Coarse-Grained Models for Automated Fragmentation and Parametrization of Molecular Databases.

We calibrate coarse-grained interaction potentials suitable for screening large data sets in top-down fashion. Three new algorithms are introduced: (i) automated decomposition of molecules into coarse-grained units (fragmentation); (ii) Coarse-Grained Reference Interaction Site Model-Hypernetted Chain (CG RISM-HNC) as an intermediate proxy for dissipative particle dynamics (DPD); and (iii) a simple top-down coarse-grained interaction potential/model based on activity coefficient theories from engineering (using COSMO-RS). We find that the fragment distribution follows Zipf and Heaps scaling laws. The accuracy in Gibbs energy of mixing calculations is a few tenths of a kilocalorie per mole. As a final proof of principle, we use full coarse-grained sampling through DPD thermodynamics integration to calculate log POW for 4627 compounds with an average error of 0.84 log unit. The computational speeds per calculation are a few seconds for CG RISM-HNC and a few minutes for DPD thermodynamic integration.

Calculation of Diffusion Coefficients through Coarse-Grained Simulations Using the Automated-Fragmentation-Parametrization Method and the Recovery of Wilke-Chang Statistical Correlation.

We introduce a model for the calculation of diffusion coefficients using dissipative particle dynamics coarse-grained molecular simulations. We validate the model on experimental diffusion data of small organics and drug-like molecules in water. The new model relies on our automated-fragmentation-parametrization protocol for cutting molecules into fragments, which are calibrated using the COSMO-RS thermodynamic model ( J. Chem. Inf. MODEL: 2016 , 56 ( 12 ), 2361 - 2377 , DOI: 10.1021/acs.jcim.6b00003 ). By simulations over the entire CULGI database of more than 11000 molecules, we recover the decades-old empirical Wilke-Chang correlation between diffusion coefficient and molar volume. We believe this is the first demonstration of the correlation by simulation or theory. From a comparison of simulated and experimental diffusion coefficients, we find that one full time unit of coarse-grained simulation equals 64 ± 13 ps real time.