Skip to main content
This page compares RevFEP with the FEP+ results Schrödinger published on the JACS benchmark set (Wang et al., 2015): eight protein–ligand systems, scored against experimental binding affinities. RevFEP has run all eight in production.
RevFEP numbers are our own production runs, scored with the same metrics as the paper. Schrödinger FEP+ numbers are the values published in Table 2 of Wang et al. (2015); we did not re-run them. The baseline methods (molecular weight, Glide SP and MM-GB/SA) come from the same table. For the current FEP+ release (21-4) and for Rowan, OpenFE, Uni-FEP and XFEP, see RevFEP benchmarks.

​
Where RevFEP is ahead

RevFEP has a lower relative-ΔΔG error than the published FEP+ results on Thrombin, TYK2, JNK1 and CDK2, and a higher correlation with experiment on Thrombin, TYK2 and CDK2. Averaged over all eight systems, RevFEP is close to the published FEP+ results on error: mean MUE is 0.87 vs 0.86 kcal/mol and mean RMSE is 1.13 vs 1.08. The four wins roughly offset the four systems (p38, PTP1B, MCL1 and BACE1) where FEP+ is ahead.

​
Scorecard

Every system, RevFEP against the published FEP+ value, on the three metrics Table 2 reports.

​
How far ahead

The gap on each system, with the sign chosen so that a bar to the right is a RevFEP win. On Thrombin and CDK2, RevFEP’s error is about 0.2 to 0.4 kcal/mol below the published value, and on CDK2 its correlation is 0.37 higher.

​
Side by side, with Schrödinger’s error bars

The same comparison as values rather than gaps. Schrödinger published an error bar for every number in Table 2, and RevFEP’s single runs have none, so the grey whiskers show how much of each gap is inside the published uncertainty.

​
Correlation against the other scoring methods

Table 2 also reports the correlation of three simpler scoring methods with experiment. RevFEP’s mean correlation is well above Glide SP (0.29) and MM-GB/SA (0.40), and below the published FEP+ value (0.74).

​
Every ligand, predicted against experiment

Pick a system to see each ligand’s RevFEP ΔG against experiment. A ✓ marks a system where RevFEP beats the published FEP+ value on at least one metric.

​
How the errors are distributed

​
Baselines from the same paper

Pearson R against experiment for the scoring methods in Table 2 of Wang et al. (2015), next to RevFEP. MW is molecular weight, Glide SP is docking score, and MM-GB/SA is an end-point free-energy estimate.

​
Full results

Schrödinger values include the ± error bars published in the paper. RevFEP values are single runs and have no error bar. “FEP+ edges” is the number of edges in Schrödinger’s network; RevFEP ran a minimal spanning tree. Bold marks the better value in each pair. MCL1 is level on MUE (1.17 vs 1.16), so neither value is bold.

​
Compute and runtime

GPU-hours and wall-clock time for the production RevFEP campaigns behind the numbers above. Campaigns that failed at the smoke stage and were re-run are excluded (CDK2 once, MCL1 three times). GPU-hours are partly estimated from sampling timings, so treat them as estimates. BACE1 has no recorded GPU-hours.

​
How the RevFEP runs were done

Every system used the standard RevFEP relative protocol with 11 λ windows, one repeat, and a minimal spanning tree of edges (one path between ligands, no closed cycles). The full settings are on the RevFEP benchmarks page.
The protocol is the same for every system. The structure preparation was not identical, and these steps were needed on four systems. BACE1 is listed last because it was run separately:
  • CDK2: a Lys89 side-chain repair of the crystal structure. The first attempt failed the smoke stage and is superseded. Schrödinger’s published CDK2 result was computed on the unrepaired structure, so this comparison is directional rather than strictly like for like.
  • MCL1: a side-chain-repaired protein and a curated network with one pose per compound (41 of the 42 published ligands). Three earlier attempts are superseded.
  • PTP1B: the Cys215 thiolate form with four crystal waters removed.
  • JNK1: both binding poses of 17 compounds are run as separate ligands (38 entries for 21 compounds). Correlation is computed per compound after averaging each pair, which gives R = 0.65; per pose it is 0.27.
  • BACE1: a 35-edge minimal spanning tree over the 36 ligands, run with the OpenFF force field and standard (not enhanced-sampling) calculations. Predictions are aligned to the mean experimental value, as for the other systems. Anchoring them to one reference ligand (CAT-13a) instead changes MUE from 0.98 to 1.02 kcal/mol. Schrödinger’s published run used the OPLS force field with enhanced sampling and a denser network (58 edges against 35), so the two runs differ in force field, sampling and network as well as in the software.
MUE and RMSE are taken over the predicted ΔΔG edges of each network, against the experimental difference. R is the Pearson correlation between the per-ligand ΔG RevFEP predicts and experiment, after the ΔΔG edges are combined into one ΔG per ligand. Schrödinger’s values are the ones in Table 2 of the paper.

​
Reading these results

They come from the original paper and reflect the software and force field of that time. Schrödinger’s current release performs better on these systems. For a comparison against their current public benchmark data, see RevFEP benchmarks.
Schrödinger ran redundant networks with closed cycles. RevFEP ran one repeat on a spanning tree, so there is no cycle-closure check and no statistical error bar. Edge errors are computed over each method’s own edges, so the two edge sets differ.
Some wins are within the error bars Schrödinger published, for example the JNK1 MUE (0.66 vs 0.78 ± 0.12, a gap of 0.12) and the Thrombin correlation (0.80 vs 0.71 ± 0.24). Thrombin has 11 ligands, so its correlations move a lot when one ligand changes. MCL1 is level on error (1.17 vs 1.16).
p38, PTP1B, MCL1 and BACE1 have higher error than the published FEP+ results, and JNK1, MCL1, PTP1B, p38 and BACE1 have lower correlation. BACE1 is the clearest gap: 0.98 against 0.84 kcal/mol MUE, and R of 0.66 against 0.78. p38 is the weakest system, with 25 of 66 legs below 0.1 overlap, which points to undersampling.
BACE1’s GPU-hours are not in the end-to-end report that the compute figures come from, so it is left out of the compute table and totals rather than estimated.

​
Sources

  • Schrödinger FEP+ and baseline methods: 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
  • RevFEP results, per-edge errors, per-ligand predictions, compute time: Revilico’s end-to-end production report for the JACS campaigns (October 2026), using only the final campaign for each system. BACE1 comes from its separate campaign (per-ligand and per-edge results, 36 ligands and 35 edges). The numbers on this page are recomputed from its edge and ligand tables.
  • Table 2 reference values as used here (including PDB IDs, network sizes and affinity ranges): Revilico’s competitor benchmark tracker, which copies them from the paper.
  • PDB structures: 2ZFF (thrombin), 4GIH (TYK2), 2GMX (JNK1), 1H1Q (CDK2), 4DJW (BACE1), 4HW3 (MCL1), 2QBS (PTP1B), 3FLY (p38).
  • Other methods on the same systems: see RevFEP benchmarks for FEP+ 21-4, Rowan FEP/TMD, OpenFE, Uni-FEP and XFEP.