Molecular electrostatic potential (ESP) surfaces are a useful tool for optimizing protein-ligand interactions in drug design. It has been shown that binding efficiency of a ligand can be greatly improved by careful modifications that lead to a better complementarity between protein and ligand electrostatics. We believe that knowledge-based scoring functions can be improved by leveraging the information about electrostatic potentials around atoms. However, time-intensive quantum mechanical calculations are required for generating molecular ESP surfaces. Therefore, to develop a fast ESP-guided knowledge-based scoring function, a quick and accurate model is required for the prediction of molecular ESPs. We will present a model for predicting ESPs around atoms (specifically, in the direction of atomic features, e.g., lone pairs, hydrogens, p orbitals, etc.) developed using graph convolutional deep neural network techniques. The network was trained on ~48,000 diverse molecules (mean heavy atom count ≈ 11). For each molecule, ESP surfaces were generated by running quantum mechanical (QM) calculations (B3LYP with 6-31G* basis set for optimization and 6-311G** basis set for energy calculation). The input layer of the model receives 209 atomic descriptors calculated using the scikit-chem library built on RDKit plus atomic connectivity matrices. The output layer returns ESP values in the direction of the atomic features. The optimized model performs very well on predicting ESP values for a validation set of ~12,000 molecules (R 2 = 0.95, p &lt;&lt; 0.001 for a correlation with ESP values derived using QM calculations). The mean absolute error in predicting ESP values is ~3 kcal/mol for the validation set suggesting that our model can provide good estimates of ESP values obtained using time-intensive QM calculations, but in a fraction of the computing time. This level of precision should also allow a successful application in guiding knowledge-based scoring functions, and we will exemplify how this may be achieved for the Protein Ligand Informatics force field (PLIff).

Artificial intelligence for predicting molecular electrostatic potentials (ESPs): A step towards developing ESP-guided knowledge-based scoring functions

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.