Molecular docking and binding free-energy calculation
Methods that pose a small molecule in a protein pocket and estimate its affinity, from fast scoring functions to alchemical free-energy perturbation and learned co-folding affinity heads.
Virtual screening and lead optimisation in structure-based drug discovery choose which compounds to make using these poses and affinity estimates.
Predicting where a drug molecule sits on a protein and how tightly it holds on.
1Protein pocket plus molecule
2Place the molecule (pose)
3Estimate grip strength
4Compare with lab measurement
5Choose molecules to make
Pose and affinity. The binding free energy is ΔG=RTlnKd, where K is the dissociation constant. At room temperature, 1.36 kcal/mol is a factor of 10 in , so 1 kcal/mol is about a factor of 5.
Go deeper
The full technical text.
What it is
Two coupled questions: where does a ligand bind (pose) and how tightly (affinity, ΔGbind=RTlnK). A 1 kcal/mol error changes by a factor of about 5.4 at 298 K, which is why 1 kcal/mol is the usual target accuracy.
Isomorphic Labs reports a closed co-folding and design engine with large gains over AlphaFold 3 on internal generalisation benchmarks. No model, weights or refereed report are available, and the circulating figures were not confirmed from a primary source.
Docking. A search algorithm samples the molecule's position, orientation and rotatable bonds; a scoring function ranks poses. Classic tools use hand-built scoring terms; deep-learning docking (DiffDock) and co-folding models (AlphaFold 3, Boltz, Chai) learn from known complexes. Tests such as PoseBusters check whether predicted poses are chemically valid, not just close to the answer.
Free-energy perturbation (FEP). Binding is a state change from molecule in water to molecule in the pocket. Instead of simulating the real process, we "alchemically" morph molecule A into B using a coupling parameter λ between 0 and 1, in the protein and in water. The difference of the two legs equals the difference in binding free energy between A and B. Estimators such as the Zwanzig formula, Bennett's acceptance ratio and MBAR combine the simulations. Quality depends on the force field, on getting charges and water right and on enough sampling.
Learned affinity. Models trained on measured affinities can be very fast, but many public datasets leak: test molecules are near-copies of training ones. Boltz-2 (2025, preprint) adds an affinity head and reports FEP-like correlation on selected sets at about a thousand times lower cost; independent prospective tests are still needed.
Noise floor. Replicate experiments from different labs disagree by about 0.9 kcal/mol RMSE, so no method can be judged much more finely.
Open problems in DARC. Accuracy of about 1 kcal/mol on new scaffolds across protein families, and co-folding that works on unseen chemistry.
d
Kd
The core machinery
Docking. Search over ligand position, orientation and torsions, scored by an empirical or knowledge-based function. Classic tools: DOCK (1982), Glide, AutoDock Vina (2010), GOLD. Deep-learning docking (DiffDock, a diffusion model over ligand poses) and co-folding models (AlphaFold 3, Boltz, Chai-1) now often rival or exceed classical docking on pose accuracy, but PoseBusters-style checks show physically implausible geometries in many learned poses, and accuracy drops for ligands and pockets dissimilar to training data.
Alchemical free-energy perturbation (FEP). Instead of simulating binding, morph ligand A into ligand B along a coupling parameter λ in both complex and solvent; the relative binding free energy is ΔΔG=ΔGcomplex−ΔGsolvent, with each leg estimated by the Zwanzig identity, the Bennett acceptance ratio or multistate variants (MBAR). Commercial FEP+ (Wang et al., 2015) and open tools (OpenFE, GROMACS/OpenMM workflows) reach root-mean-square errors near 1 kcal/mol on congeneric series of well-behaved targets. Absolute binding free energies (ABFE) remove the congeneric restriction at higher cost.
Learned affinity. Scoring from graph networks on PDBbind-style data is notoriously hampered by data leakage; Boltz-2 (2025) adds an affinity head and reports approaching OpenFE-level correlation on selected benchmark sets at over 1000x lower cost (company/developer preprint).
Why it is hard
Protonation and tautomer states, ligand strain, water networks, induced fit and protein reorganisation, force-field error, charged perturbations, and the noise floor of assays: inter-laboratory reproducibility of binding assays is itself about 0.9 kcal/mol pairwise RMSE, so no method can be judged more finely than that.
Techniques to master
Statistical mechanics of free-energy estimators, cycle-closure and hysteresis diagnostics, enhanced sampling for slow protein degrees of freedom, benchmark hygiene (time splits, scaffold splits, prospective tests), prospective blind challenges (D3R, CASP ligand categories, Polaris).
Where it bites
Two problems: affinity prediction within about 1 kcal/mol on novel scaffolds across protein families without per-target tuning, and co-folding models that generalise physically rather than recalling training poses.
Computer simulation that moves every atom according to forces, step by tiny step, to watch a molecule wiggle and change shape.
Authors report average Pearson 0.66 on a four-target FEP+ subset and approaching OpenFE on 876 measurements, at over 1000 times lower cost; variance between assays is large.
What to read before Molecular docking and binding free-energy calculation (left), and what builds on it: entries that depend on it and the open problems it hosts (right). Every box links to its page.