Parameters for the Generalized Born Model Consistent with RESP Atomic Partial Charge Assignment Protocol

Wei Zhang, Tingjun Hou, Xuebin Qiao, Xiaojie XuThe Journal of Physical Chemistry B 2003 · DOI

TL;DR

This paper introduces a new set of parameters for the Generalized Born (GB) solvation model, specifically tailored to work with the RESP atomic partial charge assignment protocol. By using a significantly larger training dataset and a genetic algorithm for optimization, they achieve improved accuracy in predicting solvation free energies for both small organic molecules and proteins, outperforming previous GB models and offering a faster alternative to Poisson-Boltzmann calculations.

What Problem Does This Paper Address?

Accurately and efficiently modeling solvation is crucial for understanding chemical and biological processes. While explicit solvent models are accurate, they are computationally expensive. Continuum solvation models offer a faster alternative by treating the solvent as a continuous medium. The Generalized Born (GB) model is a popular continuum approach that separates solvation free energy into electrostatic and nonpolar components. However, existing GB models often struggle with accuracy, particularly for small organic molecules, and their parameters are not always consistent with common charge assignment protocols like RESP used in force fields like AMBER. This paper aims to address this by developing a new set of GB parameters that are both accurate for a wide range of molecules and compatible with the RESP protocol.

Methods

The core of this work is the development of new parameters for the Generalized Born (GB) model, specifically focusing on effective atomic radii and screen factors. These parameters are optimized using a genetic algorithm (GA). The GA searches a multi-dimensional parameter space to minimize the difference between calculated and experimental solvation free energies.

The GB model used here combines an electrostatic component (described by equations 2 and 3 in the paper) with a nonpolar component modeled using the solvent accessible surface area (SASA). The atomic partial charges, essential inputs for the GB model, are derived using the RESP (Restrained Electrostatic Potential) protocol, consistent with the AMBER force field. Effective Born radii are calculated using a pairwise summation approach. The nonpolar contribution is calculated using the MSMS program, with a default coefficient from the AMBER SANDER module.

The parameter fitting was performed in two stages: first, parameters for 17 neutral atom types were determined using a training set of neutral molecules, and then, parameters for four charged atom types were refined using a training set of charged molecules.

Datasets

The study utilized two main datasets:

  • Training Set: Comprised of 402 organic molecules, with experimental solvation free energies. This set was further divided into:
  • 328 neutral molecules
    • 30 charged molecules
  • Test Set: Used for validation, consisting of:
  • 44 simple organic molecules (36 neutral, 8 charged)
    • 15 proteins randomly selected from the Protein Data Bank (PDB).

The experimental solvation free energies were obtained at 298 K and 1 atm.

Key Insights and Ideas

  • Larger Training Sets Matter: The authors emphasize that their significantly larger training set (358 molecules) compared to previous work (e.g., Jayaram et al.'s 32 molecules) is crucial for developing more robust parameters, especially for the diverse chemical environments found in organic molecules.
  • GA as an Effective Optimizer: The genetic algorithm proved to be a powerful tool for optimizing the multi-dimensional parameter space of the GB model, handling potentially coupled variables effectively.
  • RESP Compatibility is Key: By explicitly aligning GB parameters with the RESP charge assignment protocol, the model becomes directly applicable to molecular simulations using popular force fields like AMBER, bridging a gap between charge generation and solvation modeling.
  • Atom Typing is Crucial but Challenging: The development of a comprehensive atom typing scheme (21 types, including four for ions) is vital for accuracy. However, the authors acknowledge the difficulty in creating an exhaustive set that perfectly captures all chemical environments, especially for complex organic molecules.
  • GB/SA Offers a Speed Advantage: While achieving comparable or better accuracy than PB/SA for proteins, the GB/SA model is significantly faster (3 seconds vs. 33 seconds for a test set of 36 molecules on a PIV1.4GHz processor), making it more practical for large-scale simulations.

Key Results

  • Organic Molecules:
    • Training Set: Mean unsigned error of 0.85 kcal/mol for 328 neutral molecules and 4.71 kcal/mol for 30 charged molecules.
  • Test Set: Mean unsigned error of 0.73 kcal/mol for 36 neutral molecules and 3.65 kcal/mol for 8 charged molecules.
  • The new GB/SA model significantly outperformed Jayaram's GB/SA and PB/SA models for neutral organic molecules in the test set.
  • Proteins:
    • The new GB/SA model showed good agreement with PB/SA predictions and was significantly better than Jayaram's GB/SA model for the 15 tested proteins.
  • Correlation coefficient (r) of 0.97 with PB/SA for proteins.
  • Speed: The GB/SA model is considerably faster than the PB/SA model.

Limitations and Future Work

  • Atom Typing Limitations: The authors acknowledge that their atom typing scheme, while improved, may not be optimal and could be further refined. They note that the occurrence of certain charged atom types was low, potentially affecting parameter reliability.
  • Overfitting Concerns: The complexity of organic molecules makes it challenging to define enough atom types to differentiate all chemical environments without risking overfitting, especially given the limitations of available experimental data.
  • Future Directions:
    • Developing more elaborate atom typing rules for improved accuracy.
  • Exploring balanced training sets that better represent diverse chemical functionalities.
  • Applying the model to protein folding, side-chain installation, and protein-ligand binding free energy calculations.
  • Incorporating the GB/SA model into docking programs to improve scoring functions.
  • The authors are also providing a program (pdb_typing) to help AMBER users implement these new parameters in their MD simulations.

Summary updated

Abstract

Here we report a set of new parameters for the generalized Born (GB) model consistent with the RESP atomic partial charge assignment protocol. Effective atomic radii and screen factors as parameters have been obtained through genetic algorithm optimization in the parameter space to minimize the differences between the calculated and experimental solvation free energies. Here, the calculated solvation free energies are based on a GB model using partial charges fitted from the electrostatic potentials based on the 6-31G* basis set with the nonelectrostatic contributions to the free energy of solvation modeled in terms of the solvent accessible surface area (SASA). The mean unsigned error in the solvation free energies calculated by the GB/surface area calculations using the final parameters of the 328 neutral molecules in the training set is 0.85 kcal/mol, and for the 30 charged molecules the value is 4.36 kcal/mol. The refined parameters were then applied to predict the solvation free energies of 44 neutral or charged organic molecules and 15 proteins, and reliable results were obtained for both organic molecules and proteins. For the 36 neutral organic molecules in the test set, our parameters incurred an unsigned mean error of 0.73 kcal/mol, and for the eight charged molecules in the test set, our parameters incurred an unsigned mean error of 3.65 kcal/mol. For the 44 organic molecules, the performance of the GB/SA model based on our new parameters was much better than Possion-Boltzmann (PB)/SA and GB/SA based on Jayaram's parameters. For the 15 proteins randomly selected from the Protein Data Bank, the calculated results from GB/SA based on our new parameters also gave consistent results with those from PB/SA and were much better than GB/SA based on Jayaram's parameters. This model might be widely applied in molecules dynamics, protein folding, molecular docking, free energy calculations, and conformation analysis. Moreover, we are now supplying a program to help AMBER users apply our new parameters to their MD simulations.
BibTeX
@article{zhang2003parameters,
  author = {Wei Zhang and Tingjun Hou and Xuebin Qiao and Xiaojie Xu},
  title = {{Parameters for the Generalized Born Model Consistent with RESP Atomic Partial Charge Assignment Protocol}},
  journal = {The Journal of Physical Chemistry B},
  year = {2003},
  doi = {10.1021/jp034613k},
}
Figure