Premium
Combining machine learning and quantum mechanics yields more chemically aware molecular descriptors for medicinal chemistry applications
Author(s) -
Tortorella Sara,
Carosati Emanuele,
Sorbi Giulia,
Bocci Giovanni,
Cross Simon,
Cruciani Gabriele,
Storchi Loriano
Publication year - 2021
Publication title -
journal of computational chemistry
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.907
H-Index - 188
eISSN - 1096-987X
pISSN - 0192-8651
DOI - 10.1002/jcc.26737
Subject(s) - molecule , force field (fiction) , chemistry , drug discovery , density functional theory , molecular mechanics , computational chemistry , hydrogen bond , in silico , intuition , quantum chemistry , molecular descriptor , molecular dynamics , statistical physics , chemical physics , quantum mechanics , quantitative structure–activity relationship , physics , stereochemistry , organic chemistry , biochemistry , philosophy , supramolecular chemistry , epistemology , gene
Molecular interaction fields (MIFs), describing molecules in terms of their ability to interact with any chemical entity, are one of the most established and versatile concepts in drug discovery. Improvement of this molecular description is highly desirable for in silico drug discovery and medicinal chemistry applications. In this work, we revised a well‐established molecular mechanics' force field and applied a hybrid quantum mechanics and machine learning approach to parametrize the hydrogen‐bonding (HB) potentials of small molecules, improving this aspect of the molecular description. Approximately 66,000 molecules were chosen from available drug databases and subjected to density functional theory calculations (DFT). For each atom, the molecular electrostatic potential (EP) was extracted and used to derive new HB energy contributions; this was subsequently combined with a fingerprint‐based description of the structural environment via partial least squares modeling, enabling the new potentials to be used for molecules outside of the training set. We demonstrate that parameter prediction for molecules outside of the training set correlates with their DFT‐derived EP, and that there is correlation of the new potentials with hydrogen‐bond acidity and basicity scales. We show the newly derived MIFs vary in strength for various ring substitution in accordance with chemical intuition. Finally, we report that this derived parameter, when extended to non‐HB atoms, can also be used to estimate sites of reaction.