BAT2: an Open-Source Tool for Flexible, Automated, and Low Cost Absolute Binding Free Energy Calculations

Germano Heinzelmann, David J. Huggins, Michael K. Gilson

TL;DR

This paper introduces BAT2, an updated open-source software package that automates and speeds up absolute binding free energy (ABFE) calculations for protein-ligand interactions. By incorporating new features like flexible restraints and support for the OpenMM engine, BAT2 significantly reduces computational cost and time, making it a more practical tool for early-stage drug discovery and virtual screening.

What Problem Does This Paper Address?

Identifying molecules that bind to therapeutic targets is crucial in drug discovery. While experimental high-throughput screening is effective, it's costly. Virtual screening (VS) methods offer a computational alternative, but traditional VS techniques like docking can be less accurate, especially for novel targets. Physics-based methods, particularly all-atom molecular dynamics (MD) simulations for binding free energy calculations, hold great promise for accurate predictions. However, ABFE calculations have historically been hampered by high computational costs, the complexity of setting up simulations, and the need for accurate ligand conformations. This paper addresses these challenges by presenting BAT2, a tool designed to make ABFE calculations more automated, efficient, and accessible.

Methods

BAT2 builds upon its predecessor, BAT1, with several key improvements:

  • OpenMM Support: Integration with the open-source OpenMM simulation engine, alongside the existing AMBER engine, offers users more flexibility and access to GPU acceleration.
  • Flexible Restraints: Instead of relying on fixed dummy atoms, BAT2 now uses relative restraints between the protein and ligand. This simplifies the setup for new receptors and is more robust for larger systems.
  • Merged Restraint Schemes: The attachment and release of restraints (both protein and ligand) can now be performed in fewer simulation windows, significantly reducing computational time and cost.
  • Support for Complex Systems: BAT2 can now handle proteins with multiple chains and the presence of cobinders.
  • New Water Models: Support for TIP3PF⁺ and OPC water models is added, alongside existing ones.
  • Automated Ionization: The software automatically determines the number of ions needed for a chosen ion concentration.
  • Improved Alignment: The use of the lovoalign software for protein structure alignment replaces MUSTANG, enabling better handling of multi-chain proteins.
  • Simplified Workflow: The workflow is streamlined into just two stages: equilibration and free energy calculations, eliminating a separate preparation step.

The paper details the theoretical underpinnings of ABFE calculations, including the use of the double decoupling (DDM) and simultaneous decoupling and recoupling (SDR) methods. It also explains how BAT2 implements these methods, including the calculation of various free energy components (e.g., electrostatic, Lennard-Jones, restraint attachment/release).

Datasets

The paper evaluates BAT2 using two protein-ligand systems:

  1. BRD4(2) bromodomain: This system, previously used to validate BAT1, was employed to test new restraint schemes, OpenMM support, and the impact of reduced simulation times. Calculations were performed on five docked poses and the original crystal structure.
  2. Human HIV-1 protease (5ivq): This system was used to demonstrate BAT2's capabilities with multi-chain proteins and cobinders. Calculations were performed on five self-docked poses and the original crystal structure.

The paper also references experimental data for the HIV-1 protease ligand to compare calculated binding affinities.

Key Insights and Ideas

  • Significant Speedup: The merged restraint scheme in BAT2 dramatically reduces simulation time, achieving a more than 10-fold reduction compared to BAT1 for standard calculations and over 50-fold reduction for very short simulations. This makes ABFE calculations more feasible for large-scale screening.
  • Accuracy Maintained with Speed: Despite the substantial reduction in simulation time, BAT2's results remain consistent with longer calculations and previous versions, demonstrating that efficiency gains do not necessarily come at the cost of accuracy.
  • Flexibility Enhances Applicability: The introduction of relative restraints and support for multi-chain proteins and cobinders makes BAT2 applicable to a much wider range of biological systems.
  • OpenMM Integration is Key: The support for OpenMM, a high-performance, open-source engine, further democratizes access to advanced simulation capabilities.
  • Short Simulations Can Be Reliable: The study shows that even very short simulations (around 20 ns) can yield reliable binding free energy estimates, capable of distinguishing correct binding poses.

Key Results

  • Computational Cost Reduction: The merged restraint scheme reduced simulation time by over 10-fold compared to BAT1's split scheme. Very short simulations (around 20 ns) achieved a >50-fold reduction in computational time compared to previous BAT1 calculations.
  • Agreement Across Methods: Good agreement was observed between AMBER and OpenMM simulations, as well as between the split and merged restraint procedures. Results from BAT2 were generally consistent with BAT1.
  • Accurate Pose Identification: Short simulations were able to correctly identify the most favorable binding poses, correlating well with experimental data for the HIV-1 protease system.
  • HIV-1 Protease Affinity: Calculated binding free energies for the HIV-1 protease ligand were found to be within the range of experimental data (-7 to -8 kcal/mol).

Limitations and Future Work

  • Uncertainty in Short Simulations: While short simulations show good accuracy, there was a slight increase in uncertainties compared to longer simulations. The authors note that the results still generally fall within the computed uncertainties.
  • Ligand Flexibility in Short Simulations: For very short simulations (tevb procedure), the authors suggest that anchor atoms in flexible ligand regions might lead to an overestimation of restraint attachment free energy. The m*evbc approach, which keeps the ligand rigid, is recommended in such cases.
  • Future Work: The paper implies that further optimization and exploration of reduced simulation times will continue, making ABFE calculations even more accessible for large-scale screening. The open-source nature of BAT2 also encourages community contributions and further development.

Summary updated

Abstract

Absolute binding free energy (ABFE) calculations with all-atom molecular dynamics (MD) have the potential to greatly reduce costs in the first stages of drug discovery. Here, we introduce BAT2, the new version of the Binding Affinity Tool (BAT.py), designed to combine full automation of ABFE calculations with high-performance MD simulations, making it a potential tool for virtual screening. We describe and test several changes and new features that were incorporated into the code, such as relative restraints between the protein and the ligand instead of using fixed dummy atoms, support for the OpenMM simulation engine, a merged approach to the application/release of restraints, support for cobinders and proteins with multiple chains, and many others. We also reduced the simulation times for each ABFE calculation, assessing the effect on the expected robustness and accuracy of the calculations.
BibTeX
@article{heinzelmann2024bat,
  author = {Germano Heinzelmann and David J. Huggins and Michael K. Gilson},
  title = {{BAT2: an Open-Source Tool for Flexible, Automated, and Low Cost Absolute Binding Free Energy Calculations}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2024},
  doi = {10.1021/acs.jctc.4c00205},
}
Figure