Prediction of Membrane Permeation of Drug Molecules by Combining an Implicit Membrane Model with Machine Learning

Lipid membrane permeation
of drug molecules was investigated with
Heterogeneous Dielectric Generalized Born (HDGB)-based models using
solubility-diffusion theory and machine learning. Free energy profiles
were obtained for neutral molecules by the standard HDGB and Dynamic
HDGB (DHDGB) to account for the membrane deformation upon insertion
of drugs. We also obtained hybrid free energy profiles where the neutralization
of charged molecules was taken into account upon membrane insertion.
The evaluation of the predictions was done against experimental permeability
coefficients from Parallel Artificial Membrane Permeability Assays
(PAMPA), and effects of partial charge sets, CGenFF, AM1-BCC, and
OPLS, on the performance of the predictions were discussed. (D)­HDGB-based
models improved the predictions over the two-state implicit membrane
models, and partial charge sets seemed to have a strong impact on
the predictions. Machine learning increased the accuracy of the predictions,
although it could not outperform the physics-based approach in terms
of correlations.