Skip to main content
MM/GBSA and MM/PBSA estimate a ligand’s binding energy from a molecular dynamics trajectory of the complex, at a fraction of the cost of a free-energy perturbation calculation. This page asks the practical question: do the scores rank ligands the way experiment does? Each of ten protein targets has a five-ligand panel chosen to span a wide range of measured affinities (1.7 to 6.1 kcal/mol between the weakest and strongest ligand). Each ligand was simulated in complex with its target for 50 ns, and for 100 ns on three of the targets, and scored with both methods.
All results here are our runs. Experimental binding free energies come with the ligand panels, which are taken from Schrödinger’s public benchmark. Each method’s score is the average over the trajectory, and the comparison with experiment is made within a target, ligand by ligand.

​
What the scores show

MM/GBSA ranks the five ligands in the right direction on 8 of the 10 targets and gets the order exactly right on two of them, TYK2 and thrombin. Its mean rank correlation is 0.37 at 50 ns, against 0.26 for MM/PBSA. It fails on HIF-2α, where the scores run against experiment (ρ of −0.9), and is flat on SHP2 (−0.1). MM/PBSA is the better of the two on 3 of the 13 runs. On the three targets simulated for 100 ns, the longer run improves both methods, MM/PBSA more than MM/GBSA, but with only three targets that is an observation, not a rule.
Each target has five ligands, so every correlation here rests on five points. A perfect ranking of five ligands happens by chance once in 120 orderings, and anything below about 0.9 cannot be told apart from chance on a single target. Read the averages across targets, not the individual cells, and treat this as a screening-scale result.

​
By target

Spearman ρ and Pearson r between each method’s average binding energy and the experimental ΔG, over the five ligands of each target.

​
Each run, side by side

​
Which method ranks better

The difference in rank correlation for each run, MM/GBSA minus MM/PBSA.

​
Binding energy against experiment

Pick a target and a method. The energies are raw MM binding energies, without an entropy term or any calibration, so only the order and trend matter, not the absolute values.

​
For context: the published MM-GB/SA results

Six of these targets are also in the JACS set, where Schrödinger published the correlation of an MM-GB/SA score with experiment (Wang et al., 2015, Table 2). Those results are on each full series of 11 to 42 ligands, not on five.
These two columns are not a head-to-head comparison. Our panels are five ligands chosen to span a wide affinity range, which makes a correlation easier to find than on a full series with a narrower spread, and five points are far noisier than dozens. They show that the two approaches land in a similar range, not that one beats the other.

​
How the scoring was run

Each target has five ligands taken from Schrödinger’s public binding free energy benchmark, chosen to cover the range of measured affinities for that target. The targets are BACE, c-Met, Galectin-3, HIF-2α, MCL1, p38, PTP1B, SHP2, TYK2 and thrombin.
Each ligand was simulated in complex with its target for 50 ns. Galectin-3, HIF-2α and MCL1 were also run for 100 ns. MM/GBSA and MM/PBSA binding energies are averaged over the trajectory, along with their standard deviation and standard error. The score compared with experiment is the TOTAL binding energy from each method.
For each target and method, Spearman ρ and Pearson r are computed between the five average binding energies and the five experimental ΔG values. Targets are then averaged with equal weight.

​
Reading these results

The panels are deliberately small. They test whether a cheap score can pick the strongest from the weakest binder, not whether it can resolve ligands a fraction of a kcal/mol apart, which is where free-energy perturbation is used.
MM/GBSA and MM/PBSA energies carry no entropy term and are not calibrated to experiment, so they run far from measured free energies in absolute terms. This page reports ranking only.
This page does not compare these scores with FEP results or report cost. See the RevFEP benchmarks for relative free-energy results on full series.

​
Sources

  • Experimental affinities and ligand panels: Schrödinger, public binding free energy benchmark.
  • Published MM-GB/SA reference: Wang, L. et al. Accurate and Reliable Prediction of Relative Ligand Binding Potency in Prospective Drug Discovery by Way of a Modern Free-Energy Calculation Protocol and Force Field. J. Am. Chem. Soc. 137, 2695–2703 (2015), Table 2. doi:10.1021/ja512751q
  • MD and MM/GBSA, MM/PBSA results: our runs, 13 MD runs (65 ligand simulations) across ten targets, scored as described above.