Prospective protein–ligand binding free energies at 1 kcal/mol RMSE across diverse, difficult targets
Difficulty 50: routine for easy congeneric series, hard in failure classes, and the experimental noise floor sits near the target. Application value 74: lead optimisation in drug discovery has large economic stakes. Leverage 42: shapes validation of force fields and ML/MM hybrids but few other fields depend on it.
The problem
Predicting how tightly a drug candidate binds its target to within one kcal/mol, reliably and in advance.
- 1Drugs must stick to targets
- 2Morph one molecule into another
- 3Add energy of each step
- 4Good cases ~1 kcal/mol
- 5Hard cases and experiments limit
Alchemical free-energy methods. For relative binding free energy between ligands A and B, one computes with each leg obtained by morphing A into B with a coupling parameter , as explained in the free-energy entry. Absolute binding free energies instead decouple the ligand completely and add restraints for the bound state.
Built from
earlier problems → this oneThe problems this one is built from: progress or tools it needs (builds on), special cases that come first (generalizes), and weaker results on the way (implies). On the right, what builds on it.
- Start with
- One classical force field accurate for folded proteins, disordered proteins and weak protein–protein associationComputational Chemistry · D51.5
- Neural-network wavefunctions at chemical accuracy for 100+ electron molecules with transferable, amortised trainingComputational Chemistry · D54
- Predicting fold-switching proteins and both of their folds from sequenceComputational Medicine · D58
- Step 2
- Converged reference interaction energies for 100-atom noncovalent complexes: reconciling CCSD(T) and diffusion Monte CarloComputational Chemistry · D40
- Boltzmann-weighted conformational ensembles of multi-state proteins from sequenceComputational Medicine · D64
- Step 3
Path to understanding
foundations → problemThe research entries this problem rests on, in the order to read them. Each links to its explainer.
- Start with
- Step 2
- Step 3
- Free-energy methods and enhanced samplingAlgorithm · host
- Then
- Prospective protein–ligand binding free energies at 1 kcal/mol RMSE across diverse, difficult targetsThe problem
State of the art
- Established practice. Wang et al. (JACS 2015) reported FEP+ relative binding free energy errors near 1 kcal/mol over a curated set of drug-discovery targets; Ross et al. (Commun. Chem. 2023) assembled the largest public benchmark, show that with careful structure and ligand preparation FEP can reach accuracy comparable to experimental reproducibility, and quantify that reproducibility.
- Limits. Accuracy degrades for charged ligands, large changes in ligand shape, protein conformational change and poorly resolved structures; a 2025 J. Chem. Inf. Model. study is reported to show that accuracy is compromised by structure modelling and other upstream choices (reported; not re-checked against the primary paper for this entry). Absolute binding free energies are less accurate than relative ones: a retrospective fragment-optimisation study (J. Chem. Inf. Model. 2022) reported good ranking (Spearman about 0.89) but an RMSE of 2.75 kcal/mol.
- New tools. Open-source workflows (OpenFE, 2022 onward) and a recent open absolute-binding-free-energy toolkit reporting large-scale benchmarks with ranking comparable to relative methods (arXiv:2603.22274, 2026, preprint; the reported scope was not independently checked for this entry) reduce cost; machine-learned potential/MM hybrids are being tested for relative free energies (Evaluating electrostatic embedding MLIP/MM for RBFE, arXiv:2608.13355, preprint).
- Open. Systematic prospective trials with sub-kcal/mol RMSE on difficult classes, and an assessment of whether force-field error (rather than sampling, structure or protonation) is the remaining limit, are not available.
Realistic applications
Binding free energies drive lead optimisation in drug discovery: more accurate absolute predictions would allow virtual screening of unseen chemistries and reduce synthesis cycles; sub-kcal/mol relative predictions would let computation replace many assays.
What counts as solved
A pre-registered prospective campaign (or blind challenge with experimental values withheld) on at least 30 targets, covering the listed difficult classes, with RMSE at most 1 kcal/mol, performed by at least two independent groups using different engines.
How it gets verified
ObservationSettled by new measurements or observations.
Needs new experimental affinities: a pre-registered prospective campaign (or blind challenge with values withheld) on at least 30 targets covering the stated difficult classes must reach an RMSE of 1 kcal/mol or less, performed by at least two independent groups on different engines.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
Prospective trials and benchmarks, with force-field and sampling improvements.
Bottlenecks
ranked by votes- #1EmpiricalExperimental reproducibility ceiling
Measured affinities differ by about 0.8 to 1.0 kcal/mol between assays and labs.
Community grades
Sign in to add your own grade.
Connected problems
Progress
- Aug 1, 2023Maximal accuracysource
Ross et al. quantify the experimental reproducibility ceiling for FEP accuracy (date approximate).
- Jan 1, 2022Benchmark best practicessource
Hahn et al. define best practice for binding-affinity benchmarks (date given as year of publication).
- Jan 28, 2015FEP+ prospective accuracysource
Relative binding free energies with errors near 1 kcal/mol on curated sets (date approximate).