Traditional approaches to specifying a molecular mechanics force field encode all the information needed to assign force field parameters to a given molecule into a discrete set of atom types. This is equivalent to a representation consisting of a molecular graph comprising a set of vertices, which represent atoms labeled by atom type, and unlabeled edges, which represent chemical bonds. Bond stretch, angle bend, and dihedral parameters are then assigned by looking up bonded pairs, triplets, and quartets of atom types in parameter tables to assign valence terms, and using the atom types themselves to assign nonbonded parameters. This approach, which we call indirect chemical perception because it operates on the intermediate graph of atom-typed nodes, creates a number of technical problems. For example, atom types must be sufficiently complex to encode all necessary information about the molecular environment, making it difficult to extend force fields encoded this way. Atom typing also results in a proliferation of redundant parameters applied to chemically equivalent classes of valence terms, needlessly increasing force field complexity. Here, we describe a new approach to assigning force field parameters terms direct chemical perception that avoids these problems, called the SMIRKS Native Open Force Field (SMIRNOFF) format. Rather than working through the intermediary of the atomtyped graph, direct chemical perception operates directly on the unmodified chemical graph of the molecule to assign parameters. In particular, parameters are assigned to each type of force field term—e.g., bond stretch, angle bend, torsion, and Lennard-Jones—based on standard chemical substructure queries implemented via the industry-standard SMARTS chemical perception language, using SMIRKS extensions that permit labeling of specific atoms within a chemical pattern. We demonstrate the power and generality of this approach using examples of specific molecules that pose problems for indirect chemical perception, and construct and validate a minimalist yet very general force field, SMIRNOFF99Frosst. We find that a parameter definition file only ~300 lines long provides coverage of all but <0.02% of a five million molecule drug-like test set. Despite its simplicity, the accuracy of SMIRNOFF99Frosst for small molecule hydration free energies and selected properties of pure organic liquids, is similar to that of the General Amber Force Field (GAFF), whose specification requires thousands of parameters. This force field provides a starting point for further optimization and refitting work to follow.

We present a methodology for defining and optimizing a general force field for classical molecular simulations, and we describe its use to derive the Open Force Field 1.0.0 smallmolecule force field, codenamed Parsley. Rather than using traditional atom typing, our approach is built on the SMIRKSnative Open Force Field (SMIRNOFF) parameter assignment formalism, which handles increases in the diversity and specificity of the force field definition without needlessly increasing the complexity of the specification. Parameters are optimized with the ForceBalance tool, based on reference quantum chemical data that include torsion potential energy profiles, optimized gas-phase structures, and vibrational frequencies. These quantum reference data are computed and are maintained with QCArchive, an opensource and freely available distributed computing and database software ecosystem. In this initial application of the method, we present essentially a full optimization of all valence parameters and report tests of the resulting force field against compounds and data types outside the training set. These tests show improvements in optimized geometries and conformational energetics and demonstrate that Parsley's accuracy for liquid properties is similar to that of other general force fields, as is accuracy on binding free energies. We find that this initial Parsley force field affords accuracy similar to that of other general force fields when used to calculate relative binding free energies spanning 199 protein−ligand systems. Additionally, the resulting infrastructure allows us to rapidly optimize an entirely new force field with minimal human intervention.

Partition coefficients describe how a solute is distributed between two immiscible solvents. They are used in drug design as a measure of a solute’s hydrophobicity and a proxy for its membrane permeability. We calculate partition coefficients from transfer free energies using molecular dynamics simulations in explicit solvent. Setup is done by our new Solvation Toolkit which automates the process of creating input files for any combination of solutes and solvents for many popular molecular dynamics software packages. We calculate partition coefficients between octanol/water and cyclohexane/water with the Generalized AMBER Force Field (GAFF) and the Dielectric Corrected GAFF (GAFF-DC). With similar methods in the past we found a root-mean-squared error (RMSE) of 6.3 kJ/mol in hydration free energies which would correspond to an error of around 1.6 log units in partition coefficients if solvation free energies in both solvents were estimated with comparable accuracy. Here we find an overall RMSE of about 1.2 log units with both force fields. Results from GAFF and GAFF-DC seem to exhibit systematic biases in opposite directions with GAFF and GAFF-DC for calculated cyclohexane/water partition coefficients.

In the recent SAMPL5 challenge, participants submitted predictions for cyclohexane/water distribution coefficients for a set of 53 small molecules. Distribution coefficients (log D) replace the hydration free energies that were a central part of the past five SAMPL challenges. A wide variety of computational methods were represented by the 76 submissions from 18 participating groups. Here, we analyze submissions by a variety of error metrics and provide details for a number of reference calculations we performed. As in the SAMPL4 challenge, we assessed the ability of participants to evaluate not just their statistical uncertainty, but their model uncertainty – how well they can predict the magnitude of their model or force field error for specific predictions. Unfortunately, this remains an area where prediction and analysis need improvement. In SAMPL4 the top performing submissions achieved a root-mean-squared error (RMSE) around 1.5 kcal/mol. If we anticipate accuracy in log D predictions to be similar to the hydration free energy predictions in SAMPL4, the expected error here would be around 1.54 log units. Only a few submissions had an RMSE below 2.5 log units in their predicted log D values. However, distribution coefficients introduced complexities not present in past SAMPL challenges, including tautomer enumeration, that are likely to be important in predicting biomolecular properties of interest to drug discovery, therefore some decrease in accuracy would be expected. Overall, the SAMPL5 distribution coefficient challenge provided great insight into the importance of modeling a variety of physical effects. We believe these types of measurements will be a promising source of data for future blind challenges, especially in view of the relatively straightforward nature of the experiments and the level of insight provided.