A protein structure from the Protein Data Bank is a still photograph. It is averaged over millions of molecules and, in the case of X-ray crystallography, taken under conditions a cell never experiences. Molecular dynamics simulation turns that photograph into a film: it computes how every atom moves over time, letting you watch the structure breathe, flex, bind and sometimes fall apart.
This article explains what MD actually does, what it is genuinely good for, where it misleads people, and what you need to run one sensibly.
What a molecular dynamics simulation actually computes
The principle is simpler than most people expect. For every atom in the system, the simulation:
- Calculates the force acting on it from every other atom, using a force field.
- Applies Newton’s second law to convert force into acceleration.
- Advances every atom forward by a tiny time step — typically 2 femtoseconds.
- Repeats, hundreds of millions of times.
The output is a trajectory: coordinates for every atom at every saved time point. A 100-nanosecond simulation with a 2 fs time step is 50 million of these calculations, which is why MD is computationally demanding and why GPUs transformed the field.
The force field is the model — and its main limitation
A force field is a set of parameters describing bond lengths, angles, torsions, electrostatics and van der Waals interactions. Common protein force fields include AMBER, CHARMM, OPLS and GROMOS.
This matters more than newcomers realise: your results are only as good as your force field. These are empirical approximations fitted to experimental and quantum-mechanical data. They do not model bond breaking or formation, so standard MD cannot simulate a chemical reaction. They also carry known biases — some force fields historically over-stabilised helices, others made disordered regions too compact. Choosing one is a scientific decision, not a default.
Ligands are a particular pain point. Proteins are well parameterised; an arbitrary drug-like molecule is not, and needs parameters generated (via GAFF, CGenFF or similar) and ideally checked before you trust anything downstream.
What MD is genuinely good for
Testing whether a docked pose is real
This is the highest-value routine use. Molecular docking produces a ranked list of static poses scored by an approximate function. It is fast and useful, but scoring functions are known to be unreliable at ranking. Running MD on your top poses tells you which ones survive in explicit solvent at physiological temperature. A pose that drifts out of the pocket in the first nanosecond was never a binding mode, whatever its docking score said.
Understanding the effect of a mutation
If you have a variant of unknown significance, MD lets you compare wild type and mutant directly: does the mutation rigidify a loop, disrupt a salt bridge, shift the position of a catalytic residue, destabilise the fold? This is often more informative than a stability prediction score, because you see the mechanism rather than a number.
Finding cryptic and allosteric pockets
Some binding sites do not exist in the crystal structure — they open transiently as the protein moves. MD is one of the few practical ways to find them, and cryptic pockets have produced real drug programmes.
Membrane proteins
Roughly a quarter of the proteome sits in membranes, and their behaviour depends on the lipid environment. MD in an explicit bilayer is the standard computational approach for transporters, channels and GPCRs.
Binding free energy
Methods such as MM/PBSA, MM/GBSA and, more rigorously, free energy perturbation and thermodynamic integration estimate binding affinity from simulation. The rigorous methods can approach experimental accuracy for congeneric series, which is why they are now standard in lead optimisation at large pharmaceutical companies. MM/GBSA is far cheaper and correspondingly less reliable — useful for ranking within a series, not for absolute affinity.
Where MD misleads people
Three failure modes account for most bad MD in the literature.
Sampling. A simulation only shows you what happened during that run. If a conformational change takes a millisecond and you simulated 100 nanoseconds, you will not see it — and its absence is not evidence it does not occur. Timescale mismatch is the single largest limitation of the method.
Single trajectories. One simulation is one sample from a stochastic process. Rerun it with different starting velocities and you may get a different answer. Conclusions from a single replica are not trustworthy; three or more independent runs should agree before you believe a result.
Over-interpreting motion. Proteins are inherently dynamic. Loops move. Termini flail. Seeing motion is not a finding. The finding is a difference in motion between conditions you are comparing.
What you need to run one
- A structure. Experimental where possible; a predicted model from AlphaFold or similar is workable, but check the confidence scores, since low-confidence regions are frequently disordered and will dominate your flexibility analysis.
- A prepared system. Correct protonation states at your target pH, missing loops modelled, the structure solvated in a water box with counter-ions at physiological salt concentration.
- An engine. GROMACS, AMBER, NAMD and OpenMM are the common choices. GROMACS is free and fast; AMBER has excellent free energy tooling; NAMD scales well; OpenMM is the most flexible if you are scripting in Python.
- GPU compute. A single modern GPU can produce roughly 100–300 ns per day for a typical 50,000-atom system. CPU-only MD is impractical for most research questions today.
The standard workflow
- Prepare the structure: fix missing atoms and loops, set protonation, add the ligand with validated parameters.
- Solvate in a periodic water box and neutralise with ions.
- Minimise to remove steric clashes.
- Equilibrate under NVT then NPT, usually with restraints on the protein that are gradually released.
- Run production without restraints, for as long as the question requires.
- Analyse the trajectory.
Steps 1 and 6 are where projects are won or lost. The middle steps are largely mechanical and well documented; system preparation errors quietly invalidate everything downstream, and weak analysis wastes a good simulation.
Does MD replace experiments?
No, and claiming otherwise is how computational work loses credibility. MD generates mechanistic hypotheses that are hard to obtain any other way, and it explains experimental observations that a static structure cannot. But a force field is a model, and a simulation is not evidence of what a molecule does in a cell.
The productive framing is that MD tells you which experiment to run. It identifies the residue worth mutating, the pose worth pursuing, the pocket worth screening against. Used that way it saves substantial bench time. Used as a substitute for evidence, it produces confident conclusions that do not replicate.
