-25) were competitive with or superior to QuantumNAS and Elivagar across most classification metrics (Accuracy, F1, PR-AUC).
• vs. Classical ML:
  • Classification: QML circuits were competitive on some tasks (e.g., BBB_Martins F1-score of 0.92 vs. XGBoost 0.88) but generally lagged behind classical models (XGBoost, Random Forest) in PR-AUC.
  • Regression: A significant performance gap remained. Classical models achieved much lower MSE (e.g., 0.003 for SVM) compared to QML (0.307), highlighting the current limitations of NISQ hardware for regression.

C. Sim-to-Real Analysis

• Noise Impact: Transitioning from simulation to real IBM hardware generally degraded performance.
  • Regression: Uniformly detrimental; MSE increased by ~25-37%, and instability (variance) increased significantly.
  • Classification: Nuanced impact. While most tasks degraded, the Bioavailability_Ma task showed a counter-intuitive improvement on real hardware (PR-AUC increased from 0.52 to 0.77). The authors suggest hardware noise may act as a form of regularization for specific problem landscapes.

5. Significance and Conclusion

• Bridging the Gap: QCS-ADME successfully adapts quantum circuit search for the specific, complex requirements of drug discovery, moving beyond toy datasets to real-world, imbalanced, and regression-heavy biomedical problems.
• Reliability: The proposed scoring mechanism provides a reliable, training-free proxy for circuit performance, which is crucial given the high cost of training quantum circuits.
• Future Outlook: While the framework successfully identifies promising circuits, the performance gap with classical algorithms (especially in regression) remains a critical challenge. The paper highlights that hardware noise is a double-edged sword: it is a bottleneck for regression but can occasionally aid classification via regularization. Future work focuses on noise-aware pruning and robustness to label noise.

In summary, this paper establishes a foundational framework for applying Quantum Machine Learning to drug discovery by solving the specific algorithmic challenges of data imbalance and continuous regression within the constraints of Noisy Intermediate-Scale Quantum (NISQ) devices.