OpenMM

Simulations are defined once and executed transparently on NVIDIA, AMD, or Apple Silicon GPUs, or multi-core CPUs. OpenMM's runtime compiler selects the optimal compute backend without requiring code changes. This write-once, run-anywhere paradigm ensures that molecular dynamics workflows remain portable across heterogeneous computing clusters and local workstations.

The architecture supports dynamic loading of plugin libraries to introduce custom interaction potentials without modifying the core engine. This enables the integration of:

• Neural network potentials (ANI, DeepChem)
• Polarizable force fields (AMOEBA, Drude oscillators)
• Coarse-grained models (Martini)
• Custom restraint potentials for enhanced sampling

OpenMM natively implements Generalized Born and Poisson-Boltzmann implicit solvent models for rapid free energy estimation. The toolkit includes a suite of advanced integrators:

• Langevin Middle Integrator for optimal temperature control
• Variable time-step integrators for mixed-resolution systems
• Custom thermostat and barostat algorithms These are exposed through a Python API that maintains C++ performance.

OpenMM functions as the high-performance compute backend for major simulation frameworks including Amber, CHARMM, GROMACS, and DeepChem. It reads standard file formats (PDB, Amber prmtop/inpcrd, CHARMM PSF) and exports trajectories in DCD and XTC formats. This interoperability allows research groups to accelerate existing workflows without abandoning established toolchains.

The user-facing layer is a pure Python API that provides intuitive access to system construction, simulation control, and analysis. The performance-critical engine is written in C++ and compiled to native code. This design enables rapid prototyping in Jupyter notebooks while delivering simulation throughput comparable to hand-tuned Fortran or C codes running on thousands of GPU cores.

A class of molecular dynamics techniques that apply external biases to accelerate exploration of a system's free energy landscape. These methods enable observation of rare events—such as protein folding or ligand unbinding—within computationally feasible timescales. Key approaches include:

• Metadynamics: Deposits history-dependent Gaussian bias potentials
• Umbrella Sampling: Uses harmonic restraints along reaction coordinates
• Replica Exchange MD: Runs parallel simulations at different temperatures
• Gaussian Accelerated MD: Smoothens the potential energy surface without predefined collective variables

Machine-learned interatomic potentials that regress the potential energy surface from high-level quantum mechanical data. These models provide ab initio accuracy at a fraction of the computational cost of traditional quantum chemistry methods. Notable frameworks include:

• Deep Potential Molecular Dynamics (DeePMD-kit): Learns local atomic environment descriptors from first-principles data
• ANI: Utilizes symmetry functions as atomic environment descriptors
• SchNet: Employs continuous-filter convolutional layers for molecular representations
• NequIP: Implements equivariant message passing for state-of-the-art accuracy

Computational techniques that calculate the free energy difference between two states by simulating a non-physical pathway of intermediate states. One molecule is gradually morphed into another through coupling parameters. Essential methods include:

• Thermodynamic Integration (TI): Integrates the derivative of the Hamiltonian along the alchemical path
• Free Energy Perturbation (FEP): Uses exponential averaging of energy differences
• Bennett Acceptance Ratio (BAR): Provides statistically optimal estimates from bidirectional work distributions
• Multistate Bennett Acceptance Ratio (MBAR): Combines data from all intermediate states simultaneously to minimize variance

Kinetic network models that discretize phase space into metastable states and estimate a transition probability matrix to describe long-timescale dynamics from many short, parallel simulations. The workflow involves:

• Featurization: Reducing trajectory frames to informative collective variables
• Dimensionality reduction: Applying Time-lagged Independent Component Analysis (TICA) to identify slow degrees of freedom
• Clustering: Grouping conformations into discrete microstates
• Transition matrix estimation: Computing state-to-state transition probabilities at a chosen lag time
• Macrostate decomposition: Using Perron-Cluster Cluster Analysis (PCCA+) to identify kinetically distinct macrostates

A simulation approach that reduces computational cost by grouping atoms into pseudo-particles or beads, enabling the study of larger systems and longer timescales at the expense of atomic resolution. The Martini force field is the most widely used coarse-grained model, mapping approximately four heavy atoms to a single interaction site and parameterized to reproduce thermodynamic properties like partitioning free energies. Applications include:

• Lipid bilayer self-assembly and phase behavior
• Protein-protein association kinetics
• Large-scale membrane remodeling processes
• Viral capsid assembly simulations