Predicting the sites and energies of noncovalent intermolecular interactions using local properties.

Feed-forward artificial neural nets have been used to recognize H-bond donor and acceptor sites on drug-like molecules based on local properties (electron density, molecular electrostatic potential and local ionization energy, electron affinity, and polarizability) calculated at grid points around the molecule. Interaction energies for training were obtained from B97-D and ωB97X-D/aug-cc-pVDZ density-functional theory calculations on a series of model central molecules and H-bond acceptor and donor probes constrained to the grid points used for training. The resulting models provide maps of both classical and unusual H- and halogen-bonding sites. Note that these reactions result even though only classical H-bond donors and acceptors were used as probes around the central molecules. Some examples demonstrate the ability of the models to take the electronics of the central molecule into consideration and to provide semiquantitative estimates of interaction energies at low computational cost.