Hybrid Quantum–Classical Machine Learning for Sustainable Drug Discovery: Heat Capacity-Based Drug-Likeness Evaluation

Abstract

Sustainable approaches to molecular analysis can support the development of computational methods for early stage drug discovery. This work explores whether combining quantum derived molecular representations with a classical regressor improves the prediction of heat capacity, and considers what that might mean for drug-likeness assessment. The QM9 molecular dataset is first prepared through numerical preprocessing, feature scaling, and dimensionality reduction, after which the resulting representations are encoded into a quantum circuit to produce quantum-derived features. These are combined with conventional molecular descriptors and fed into an XGBoost regressor to generate the final heat-capacity (Cv) predictions. To gauge whether this hybrid design offers any real advantage, its output is compared directly against a purely classical model, using R², MAE, and MSE as the basis for comparison. The hybrid configuration reaches an R² of approximately 0.947, versus roughly 0.923 for the classical model a gap that suggests quantum-derived representations can meaningfully sharpen prediction of this thermodynamic property, at least under the conditions tested here. Overall, the framework offers a computationally grounded approach that could support molecular screening and contribute to more sustainable early-stage drug-discovery strategies.