Toward Automated Benchmarking of Atomistic Force Fields: Neat Liquid Densities and Static Dielectric Constants from the ThermoML Data Archive

Abstract

Atomistic molecular simulations are a powerful way to make quantitative predictions, but the accuracy of these predictions depends entirely on the quality of the forcefield employed. While experimental measurements of fundamental physical properties offer a straightforward approach for evaluating forcefield quality, the bulk of this information has been tied up in formats that are not machine-readable. Compiling benchmark datasets of physical properties from non-machine-readable sources requires substantial human effort and is prone to the accumulation of human errors, hindering the development of reproducible benchmarks of forcefield accuracy. Here, we examine the feasibility of benchmarking atomistic forcefields against the NIST ThermoML data archive of physicochemical measurements, which aggregates thousands of experimental measurements in a portable, machine-readable, self-annotating IUPAC-standard format. As a proof of concept, we present a detailed benchmark of the generalized Amber small molecule forcefield (GAFF) using the AM1-BCC charge model against experimental measurements (specifically bulk liquid densities and static dielectric constants at ambient pressure) automatically extracted from the archive, and discuss the extent of data available for use in larger scale (or continuously performed) benchmarks. The results of even this limited initial benchmark highlight a general problem with fixed-charge forcefields in the representation low dielectric environments such as those seen in binding cavities or biological membranes.

Blinded predictions of distribution coefficients in the SAMPL5 challenge

“…In this case charging corrections fail to improve the estimations. As pointed out by Beauchamp et al [43], the solvation of polar solutes in a non-polar solvent such as cyclohexane is expected to be systematically underestimated since the lack of polarisability yields a cyclohexane model with a static dielectric constant of cyclohexane of about 1.0, whereas experimental data indicates a value closer to 2.0. This is expected to cause a significant error in the computed solvation free energy of polar solutes in a non-polar solvent.…”

Calculating Partition Coefficients of Small Molecules in Octanol/Water and Cyclohexane/Water

“…In an effort to automate our simulations, we have created Solvation Toolkit to generalize workflows across different software packages. It aims to deliver a simple tool to set mixtures of arbitrary combinations of solutes and solvents for use in popular molecular dynamics software packages such as GROMACS 69–75 and AMBER This piece of software relies on other software packages including OpenMolTools, 76,77 OpenEye’s Python toolkits, 78–88 AmberTools, 89–94 Packmol, 95 ParmEd, 96 and MDTraj. 97 …”

Open Force Field Consortium: Escaping atom types using direct chemical perception with SMIRNOFF v0.1

“…To further validate the SMIRNOFF implementation on a more diverse set of molecules when our prototype SMIRNOFF99Frosst force field was available (Section 5.2), we computed densities and dielectric constants for a diverse set of small molecules previously considered by Beauchamp et al [6], as further discussed in Section 5.2.6. To do this validation, we updated the pipeline of the previous work to optionally employ a SMIRNOFF force field (and put the resulting code in https://github.com/kyleabeauchamp/solutionffbench and our SI) then used it to repeat the benchmark both with GAFF and our SMIRNOFF99Frosst (version 1.0.7).…”

“…To test our SMIRNOFF99Frosst prototype force field, we performed an extensive benchmark on density and dielectric constant calculations of pure solvents, as well as an even larger test on the FreeSolv database of hydration free energies, comparing SMIRNOFF99Frosst results with those from GAFF. As discussed in Section 4.5.2, density/dielectric constant calculations used the test set from prior work [6] with minor updates to the code; we repeated all calculations with both GAFF and SMIRNOFF99Frosst to ensure the only difference was the force field. Results are shown in Figure 10.…”

“…Thus, force fields demonstrate their considerable success even in blind predictions, such as on host-guest [64,65,105] and protein-ligand binding affinities [1,7,52,61,75,77], hydration free energies [21,63], partition and distribution coefficients [3], ligand binding modes and activity [15], and others. Extensive retrospective tests for other properties such as dielectric constants [6,19,32,69] and perturbations in protein stability [80] are also worth noting given their success. It seems safe to say these relatively simple models have succeeded far beyond original expectations, likely in part because of their strong physical basis, careful and time-consuming attention paid in parameterization, and a reasonable balance of speed versus accuracy for many applications.…”