Using quantum atomics and machine learning to advance picotechnology

We explore the use of machine learning to predict spectroscopic properties and interaction energies of the carbonyl groups in 225 ketones, aldehydes, imides, and amides. In the combined spirit of Density Functional Theory (DFT) and the Quantum Theory of Atoms in Molecules (QTAIM), but with an eye toward eventually using databases of transferable fragment densities, we limit the training data to small sets of descriptors (from 18 to 48 per molecule) that are based on topological features in the total charge density, rho, and/or its Laplacian, del 2 rho. We obtain a mean absolute error under 1% for carbonyl stretching frequencies, and just over 1% for C-13 NMR shifts. Predicting interaction energies with a model nucleophile (F-1 ion) is significantly more challenging. Mean absolute errors just over 3 kcal/mol were obtained for covalent bond formation energies. Similar mean absolute errors were obtained for much weaker van der Waals interaction energies. We also conducted a stress test to see if our small molecule-based machine learning could predict covalent bond formation energy (by nucleophile F-1 to substrate 3-hydroxypropanal) in a model of the active site of the E. coli enzyme, D-fructose-6-phosphate aldolase. We again found an error of 3.0 kcal/mol in the prediction of this covalent binding energy using just Laplacian data as descriptors and the same machine learning model developed for isolated small carbonyl molecules.