Researchers at the Cleveland Clinic have developed a hybrid quantum-classical machine learning model to predict proton affinities (PAs) of molecules with improved efficiency and competitive accuracy. The model, described in the Journal of Chemical Theory and Computation, integrates classical molecular descriptors with quantum circuits serving as feature encoders, offering a practical approach compatible with current noisy intermediate-scale quantum (NISQ) hardware.
The team trained classical ensemble models on over 1,100 molecules using a feature set of 186 descriptors, achieving a mean absolute error (MAE) of 2.47 kcal/mol—within experimental uncertainty. To reduce complexity, they encoded subsets of these features into quantum states using low-depth, parameterized circuits. This encoding revealed latent structure in the data: one quantum-encoded feature showed a two-order-of-magnitude stronger correlation to PA values than any original descriptor. The hybrid model reached an MAE of 3.29 kcal/mol in simulation and 3.63 kcal/mol on IBM’s “IBM-Cleveland” device, with significantly fewer trainable parameters than its classical counterpart.
The work highlights the utility of quantum circuits as encoders in chemistry-focused ML pipelines—offering dimensional lift and improved expressivity without requiring full quantum solvers. It supports the view that NISQ hardware can contribute meaningful enhancements when integrated strategically into hybrid workflows.
Read the full study here.
April 4, 2025
Leave A Comment