Software Tools for Force-Field Development

The optimization of force-field parameters is usually a multi-dimensional task in terms of the number of variables and also a multi-objective problem in terms of the desired target properties. It is also a problem that is usually broken down in a few steps, such as the optimization of covalent terms and the optimization of non-bonded interactions. In the optimization philosophy of most biomolecular force fields, the parameters corresponding to the bonded terms, i.e. bond-stretching, angle-bending and torsional potentials, are usually adjusted based on target values derived from quantum mechanics (equilibrium geometries, force constants, torsional-energy profiles). Among those, only the torsional potentials impose a reasonable challenge for fitting the classical potential to match the desired quantum-mechanical energy profile. Although there are many automated protocols available for carrying out the fitting of the torsional potential, the coupling with the 1-4 Lennard-Jones (LJ) interactions has been ignored or treated in a sub-optimal fashion. In order to address this problem, we have recently developed the profilerTools, a computational tool  to simultaneously optimize both the 1-4 Lennard-Jones interactions and the torsional profiles. The code implements two different optimization methods: a deterministic linear least-squares based self-consistent scheme and a stochastic covariance matrix adaptation evolutionary strategy. Both schemes were validated and provide similar converged solutions. The optimization of non-bonded interaction parameters is usually a more difficult problem since it involves the reproduction of condensed-phase properties due to the effective nature of the pairwise potential. To facilitate this procedure, we have developed a grid-based inference strategy. In principle, the method can be used to optimize any combination of parameters. Consider for instance the optimization of the C₆ and C₁₂ LJ-interaction parameters of a given atom type and suppose that the target properties are the liquid density, the heat of vaporization and the surface-tension coefficient. The simplest, but most computationally expensive approach, would be the brute-force screening by simulating all combinations of C₆ and C₁₂ in a given 2D grid that would hopefully contain the optimal solution. Unfortunately, this approach is extremely expensive and could also fail if the optimal solution is outside the defined grid space. Our solution to this problem is based on the simulation only a few grid points, followed by the inference of all other grid points using a machine-learning algorithm. The entire procedure is automated, including the submission of simulations, property calculations, the addition of new simulation points to the grid when necessary, and the adaptation of the grid region in case the optimal solution is inferred to be outside the initial grid. Both tools presented here aim to help force-field developers to achieve optimized parameters with reduced human effort and are available in our github repository (http://github.com/mssm-labmmol).