Automated Force Field Developer and Optimizer Platform: Torsion Reparameterization

Madushanka Manathunga,Alejandro Blanco-Gonzalez,Ryan Snyder,Shi Zhang,Timothy J. Giese,Andreas W. Goetz,Kenneth M. Merz, Jr.,Darrin M. York,Hasan Metin Aktulga,William Betancourt

DOI: https://doi.org/10.26434/chemrxiv-2024-lcnx1

2024-09-11

Abstract:Generic force fields such as Generalized Amber Force Field (GAFF) are widely used in protein-ligand binding simulations in structure-based drug discovery. However, the force field parameters are not always transferable across ligand molecules, and reparameterization is necessary for accurate binding free energy simulations. This is especially true for torsion parameters which are highly dependent on stereoelectronic and steric effects. Here we report a novel, flexible, and user-friendly computational tool called the Automated Force Field Developer and Optimizer (AFFDO) platform that allows generating accurate GAFF2 torsion parameters for drug-like molecules. For a given ligand, AFFDO selects the most important torsions, carries out GPU-accelerated density functional theory calculations to collect reference data and fits torsion terms using a fast gradient-based optimizer that leverages automated differentiation. We benchmark AFFDO by parameterizing a series of drug-like molecules and carrying out protein-ligand relative binding free energy (RBFE) simulations. The results show that our tool is capable of significantly improving GAFF2 torsion parameters and RBFE values within a reasonable amount of time.

Chemistry

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that in structure - based drug design, the general force field parameters do not have good transferability among different ligand molecules, especially for torsion parameters. Torsion parameters are highly dependent on stereoelectronic and steric effects. Therefore, when accurately simulating protein - ligand binding free energy, these parameters need to be re - parameterized. Although the existing re - parameterization methods can improve the accuracy of free - energy prediction, the process is complex and the computational cost is high.

To this end, the authors have developed a new platform named Automated Force Field Developer and Optimizer (AFFDO), aiming to generate accurate GAFF2 torsion parameters to improve the protein - ligand relative binding free energy (RBFE) simulation of drug - like molecules. AFFDO achieves this goal through the following steps:

1. **Ligand Fragmentation**: In order to reduce the computational cost and avoid steric clashes that occur during the torsion - scan process, AFFDO uses the proprietary FragMentor tool to break large ligands into fragments, parameterize these fragments separately, and finally transfer the torsion parameters to the original ligands.
2. **Identification of Important Torsions**: AFFDO generates a large number of molecular conformations, uses density functional theory (DFT) calculations to collect reference data, and selects torsions that have a significant impact on structural changes.
3. **Generation of Reference Data**: Use the low - cost semi - empirical method (xTB) for preliminary torsion scans, and then use GPU - accelerated DFT calculations to further optimize these scan results.
4. **Torsion Fitting**: Adopt a gradient - based optimizer and use automatic differentiation techniques for rapid convergence to minimize the squared error between quantum chemistry (QC) and GAFF2 torsion energy profiles.
5. **Generation of Output Files**: Finally, AFFDO generates a series of input files that can be directly used in the ProFESSA framework to run binding free - energy simulations.

Through the above process, AFFDO can significantly improve the quality of GAFF2 torsion parameters within a reasonable time, thereby improving the accuracy of RBFE values. This not only improves the efficiency of drug design but also reduces the need for manual intervention, enabling researchers to obtain reliable results more quickly.