Skip to main content
RevFEP results on two public protein–ligand benchmark sets, the JACS set (Wang et al., 2015) and the Merck set (Schindler et al., 2020), using the curated structures from Ross et al. (2023), next to the published results of Schrödinger FEP+, Rowan FEP/TMD, OpenFE, Uni-FEP and XFEP. The page covers 20 targets and 475 ligands. Every RevFEP number except BACE1 comes from a production RevFEP run with the default settings in How the Benchmarks Were Run. BACE1 comes from a separate RevFEP run with three repeats per edge. Nothing was tuned per target.
RevFEP numbers are our own results: Revilico ran every target through the RevFEP pipeline. FEP+ numbers are not our runs. They are Schrödinger’s published FEP+ (release 21-4) predictions for the same ligands, downloaded from their public benchmark repository and scored here with the same metrics, so the two can be compared directly.Rowan FEP/TMD, OpenFE and Uni-FEP numbers are the summary statistics those groups publish for the same benchmark sets. We did not re-run them or see their per-ligand predictions, so they appear in the tables and per-target charts, but not in the ligand-level plots. XFEP publishes only an aggregate error distribution on the Merck set, so it appears in the error-bin chart and shows - in the per-target tables.

FEP benchmark overview

Every target against every competitor, one metric at a time. A bold RevFEP value beats at least one other method in its row, and the italic, underlined value is the best in the row. A - means no comparable result. The set and overall averages are means of the per-target statistics, so each target counts once whatever its size. Click a target to open it below.

Benchmark details

Pick a target to plot every ligand’s predicted against experimental binding free energy. Click a point or a ligand row to see the molecule and its values.
  • JNK1: the RevFEP run treats the two flipped binding poses of 17 molecules as separate ligands (38 entries for 21 molecules). FEP+ combines each pair into one value per molecule, so this target is not a like-for-like comparison and is left out of the averages.
  • c-Met (+1, redocked): the same six ligands as c-Met (+1), started from redocked poses instead of the poses supplied with the benchmark. It is shown to illustrate how much the starting pose matters, and is left out of the averages.
  • BACE1: run separately from the production campaigns, with three repeats per edge. RevFEP’s 35 edges form a minimal spanning tree over the 36 ligands, run with the OpenFF force field and standard (not enhanced-sampling) free-energy calculations. As for every target, the ΔG values are shifted to the mean experimental value. Anchoring them to a single reference ligand (CAT-13a) instead gives almost the same result (MUE 1.02 against 0.98 kcal/mol). Schrödinger’s published BACE1 run used the OPLS force field with enhanced sampling and a denser network.
  • Ligands without an FEP+ prediction: CDK8 (1 ligand), SHP2 (1), SYK neutral (3) and SYK −1 (1) have ligands that are missing from the public FEP+ data. The plot shows them for RevFEP only, and the statistics use only the ligands with both predictions.

Charts

Averages of the per-target statistics for each benchmark set, over the targets every method has a result for (the shared targets rows of the overview table). RevFEP numbers are our runs; every other method is published data. Uni-FEP covers the JACS set only and does not report ρ.

How the Benchmarks Were Run

Every target was run through the standard RevFEP pipeline with its default settings. Nothing was tuned per target. Structures, ligand poses and protonation states are the ones published with the benchmark. RevFEP does not change the net charge of a ligand along an edge, so series that mix charge states (c-Met, Eg5, SYK) were run as separate same-charge subsets. The published FEP+ results include charge-changing edges and run each of these series as one network.

How the Numbers Are Calculated

RevFEP’s relative free energies (ΔΔG) are combined into one ΔG per ligand. Because relative calculations fix ΔG only up to a constant, the predictions for each target are shifted so that their mean equals the mean experimental value. The FEP+ values are shifted the same way, over the same ligands.
  • RMSE and MAE are the root-mean-square and mean absolute errors of ΔG, in kcal/mol. Lower is better.
  • R² is the square of the Pearson correlation coefficient. Higher is better.
  • ρ is Spearman’s rank correlation, which measures how well the ligands are ranked. Higher is better.
Rowan FEP/TMD, OpenFE and Uni-FEP values are the per-target statistics from each group’s published benchmark tables (Rowan and OpenFE through the Rowan benchmark site), recorded in our competitor tracker. They are not shifted or re-scored. A competitor result is shown only when its ligand set matches the one RevFEP ran to within one ligand, so these series have no value (-) for Rowan, OpenFE and Uni-FEP: c-Met (+1), SYK (−1) and TNKS2, which they split by charge state differently from us, and CDK8, where the published figure is a pairwise RMSE that is not comparable to an RMSE of ΔG. Uni-FEP reports RMSE and R² but not MAE or ρ. XFEP has no per-target results.
Ligand-level FEP+ (release 21-4) predictions are taken from Schrödinger’s public benchmark repository, matched to RevFEP ligands by name. The experimental values in both data sets agree to within 0.01 kcal/mol for every matched ligand.

Reading These Results

Each edge was simulated once (three times for BACE1), and a minimal spanning tree has no closed cycles, so apart from BACE1 these results come without a statistical uncertainty, and none has a cycle-closure check. Repeat runs and networks with cycles would add an error bar and a consistency check.
R² and ρ computed on fewer than about 15 ligands (c-Met, Eg5 neutral, SYK −1, Thrombin) can change a lot when one ligand moves.
They were produced with a different force field (OPLS4), a different sampling protocol, and per-target corrections for protonation states, tautomers and alternative ring flips that were applied by experts. RevFEP runs every target the same way with no manual corrections.
The largest errors are on the +1 series (c-Met, Eg5) and on SHP2. The redocked c-Met run shows that the starting pose alone can move the RMSE from 1.87 to 0.70 kcal/mol for the same ligands.
The other groups in Ross et al. (2023) have not been run with RevFEP.

Sources

  • 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). doi:10.1021/ja512751q
  • Schindler, C. E. M. et al. Large-Scale Assessment of Binding Free Energy Calculations in Active Drug Discovery Projects. J. Chem. Inf. Model. 60, 5457–5474 (2020). doi:10.1021/acs.jcim.0c00900
  • Ross, G. A. et al. The maximal and current accuracy of rigorous protein-ligand binding free energy calculations. Commun. Chem. 6, 222 (2023). doi:10.1038/s42004-023-01019-9
  • Rowan Scientific, protein–ligand benchmarks (FEP/TMD and OpenFE results on the Ross et al. sets): benchmarks.rowansci.com/protein-ligand, accessed 21 September 2026.
  • Uni-FEP: Breaking Barriers in FEP Benchmarking. ChemRxiv (2025). doi:10.26434/chemrxiv-2025-rf8mf
  • XFEP: Molecular Solutions Software, FEP case studies (2026), error distribution on the Merck set.
  • FEP+ 21-4 ligand predictions: schrodinger/public_binding_free_energy_benchmark, 21_4_results/ligand_predictions/