Abstract

📷

AmberTools is a free and open-source collection of programs used to set up, run, and analyze molecular simulations. The newer features contained within AmberTools23 are briefly described in this Application note.

This publication is licensed under

CC-BY-NC-ND 4.0.

• 📷
• 📷
• 📷
• 📷

Introduction

ARTICLE SECTIONS

Jump To

The present status of the Amber (Assisted Model Building and Energy Refinement) suite of programs has been the product of decades of effort from a broad range of research groups, starting with the group of the late Peter Kollman in the early 1980s. (1) Amber contains tools for energy minimization (EM), molecular dynamics (MD) simulations, free energy (FE) calculations, potential of mean force (PMF) capabilities, and all the needed tools to set up the modeling effort. The software stack has been reviewed in the past, (2−4) and the manual contains detailed descriptions of all the algorithms in Amber as well as a full list of contributors to Amber over the years (see https://ambermd.org). Besides the actual code, Amber is used to describe a series of highly regarded force fields (5) for proteins, (6−12) carbohydrates, (13,14) nucleic acids, (7,8,15) and lipids. (16) The present Application Note will only describe the latest additions to the open-source AmberTools23 and as such is not meant to give a thorough exposition of all the methods and capabilities of AmberTools and Amber.

Overview of Amber and AmberTools

Amber and AmberTools form a collection of programs that are designed to work together to facilitate system setup, MD simulations, and trajectory analysis for biomolecules. It is useful to note that the Amber force fields mentioned above can be used in a variety of molecular dynamics codes outside of AmberTools and Amber. The Amber code is updated in even-numbered years, and it uniquely includes the base MD code known as pmemd, which offers parallel and graphics processing unit (GPU)-accelerated versions of the MD codes along with some free-energy-based methods not implemented in AmberTools. Analogous MD function is available in sander in AmberTools. AmberTools is distributed under an open-source license, primarily the GNU General Public License, with some portions covered by other compatible open-source licenses. The Amber force fields are in the public domain and are distributed with AmberTools. The pmemd code is distributed as source code but has a separate license that contains restrictions on use and redistribution; there is no license fee for noncommerical use of pmemd. Full details on licensing and distribution can be found at https://ambermd.org.

Typical Workflow

The basic workflow for AmberTools is shown in the accompanying (see Figure 1), and it describes preparation, simulation, and analysis steps. Preparation starts at the top, since all MD simulations require some sort of starting three-dimensional (3D) structure, which for biomolecules is usually in the form of a PDB-format file; AmberTools has some model-building capabilities (e.g., PACKMOL-Memgen, see below), but other codes are generally used if experimental structures are not available. The prepareforleap step, which is recent and still under development, carries out tasks to map components in the input file to Amber nomenclature (especially useful for carbohydrates), add hydrogens, identify cross-links, assign histidine protonation states, and similar tasks. Next is the LEaP program, which is a workhorse program that connects the nascent structure to Amber’s built-in force fields for proteins, nucleic acids, carbohydrates, lipids, and common solvents and to bespoke force fields for other components like ligands and cofactors that can be created by programs like antechamber and mdgx (for general organic molecules) and pyMSMT (for metal ions). The LEaP code creates two files: an “inpcrd” file that has complete three-dimensional coordinates and a “prmtop” file that contains all other information needed for force field-based analyses of the system. The latter file can be examined and edited via parmed, which can also export similar files in the GROMACS or CHARMM format.

Figure 1 📷Figure 1. Common workflow in AmberTools. Flow went from top to bottom. Black boxes are for preparation, gray indicates an optional preparation step specific for membrane systems, blue for simulation, and red for analysis.

The simulation phase is primarily the province of sander or pmemd. The “mdin” file contains a large number of parameters that control the type and length of the simulation to be carried out, the integration method, the use of a QM/MM (quantum mechanics/molecular mechanics) model, specification of enhanced-sampling and thermodynamic integration methods, and the like. Restraints on the system, often from NMR or X-ray data but more recently from cryogenic electron microscopy (cryoEM) and other sorts of integrative modeling, can also be input at this point.

Snapshots of conformations are generally stored at regular intervals during a simulation and then serve as input for an analysis phase. The cpptraj program is the workhorse code here, providing geometric and energetic analyses, clustering algorithms, and many other routines. Three other codes, MMPBSA.py, FE_Toolkit, and FEW (Free Energy Workflow) (17) are devoted to estimating free energy changes. More complete descriptions of all of this, including a full list of programs, encompassing nearly 1000 pages of text, are in the Amber23 Reference Manual.

AmberTools23 Updates

We have a number of significant new features for AmberTools23 which include automated building of membrane-protein–lipid-bilayer systems, enhancements to the polarizable Gaussian multipole method, extensions to the Poisson-Boltzmann surface area (PBSA) method, enhanced free energy capabilities, enhanced QM and QM/MM capabilities, and a significant upgrade of the Amber Web site and tutorials. Each of these additions is summarized below.

1. Polarizable Gaussian Multipole Model in the SANDER Program

The polarizable Gaussian Multipole (pGM) model is a next-generation induced-dipole polarizable model aiming to balance accuracy and efficiency for molecular simulations of biomolecular systems. (18−22) We recently developed a new framework for efficient computation of analytical atomic gradient for the pGM model. (18) The pGM virial for constant pressure molecular dynamics simulations was also implemented in previous releases of Amber. (19) The accuracy and robustness of the pGM model have also been validated on various molecular properties. (20−22) In the AmberTools23 release, we further optimized the induced-dipole iteration algorithm. Specifically, we introduced maximum relative error as the convergence criterion to ensure energy conservation in molecular dynamics simulations. We also designed and implemented multiorder extrapolation (MOE) and local preconditioning conjugate gradient (LPCG) schemes to accelerate the induced-dipole iteration. (23) Given the new developments, MD simulations with the pGM model are able to achieve a similar level of energy conservation as those with the point charge additive models, within 2–3 induction iterations.

2. New Features in the PBSA Program

MM/PB(GB)SA (24) is an end-point method for calculating the free energies of molecules in implicit solvent, i.e., Poisson–Boltzmann (PB) and generalized Born (GB). Solvation interactions, especially solvent-mediated dielectric screening and Debye–Hückel screening, are essential determinants of the structure and function of biomolecules. Several efficient finite-difference numerical solvers, both linear (25−27) and nonlinear, (28) are implemented in pbsa for various applications of the Poisson–Boltzmann method. The GPU support of those solvers is also implemented in pbsa.cuda. (29−31) In the 2023 release, improvements to the pbsa program include the integration of the Machine-Learned Solvent Excluded Surface (MLSES) model, (32) which provides a highly efficient and differentiable molecular surface for continuum solvation modeling of biomolecules. Various options for the MLSES model have been implemented, allowing users to optimize performance on both central processing unit (CPU) and GPU platforms using Fortran, the CUDA kernel, and LibTorch. This flexibility enables users to choose the best-suited hardware and software environments for their needs. Additionally, an MBAR/PBSA strategy has been developed combining the PBSA continuum solvent model with the Multistate Bennet Acceptance Ratio (MBAR) approach. This coupling allows for more accurate modeling of electronic polarization, leading to improved accuracy in absolute binding free energy simulations of highly charged ligands. (33)

To date, the GB model in AmberTools could be specified with the following “igb” values: 1, (34) 2, (35) 5, (36) 7, and 8. (37) In 2017, an accurate yet efficient grid-based surface GB model was introduced (38) which is currently available in AmberTools as a stand-alone application named GBNSR6 ($AMBERHOME/bin/gbnsr6). (39) GBNSR6 calculates the solvation free energy of an input structure on a single snapshot. In AmberTools23, GBNSR6 has been integrated into MMPBSA.py (40) such that it runs over multiple snapshots extracted from the trajectories of protein, ligand, and complex structures. To run this model, “igb = 66” is now available in MMPBSA.py. All input parameters of the stand-alone GBNSR6 program can be modified through the MMPBSA.py input file.

3. PyRESP and PyRESP_GEN

Accurate modeling of electrostatic and polarization effects is crucial in molecular simulations. Many polarizable force fields have been developed to account for these important effects. Among these models, the polarizable Gaussian Multipole (pGM) model has emerged as a self-consistent approach in handling both short-range and long-range interactions. (18−23) We have recently developed the PyRESP program (41) for electrostatic parametrizations for point charge additive models and induced-dipole models, including the pGM model. By performing least-squares fittings to electrostatic potentials surrounding molecules, the PyRESP program extends functionalities of the ancestor RESP program that only perform parametrizations for point charge additive models. (42) However, the process of generating input files for PyRESP is tedious and error-prone. In the AmberTools23 release, we implemented a flexible and user-friendly program, PyRESP_GEN, (82) to minimize the user’s efforts to set up a PyRESP run. In addition, we also optimized the restraint weights for the pGM models with and without permanent dipoles. For the pGM-perm model, the optimal strategy for electrostatic potential fitting is also proposed.

4. 3D-RISM

The 3D reference interaction site model (3D-RISM) of molecular solvation (43) is an implicit solvent model that calculates equilibrium density distributions and thermodynamics of explicit solvent models. The implementation in AmberTools permits MD, energy minimization, and trajectory analysis through sander, while rism3d.snglpnt provides standalone trajectory analysis. (44) Recently, periodic boundaries were introduced, allowing application to crystal structure refinement and other periodic systems. (45) In addition, computational scaling for open boundaries was improved via treecode summation for electrostatic interactions, providing a 2–4 times speedup for typical proteins and enabling application to large biomolecular complexes with more than 1 million atoms. (46)

5. LibTorch Interface to Amber

We introduced a LibTorch interface to the 2023 release of AmberTools, which is a cutting-edge C++ runtime library developed by the PyTorch team. (47) This library enables flexible tensor computations and dynamic deep neural network modeling. Amber now provides two options for enabling the LibTorch library: a built-in mode and a user-installed mode. With the LibTorch integration, the pbsa program supports both CPU and GPU environments, making it highly versatile. Additionally, user instructions and tutorials have been provided for configuring the LibTorch library, making it more accessible to researchers and developers working in Amber and AmberTools.

6. Free Energy

Free energy methods have been a mainstay of Amber for decades. (48,49) Besides our existing free energy technology base this latest release of AmberTools includes a collection of new software tools for the robust analysis of free energy simulations (FE-ToolKit) (50,51) as well as workflow tools for production free energy simulation setup and analysis (ProFESSA) (52) using the GPU-accelerated Amber free energy engine with enhanced sampling features. This software is part of the Amber Drug Discovery Boost package. (53)

6.1. FE-ToolKit

The FE-ToolKit contains two main utilities: edgembar (51) for analysis of alchemical free energy simulations (e.g., such as those used for prediction of ligand-protein absolute and relative binding free energies in drug discovery (54)) and ndfes (50) for analysis of multidimensional free energy profiles (e.g., such as those used for prediction of minimum free energy pathways in studies of enzyme mechanisms (55,56)).

6.2. Edgembar

The edgembar program performs analysis of alchemical free energy simulations using the multistate Bennett acceptance ratio (MBAR) method, (57) the Bennett acceptance ratio (BAR) method, (58) exponential averaging, (59) thermodynamic integration, (60) or combinations of these approaches. Alchemical free energy simulations often calculate a network of relative free energy differences between two environments. For example, ligand binding applications in drug discovery use a network of alchemical transformations between ligands, termed a “thermodynamic graph”, where each ligand represents a “node” in the graph and each “edge” represents an alchemical transformation between ligands bound to their target relative to that in aqueous solution. Given the alchemical simulation outputs from the independent trials in both environments, edgembar will perform a “network-wide” free energy analysis, (51) including the imposition of cycle closure and, optionally, experimental constraints. The analysis produces a comprehensive report of the results, including uncertainties and warnings. The report identifies potential problems with simulations that may require further attention. The issues include: a lack of convergence, the analysis of too few statistically independent samples, poor phase space overlap between adjacent alchemical states, (61) and poor reweighting entropy. (62)

6.3. Ndfes

The ndfes program evaluates multidimensional free energy surfaces from umbrella sampling. (50) The analysis can be performed with the variational free energy profile (vFEP) method, (63,64) MBAR, (57) the weighted thermodynamic perturbation method (wTP), (65) and the generalized weighted thermodynamic perturbation method (gwTP). (66) The wTP and gwTP methods estimate the free energy surface of an expensive target-level of theory from the sampling performed with inexpensive reference potentials. (66) The estimation of ab initio QM/MM free energy surfaces in condensed-phase environments has become more practical in the latest version of AmberTools with the combined introduction of the GPU-accelerated QUICK software (67) and ndfes analysis program.

6.4. ProFESSA

The ProFESSA workflow (52) uses the GPU-accelerated AMBER free energy engine. The workflow establishes a flexible, end-to-end pipeline for performing alchemical free energy simulations that brings to bear technologies including new smoothstep softcore potentials and optimized alchemical transformation pathways, (68) the alchemical enhanced sampling (ACES) method, (69) and a network-wide free energy analysis (51) with optional imposition of cycle closure and experimental constraints implemented in FE-ToolKit.

7. Quantum Mechanical/Molecular Mechanical Methods

Amber has had a long tradition of QM/MM methods and implementations, (70) with the most recent additions being the QUICK/sander QM/MM implementation in AmberTools23. (67,71−73) QUICK/sander has been extensively updated, and its performance has been significantly improved. QUICK, as distributed with AmberTools23, can also be used as a standalone QM program for single point calculations or geometry optimizations.

7.1. Performance Improvements/AMD Implementation

With the second-generation electron repulsion integral code and other performance enhancements recently introduced into QUICK, (67,71−73) higher ps/day can be obtained in QM/MM simulations. (73) For instance, with respect to AmberTools21, (74) up to 2× speedups have been observed for benchmark simulations with different QM regions of photoactive yellow protein on NVIDIA V100 GPU. (73) Furthermore, support for AMD GPUs has been enabled. Users can now make use of AMD data center cards such as MI50, MI100, MI200, and MI250 for simulations. According to benchmark studies, the performance on the MI100 is similar to that of NVIDIA V100. (73) The implementation runs properly on MI200 and MI250 cards; however, the performance is not yet optimized for these cards. The recommended AMD GPU for the current version is MI100. An optimized version for MI2XX will be available to users in the next AmberTools release.

7.2. Long-Range Electrostatics

For the treatment of long-range electrostatics in QM/MM, the ambient-potential composite Ewald method (CEw) developed by Giese and York (75) has been integrated. The performance penalty for turning on CEw in the GPU version is