MM-GBSA rescoring: when physics beats the docking score

A docking score is a fast, cheap guess. It has to be: a scoring function evaluates millions of poses per screen, so it trades physical realism for speed. MM-GBSA is the next rung up the ladder - a physics-based endpoint you apply to a handful of survivors to get a better-correlated ranking. It is one of the most useful tools in a structure-based workflow and one of the most over-trusted, so it is worth being precise about what it actually buys you.

What MM-GBSA computes

MM-GBSA stands for Molecular Mechanics with Generalized Born and Surface Area solvation. Conceptually it estimates a binding free energy as the energy of the complex minus the energy of the free protein and free ligand, where each term combines a molecular mechanics gas-phase energy with an implicit-solvent correction. The Generalized Born model approximates the electrostatic cost of desolvation, and the surface-area term approximates the nonpolar contribution.

The key difference from a docking score is that MM-GBSA uses a real forcefield and an explicit solvation model rather than an empirical sum of pre-fit terms. That makes it slower by orders of magnitude, but it captures desolvation and electrostatics in a way most fast scoring functions only approximate. The sibling method MM-PBSA swaps the Generalized Born term for a Poisson-Boltzmann solvation calculation - more rigorous, slower, and in practice often no more accurate for pose ranking.

Where it genuinely helps

• Pose discrimination. In a systematic benchmark of 98 protein-ligand complexes, MM/GBSA identified the correct binding conformation about 69% of the time, beating MM/PBSA (around 46%) and several popular docking scoring functions (Hou et al., 2011). That is its core job: of the poses docking gave you, which is the real one.
• Congeneric series ranking. For a set of closely related analogs - the situation in lead optimization - MM-GBSA tends to give cleaner rank correlation with measured affinity than a raw docking score, because systematic errors partly cancel across similar molecules.
• Ensemble averaging. Rescoring over multiple snapshots rather than a single minimized structure can lift the correlation substantially - in one antithrombin study the R-squared went from 0.36 to 0.69 moving from single-structure to ensemble-averaged MM/GBSA.

Where it quietly fails

MM-GBSA is not a free-energy method in the rigorous sense. It ignores configurational entropy unless you bolt on a separate (noisy) normal- mode or quasi-harmonic estimate, and the absolute numbers it produces are not real binding free energies - they are a scale that happens to correlate, sometimes. Treat an MM-GBSA value of -45 kcal/mol as a ranking coordinate, not a ΔG you can convert to a Kd.

Two failure modes matter most in practice. First, results are highly sensitive to the choice of internal dielectric and GB model, so a number is only meaningful relative to others computed the same way. Second, if you feed it a docked pose rather than a crystal structure, you inherit the docking error on top of the MM-GBSA error - garbage pose in, confident-looking garbage energy out. As Genheden and Ryde put it in their widely cited review, the methods are useful but their accuracy is system-dependent and easily overstated.

How it fits in a docking workflow

The sensible pattern is a funnel. Use fast docking to screen broadly and generate poses, keep the top survivors, then rescore that short list with MM-GBSA to re-rank. Reserve true alchemical free-energy methods (FEP) for the final few candidates where the extra compute is justified. MM-GBSA sits in the middle: cheaper than FEP, more physical than a docking score, and most valuable when you are comparing similar molecules against the same target rather than ranking diverse scaffolds.

The recurring theme - the same one behind comparing Vina against GNINA, or watching ΔΔ instead of absolute scores - is that no single number is the answer. Each tier of method corrects a different bias, and MM-GBSA earns its place by fixing the pose-ranking step that fast scoring functions get wrong most often.

Open Studio to dock a ligand against your target and inspect the top poses. When two poses score within a kcal/mol of each other, that is exactly the situation where a physics-based rescore changes the answer - so it is worth looking at the interaction pattern, not just the number.

Liganx puts molecular docking online and free in the browser, which makes it easy to generate the candidate poses that a downstream MM-GBSA or FEP step then re-ranks. Run molecular docking first, then spend the expensive compute only on what survives.

Primary sources

• Hou T, Wang J, Li Y, Wang W. Assessing the performance of the MM/PBSA and MM/GBSA methods. II. The accuracy of ranking poses generated from docking. J Comput Chem 32, 866-877 (2011). doi:10.1002/jcc.21666
• Genheden S, Ryde U. The MM/PBSA and MM/GBSA methods to estimate ligand-binding affinities. Expert Opin Drug Discov 10, 449-461 (2015). doi:10.1517/17460441.2015.1032936
• Sun H, Li Y, Tian S, Xu L, Hou T. Assessing the performance of MM/PBSA and MM/GBSA methods. 4. Accuracies of MM/PBSA and MM/GBSA methodologies evaluated by various simulation protocols. Phys Chem Chem Phys 16, 16719-16729 (2014). doi:10.1039/C4CP01388C