Molecular Dynamics Fingerprints (MDFP): Machine Learning from MD Data To Predict Free-Energy Differences

While the
use of machine-learning (ML) techniques is well established in cheminformatics
for the prediction of physicochemical properties and binding affinities,
the training of ML models based on data from molecular dynamics (MD)
simulations remains largely unexplored. Here, we present a fingerprint
termed MDFP which is constructed from the distributions of properties
such as potential-energy components, radius of gyration, and solvent-accessible
surface area extracted from MD simulations. The corresponding fingerprint
elements are the first two statistical moments of the distributions
and the median. By considering not only the average but also the spread
of the distribution in the fingerprint, some degree of entropic information
is encoded. Short MD simulations of the molecules in water (and in
vacuum) are used to generate MDFP. These are further combined with
simple counts based on the 2D structure of the molecules into MDFP+.
The resulting information-rich MDFP+ is used to train ML models for
the prediction of solvation free energies in five different solvents
(water, octanol, chloroform, hexadecane, and cyclohexane) as well
as partition coefficients in octanol/water, hexadecane/water, and
cyclohexane/water. The approach is easy to implement and computationally
relatively inexpensive. Yet, it performs similarly well compared to
more rigorous MD-based free-energy methods such as free-energy perturbation
(FEP) as well as end-state methods such as linear interaction energy
(LIE), the conductor-like screening model for realistic solvation
(COSMO-RS), and the SMx family of solvation models.