Transferable Neural Network Potential Energy Surfaces for Closed-Shell Organic Molecules: Extension to Ions

Transferable high dimensional neural network potentials(HDNNPs) have shown great promise as an avenue to increase the accuracy anddomain of applicability of existing atomistic forcefields for organic systems relevantto life science. We have previously reported such a potential (Schro''dinger-ANI) thathas broad coverage of druglike molecules. We extend that work here to cover ionicand zwitterionic druglike molecules expected to be relevant to drug discoveryresearch activities. We report a novel HDNNP architecture, which we call QRNN,that predicts atomic charges and uses these charges as descriptors in an energymodel that delivers conformational energies within chemical accuracy whenmeasured against the reference theory it is trained to. Further, wefind that deltalearning based on a semiempirical level of theory approximately halves the errors. We test the models on torsion energy profiles, relative conformational energies, geometric parameters, and relative tautomer errors

Automated summary

Transferable high dimensional neural network potentials (HDNNPs) have shown great potential in improving the accuracy and applicability of existing atomistic force fields for organic systems in life science. This study extends the previous work on a potential called Schro''dinger-ANI to cover ionic and zwitterionic druglike molecules relevant to drug discovery. A novel HDNNP architecture called QRNN is introduced, which predicts atomic charges and uses them as descriptors in an energy model to accurately calculate conformational energies. Additionally, delta learning based on a semiempirical level of theory reduces errors by approximately half. The models are tested on various aspects, including torsion energy profiles, conformational energies, geometric parameters, and tautomer errors.