A Hardware-Feasible Quantum Machine Learning Framework for Structure-Based Virtual Screening

In structure-based virtual screening, evaluating the binding free energy of protein-ligand complexes requires accounting for both molecular conformations and spatial transformations, such as shifts and rotations, which can lead to an exponential increase in possible configurations. Classical computing approaches are limited in handling this combinatorial explosion, whereas quantum computing offers a promising alternative due to its inherent parallelism. In this study, we propose a quantum machine learning framework that encodes molecular features into quantum states and processes them through parametrized quantum gates, with all architectural and representational choices deliberately guided by near-term hardware feasibility and an explicit focus on minimal qubit counts, shallow circuit depth, and compact input representations. The model is implemented and optimized in PyTorch, and its predictive performance is examined under three conditions: ideal simulation, finite-shot sampling, and quantum-noise simulation. With six quantum circuit units, the model achieves a root-mean-square deviation of 2.37 kcal/mol and a Pearson correlation coefficient of 0.650. The predictions remain stable with 100,000 measurement shots, demonstrating compatibility with near-term quantum hardware. Although the introduction of noise slightly reduces absolute accuracy, the Pearson correlation coefficient remains stable, indicating that the ranking of ligand affinities is preserved. These results highlight a practical, scalable quantum approach that balances predictive power and robustness, providing a feasible pathway to accelerate virtual screening using moderately deep quantum circuits.