Choosing an operation

Start from what you have in hand. Runtime is indicative and depends on system size.

The single largest practical difference in this table is cost. A physics-based trajectory in GROMACS or OpenMM is hours to days of compute; a learned ensemble from BioEmu is seconds. That gap, rather than any accuracy argument, usually decides which route a project takes. Use the learned models to explore and the physics-based engines to confirm.

[Small molecules are out of scope for this page. If you are starting from a SMILES string with no 3D coordinates, generate conformers first → see the ligand preparation page.]

The operations

Molecular dynamics simulation

Use case replayMolecular dynamics simulationSee this use case

Predicts time-resolved atomic coordinates of a protein or protein-ligand complex under a given force field, by integrating Newton's equations of motion.

This is the reference method. It produces nanosecond to microsecond trajectories at atomic resolution, capturing conformational dynamics, solvation effects and transient states that no single structure contains. Use it when you need quantitative dynamics, thermodynamic quantities, or validation of a docked pose. It is the slowest option here by a wide margin.

Input

Protein structure (PDB or CIF), force field (AMBER, CHARMM), water model, simulation parameters including temperature, pressure, duration and timestep.

Output

Trajectory frames (XTC, TRR, DCD), energy logs, final structure, and standard metrics (RMSD, RMSF, radius of gyration).

ModelsGROMACSOpenMMAmbertools-OpenMM-MD

Trajectory analysis

Computes structural and dynamical metrics from a trajectory you already have, without re-running the simulation.

Use this to extract stability (RMSD), per-residue motion (RMSF), interaction networks, secondary structure evolution or principal components from an existing run. It applies to a single continuous trajectory; for comparing discrete snapshots, use ensemble analysis instead.

Input: Topology file (PDB, PSF, GRO, TPR, PRMTOP), trajectory file (XTC, TRR, DCD, NetCDF), atom selection, and the analyses required.

Output: Time-series plots, per-residue B-factors, secondary structure evolution, contact frequency matrices, principal components with explained variance, and an interactive 3D viewer.

Models: MDAnalysis, MDTraj.

Trajectory generation

Predicts a sequence of future conformations from an initial structure using a learned generative model, without solving any equations of motion.

This is for rapid exploration, for filling gaps in a sparse trajectory, or for transition-path inference. It returns results in seconds rather than hours. It is not a substitute for physics-based MD where the numbers matter, for reasons set out under the model description below.

Input: Initial protein structure (PDB, 5 to 256 residues), frames per rollout, number of autoregressive rollouts, and simulation mode.

Output: Generated trajectory as multi-model PDB or XTC, with a per-frame confidence or energy proxy.

Models: MDGen.

Structure ensemble analysis

Computes metrics and alignments across a set of conformations to identify metastable states, conformational diversity and per-residue fluctuation.

Use this to compare multiple experimental structures, to characterise a generated ensemble, or to summarise MD snapshots without treating them as a time series. It works on discrete structures rather than continuous dynamics.

Input: Ensemble archive of PDB files, optional reference structures for two conformational states, clustering method, and CPU cores for parallel alignment.

Output: Ensemble-wide RMSD and clustering dendrogram, per-residue B-factors and RMSF, state classification, PCA projection, and an interactive alignment viewer.

Models: AFsample2 in analyse mode.

Protein ensemble generation

Produces a diverse set of conformations sampling the protein's accessible states, from sequence or from an initial structure.

This replaces a single-point prediction with a distribution, which is what ensemble-based docking and flexibility-aware design require. Use it when you need multiple binding poses, allosteric states or a population estimate. The four models available differ substantially in cost and in what their output represents, so read the descriptions before choosing.

Input: Protein sequence, or an initial structure for RAVE. Number of seeds, MSA parameters, simulation settings, and an optional reference MSA.

Output: Multi-model PDB or ensemble archive of 10 to 128 structures, per-model confidence, optional per-residue RMSF and state assignment.

Models: AlphaFold2-RAVE, BioEmu, AFsample2, AlphaFlow.

Model reference

Physics-based engines

• GROMACS is a GPU-accelerated engine optimised for large biomolecular systems. It supports replica-exchange MD for enhanced sampling and computes RMSD, RMSF and Rg during the run. Choose it for throughput on large systems, or when you already have a GROMACS TPR file.

• OpenMM is a flexible toolkit with a Python API over a C++ core, supporting AMBER, CHARMM and AMOEBA force fields with both explicit and implicit solvent. Choose it when you need custom constraints, non-standard integrators or implicit solvent without full parameterisation. It is more approachable than GROMACS for bespoke workflows.

• Ambertools-OpenMM-MD bundles AmberTools parameterisation, including H++ protonation and GAFF ligand handling, with the OpenMM engine. Choose it for protein-ligand systems where you want automatic preparation and an end-point binding free energy estimate.

Trajectory analysis libraries

MDAnalysis and MDTraj are both Python libraries for reading and analysing trajectories, and their capabilities overlap heavily. Both compute RMSD, RMSF, hydrogen bonds, secondary structure and principal components, and both run identically in a notebook or a script.

The practical differences are narrower than they are often presented. MDAnalysis has a richer atom selection language and a larger ecosystem of analysis modules, which matters for complex selections and less common analyses. MDTraj has faster file I/O and a lighter API, which matters when processing many trajectories in a pipeline. Either will handle a standard RMSD and RMSF workflow, so pick on selection syntax and format support rather than on capability.

Learned trajectory and ensemble models

• MDGen is a learned trajectory predictor. It takes a single structure and outputs subsequent conformations based on patterns learned from MD datasets, supporting forward simulation, transition-path inpainting and upsampling of sparse trajectories.

An important limitation. MDGen does not integrate equations of motion, so it does not conserve energy and its output is not guaranteed to be physically realisable. Timesteps are nominal rather than physical. It is suitable for exploring conformational space cheaply and for filling gaps in an existing trajectory. It is not suitable for anything where kinetics, thermodynamics or solvation-dependent behaviour matter, and it should not be read as a cheap replacement for GROMACS or OpenMM.

• AlphaFold2-RAVE combines reduced-MSA structure prediction with unbiased MD and metadynamics to explore the conformational landscape, using learned collective variables and a well-tempered bias. The output is reweighted toward equilibrium populations, which is a stronger basis for population estimates than the purely generative models, though it should not be taken as a converged free-energy landscape. Full RAVE takes hours; an AF2-only mode runs in minutes with correspondingly weaker guarantees.

• BioEmu is a diffusion model that produces a backbone-level ensemble from sequence in seconds, trained on structural databases together with MD data. Its speed makes it the default first look at whether a protein is flexible and where. Note that it samples an approximation to the equilibrium distribution rather than the distribution itself, and that it returns backbone coordinates rather than all-atom structures.

[Confirm whether the deployed BioEmu path requires an MSA. The source document states both that it wraps ColabFold MSA embedding and that it needs no MSA.]

• AFsample2 runs AlphaFold2 with randomised MSA columns across seeds, using stochasticity in the input alignment to produce a diverse ensemble. It also provides the analysis mode used for ensemble clustering and state classification, which is why it appears both as a generator and as an analyser.

AlphaFlow is AlphaFold2 fine-tuned under a flow matching framework, and it ships as distinct checkpoints that produce different distributions. The PDB checkpoint was trained on experimental structures and gives better precision and diversity than AlphaFold with MSA subsampling. The MD checkpoint was further fine-tuned on all-atom MD ensembles and reproduces conformational flexibility and per-residue fluctuation more faithfully. Choose the checkpoint that matches the distribution you want to sample, since they are not interchangeable.

Next step

Once you have characterised the dynamics or generated an ensemble, the results feed directly into structure-based design. Use the ensemble for ensemble docking, to screen for compounds that bind across several conformations rather than one, then apply ADMET filtering and binding free energy calculations to the survivors. Screening against an ensemble rather than a single structure prioritises compounds that are robust to conformational change, which tends to improve hit rates in follow-up assays.