Synergies between Quantum Mechanics and Machine Learning for Advancing Pharmaceutical Research

The drug development process is resource-intensive, often costing billions and taking over a decade, yet many candidates still fail in late-stage trials. This thesis addresses key bottlenecks in early-stage drug discovery—such as navigating chemical space, modeling molecular interactions, and predicting biological properties—by integrating quantum chemistry and machine learning to develop more accurate and scalable computational methodologies. The analysis of the Aquamarine (AQM) dataset, designed to capture the interplay between molecular conformations, solvation effects, and non-covalent interactions, is presented as a key milestone for future machine learning models dealing with solvation effects for relevant molecules in medicinal chemistry. The results of the analysis reveal that many-body dispersion effects and implicit solvation significantly influence molecular geometries, reinforcing the necessity of accurate modeling for reliable predictions in biological environments. In a similar direction, the thesis introduces also a photonic quantum simulation framework for studying full Coulomb interactions between quantum Drude oscillators as a way to study dis- persion beyond the dipole approximation typical of current models. This study uncovers non-trivial quantum effects, including the formation of entangled Schrödinger cat states during binding and offering insights into the fundamental nature of dispersion interactions. Moving from fundamental problems to more practical applications, the Quantum Inverse Mapping (QIM) framework is introduced to establish a direct, differentiable connection between quantum mechanical properties and molecular structures. This enables multi-objective molecular design and generation of transition path initializations, demonstrating its utility in navigating chemical spaces for different tasks. Finally, the thesis explores the role of quantum chem- istry data in enhancing deep learning models for ADMET property modeling. A systematic study on Graph Transformer reveals that pretraining on atom-level quantum properties improves the model’s representation, leading to superior performance. Collectively, these contributions bridge quantum chemistry with machine learning to address key challenges in molecular exploration, electronic structure calculation, and biological property modeling, advancing computational methodologies for rational drug discovery.