Molecular dynamics for binder design
Molecular dynamics integrates Newton's equations for every atom in a system, using a force field to supply the forces. Given a starting structure, it produces a trajectory: a movie of the system moving under thermal energy. For binder design, the useful question is narrow and answerable: does this predicted complex hold together, and where does it hold?
The question MD is good at
A designed binder usually arrives as a static pose from a structure prediction or docking method. That pose is a hypothesis, and it is often wrong in ways that are not visible in a single frame. Simulation puts the complex in water at temperature and lets it come apart if it wants to.
The readouts that repay the compute are the ones about persistence rather than energy:
- Does the binder stay on the epitope? Align on the antigen and watch the binder. Drift, rotation or a clean departure is a result.
- Which contacts survive? Contacts present in the starting pose that vanish in the first nanoseconds were packing artifacts. Contacts that persist across replicas are the interface.
- What is flexible? Loops that thrash in every replica are unlikely to be doing the work they were designed to do.
- Do the replicas agree? Agreement is itself the measurement. Three replicas that disagree tell you the pose is marginal, which is more useful than one replica that looks good.
The questions MD is not good at
Plain simulation on accessible timescales does not give you binding affinity. Free energy methods exist and they work, but they are a different undertaking with their own controls and their own failure modes, and they are not what a short unbiased trajectory delivers.
It also does not sample rare events. A binder that would unbind on a millisecond timescale will look perfectly stable for a hundred nanoseconds. Absence of unbinding is weak evidence of tight binding, and it should be reported as what it is.
And it will not rescue a wrong pose. Simulation refines and tests, it does not search. If the epitope is wrong, a longer trajectory gives you a well-equilibrated wrong answer.
Setup choices that decide the answer
- Force field and water model. Modern protein force fields and their matched water models are the baseline. Mixing a force field with a water model it was not parameterized against is a quiet source of nonsense.
- Protonation and termini. Histidine states and capped or charged termini change interfaces. Decide deliberately rather than accepting a default.
- Solvation, ions and box size. Enough water that the complex never sees its own periodic image, and physiological ionic strength unless there is a reason to do otherwise.
- Equilibration. Minimize, heat, and restrain the solute while the solvent relaxes. Production started too early produces artifacts that look like biology.
- Replicas over length. Several independent runs with different starting velocities beat one long run for almost every design question, because they sample the variability that matters.
Analysis traps worth knowing
The one that bites everyone is periodic boundary imaging. Molecules that cross the box edge appear to teleport, and a complex can look dissociated when it is intact. Image and center the trajectory on the antigen before measuring anything, and look at the movie before trusting a metric.
The second is alignment. Root mean square deviation computed after aligning on the whole complex hides binder motion inside a global fit. Align on the antigen, then measure the binder.
The third is reporting a mean without the spread. A single average over replicas that disagree is the least informative number in the analysis.
How we run it
We simulate the shortlist rather than the whole panel, run replicas rather than one long trajectory, image and align on the antigen before analysis, and report contact persistence and the spread across replicas rather than one summary number.