Equivariant learning leveraging geometric invariances in 3D molecular conformers for accurate prediction of quantum chemical properties

Deciphering the intricate interplay between the three-dimensional geometrical conformation of molecules and their thermodynamic properties is a central quest in molecular chemistry, with far-reaching implications spanning diverse domains from molecular biology to medicine. In this study, we present a computational framework termed 3D molecular structure enhanced (3DMSE) that seamlessly integrates the rich structural information inherent in 3D molecular geometries with state-of-the-art machine learning algorithms to enable highly precise and computationally efficient prediction of crucial quantum chemical properties. The foundation of the 3DMSE approach lies in an equivariant learning module that adeptly captures the subtle geometric intricacies of molecular conformers while ensuring invariance to rotations and permutations. By leveraging these structurally-informed 3D embeddings, 3DMSE constructs a robust and interpretable model capable of unraveling the delicate patterns that bridge molecular geometry and thermodynamic behavior. Rigorous experimental evaluations on the widely-adopted QM9 benchmark dataset highlight the exceptional performance of our 3DMSE methodology in predicting pivotal properties such as HOMO-LUMO energy gap, dipole moment, and polarizability, markedly surpassing methods that rely solely on 2D topological features or raw 3D atomic coordinates. Through the lens of the 3DMSE paradigm, we illuminate the profound impact of molecular structure on thermodynamic properties, providing fresh insights into the underlying principles that dictate the behavior of molecular systems.