Proteins are not statues
A crystal structure freezes one conformation, but a real protein jitters constantly. Molecular dynamics (MD) simulates that motion: it places the protein, ligand, and explicit water in a box, applies a force field describing all the atomic forces, and steps forward in tiny time increments to watch the system evolve. MD reveals conformational flexibility that a single static picture hides — side chains that swing open, loops that gate a pocket, and the induced fit that rigid docking cannot see.
MD is also the honest way to ask about water. By watching where waters sit and how tightly they are held, you can identify the displaceable waters worth targeting and the structural waters you must keep. And because MD reports on fluctuation, it speaks directly to binding entropy — the often-ignored half of binding that comes from changes in flexibility and disorder when a ligand locks in.
Binding as a free-energy balance
Affinity is governed by the binding free energy, ΔG, which combines enthalpy (the interactions you drew) and entropy (order, flexibility, and water). A scoring function only crudely estimates ΔG; that is why it ranks poorly. To predict potency well enough to guide synthesis, you need a method grounded in real statistical thermodynamics that accounts for protein motion and explicit solvent.
Free-energy perturbation (FEP) is that method. Instead of computing one molecule's absolute ΔG — which is very hard — FEP computes the difference in binding free energy between two close analogs by gradually "morphing" one into the other inside the pocket and in water, using MD all the way. Because both legs share most of their structure, errors cancel, and modern FEP can predict relative potency to roughly 1 kcal/mol — accurate enough to rank which analog to make next.
Why a difference is easier than an absolute: ligand_A --(alchemical morph in WATER)--> ligand_B : dG_water ligand_A --(alchemical morph in POCKET)--> ligand_B : dG_bound ddG_binding(A->B) = dG_bound - dG_water A simple edit -- swap H for F, add a methyl -- changes little, so the shared parts cancel and only the change is 'paid for'.
Using FEP without wasting it
FEP is accurate but costly, and it is only as good as the pose you start from. It pays off in the right place — late, careful optimization of a known series — and wastes compute if you ask it to compare wildly different molecules or rely on a pose you are not sure of. Use it as a precision filter, and keep it honest by checking it against compounds you have already measured.
- Use FEP in lead optimization, where you have a trusted binding pose and are exploring small edits on a fixed scaffold.
- Keep perturbations small and chemically sensible — H→F, adding a methyl, a single ring swap — so the cycle stays reliable.
- Anchor the calculations to measured analogs so you can check the predictions against real data.
- Rank the virtual ideas, synthesize the top few, feed the results back, and repeat — FEP is a filter inside the design cycle, not a replacement for the lab.