Free-Energy Perturbation
Python API · stjames models · Graph example · FEP example
How it works
Relative binding free-energy perturbation (RBFE) compares the binding affinities of related ligands by alchemically transforming one into another in both the protein complex and solvent. The difference between these transformations gives the relative binding free energy, ∆∆G.
How to submit
The web workflow has three stages: Analogue docking → RBFE graph → Relative binding free-energy perturbation. Resubmit the output of each stage to carry the selected structures into the next submission form.
What you need
Start with a prepared protein, a bound reference ligand pose, and a list of related ligands in SMILES format. The ligands should have substantial structural overlap and the same net charge. Experimental poses are preferred; docking or co-folding can introduce pose errors that affect the FEP results. Unrelated ligands may need endpoint methods or ABFE.
If you already have aligned bound poses for every ligand, skip analogue docking and add those structures directly to an RBFE graph workflow in Step 2. The screenshots below use a Rock2 system from Uni-FEP.
Step 1: Generate aligned analogue poses
Open Analogue docking and supply the target protein, the bound reference pose, and the analogue SMILES. If the protein needs preparation, use the protein editor to resolve the reported issues before submitting; see protein preparation. Include the reference ligand's SMILES explicitly if you want it in the final affinity comparison: supplying a reference pose does not automatically add it to the analogue list.
Submit the analogue docking workflow. When it finishes, inspect the poses for each analogue, compare docking scores and MCS RMSD, and overlay the ligands to check that their shared cores align in the binding pocket. Choose one suitable pose per ligand for the FEP campaign. Resolve failed or poorly aligned poses before continuing; poses from separate docking runs can also be added at the graph-building stage.
Step 2: Resubmit poses to build an RBFE graph
Select the poses you want to carry forward and click Resubmit. Confirm the selected ligand structures and protein in the resubmission dialog, then choose RBFE graph workflow. The graph submission form opens with the selected ligand geometries; check that every intended compound is present, give the ligands recognizable names, and add any additional aligned poses.
The graph determines which pairs of ligands will be transformed into one another: each node is a ligand and each edge is an alchemical transformation. Start with Greedy graph mode and the default scoring and redundancy settings, or choose Star Map and a Hub compound to connect every ligand directly to one reference. The graph settings are described below.
Submit the RBFE graph workflow and inspect the completed graph in the Graph explorer. This stage builds the transformation network; the free-energy simulations run in Step 3.
Step 3: Resubmit the graph to run RBFE
On the completed graph workflow, click Resubmit to FEP to open the relative binding free-energy perturbation submission form. Resubmission carries over the graph and ligand poses and attempts to recover the protein from the preceding workflow. Verify the target protein; if you built the graph directly from ligand structures, add the prepared protein now.
Review the graph and use its pencil icon to add or remove edges if needed. Choose a TMD settings preset; Recommended (Default) is the starting point for a first run. Expand show details to adjust individual simulation settings.
List any cofactors that need parameterization, then run the System compatibility check. Resolve any reported protein-preparation or parameterization errors and repeat the check. Once it passes, submit the RBFE workflow.
As simulations finish, inspect the per-edge results and lambda overlap. Once the workflow completes, review the per-ligand relative affinities together with convergence, cycle closure, and uncertainty diagnostics before ranking compounds.
Settings
- Graph mode: "Greedy" connects similar ligands with redundant paths. "Star Map" connects every ligand to the selected "Hub compound," using fewer edges but providing no cycle checks.
- Scoring method: "Jaccard" uses atom-mapping similarity; "Dummy atoms" uses the number of unmatched atoms. "Best" (default) tries both and selects the graph with fewer dummy atoms.
- K min cut: target graph redundancy, default 3, for greedy graphs. Higher values generally add edges and cost. "Generate intermediate ligands?" can split large transformations into smaller ones.
- Cofactors: list fixed small molecules in the protein structure that need parameterization, with residue identifiers and SMILES. Known ions and waters do not need SMILES.
- Forcefield / Charge method: choose ligand parameters and partial charges. The web default uses OpenFF Sage 2.0.0 with NAGL; NAGL is faster than AM1-BCC for large ligands. The web form requires NAGL for Mango.
- Equilibration steps / Num frames: defaults are 200,000 equilibration steps (0.5 ns) and 2,000 production frames (2 ns) per window, at 400 steps per frame and a 2.5 fs timestep.
- Initial lambda windows / Minimum overlap / Target overlap: defaults are 48 maximum windows and 0.667 for both overlap targets. Adaptive scheduling uses fewer windows for easier transformations.
- Local MD steps: default 390 of each 400-step frame; 0 disables local resampling. "REST max temperature scale" defaults to 1 (disabled); higher values enable solute tempering.
"Recommended (Default)" uses the settings above; "Fast" shortens sampling and lowers overlap targets. "Rigorous" disables local resampling and uses AM1-BCC charges.
Notes
Run the "System compatibility check" to catch missing atoms, nonstandard residues, or incompatible protein naming. Membrane embedding and covalent bond formation are unsupported; compare only the noncovalent contribution of covalent inhibitors. Compare different ligand charge states in separate groups.
Rowan uses TMD, derived from Time Machine, with local resampling, adaptive lambda scheduling, water sampling, and optional REST.
Results include per-edge ∆∆G and uncertainties, per-ligand relative affinities in kcal/mol, cycle closure, overlap, convergence, and replica mixing. Saved trajectories cover the complex and solvent endpoints. Inspect these diagnostics before interpreting affinities; uncertainty estimates do not capture all pose, force-field, or sampling errors.
For comparison with experimental absolute affinities, add a constant to align dataset means before calculating MAE or RMSE. Rowan does not propose analogues automatically; CReM is one possible source.
Submission video
Benchmarks and validation
Compare binding-affinity predictions across JACS, Merck, Waterset, and additional systems. Select a benchmark to inspect the individual compounds: free-energy-perturbation benchmarks.