Small drug discovery and chemical biology teams often ask whether they can use computational docking without a computational scientist on staff. Yes, but the method's outputs and limits must be understood, with interpretation handled by a team biologist or chemist.
This note is for PIs and medicinal chemists with biological or synthetic chemistry expertise who have not run docking calculations. It explains the core concepts, output meaning, and key points for evaluating results.
What protein-ligand docking calculates
Molecular docking takes 3D protein and small-molecule structures, then searches for the small molecule pose in the binding site that minimizes a scoring function. The function approximates binding free energy using van der Waals interactions, electrostatics, hydrogen bond geometry, and desolvation penalties. Results contain one or more ranked poses with scores, usually in kcal/mol or software-specific arbitrary units.
A docking score is not thermodynamic binding free energy. It is a fast empirical approximation trained to distinguish binders from non-binders. Absolute affinity prediction requires methods such as free-energy perturbation (FEP) or thermodynamic integration, which cost substantially more compute. Docking suits library rank-ordering and binder identification, not predicting Kd values such as 50 nM for compound A versus 120 nM for compound B.
What you need to run docking
You need a protein structure, defined binding site, and prepared ligand. The protein is usually a Protein Data Bank crystal structure, with resolution better than 2.5 angstroms generally suitable. Without an experimental structure, an AlphaFold model can be used, but its binding-site geometry is less reliable, particularly for induced-fit sites requiring a conformational change.
A co-crystallized inhibitor can define the site through its position. Without one, cavity tools such as FPocket or Fpocket2 use protein-surface geometry to identify likely pockets. For allosteric sites, literature gives the approximate location, while homologous structures may help infer exact geometry.
Ligand preparation causes many first-use errors. A database or drawing-tool SMILES string needs a 3D conformation, relevant-pH protonation states, and partial charges. LigPrep from Schrodinger, OpenBabel, or RDKit with ETKDG conformer generation can handle this. A flat 2D input to a program expecting 3D produces meaningless results.
Interpreting docking output
Do not sort by score and send the top compounds straight to synthesis. First inspect poses in PyMOL, UCSF Chimera, or similar software. For each high-scoring compound, ask whether it sits inside the site, forms at least one hydrogen bond with a literature-supported binding residue, and adopts a geometrically reasonable rather than highly strained conformation.
A strong numerical score can accompany an obviously incorrect pose, due to a spurious local minimum or incomplete penalty for protein-surface overlap. Visual inspection catches this. An experienced medicinal chemist needs five to ten minutes per compound, little compared with a failed synthesis.
Next review molecular weight, logP, hydrogen bond donors and acceptors, and rotatable bonds. A well-scoring compound with five hydrogen bond donors, logP above 5, and molecular weight above 500 may face major oral bioavailability challenges. Docking is not the stage to optimize bioavailability, but removing compounds that break Lipinski's rule of five or target-specific criteria avoids obvious dead ends.
Methods beyond basic docking
Rigid-receptor docking fixes the protein, which is an approximation. Flexible side chains and ligand-induced loop movement can matter. Induced-fit docking (IFD) permits limited receptor flexibility and often improves pose prediction for flexible sites, but takes longer. For 20 to 30 compounds, run IFD when site flexibility is plausible.
Molecular dynamics (MD) can test whether a proposed pose remains stable. A compound that docks well but leaves the site during a 50-nanosecond MD run may not bind or may use a mode docking missed. MD is unnecessary for every compound in a large screen, but adds a useful check for the final 5 to 10 candidates before synthesis, at a few hours of compute per compound.
Where docking does not replace expert judgment
Docking cannot assess synthetic accessibility. An ideal in silico compound may require a 12-step synthesis. A bench scientist must judge which chemical series the team's synthetic capabilities can feasibly elaborate, rather than relying on the score.
Docking cannot reliably predict proteome-wide selectivity without screens against all structural homologs. Even then, low-similarity off-targets such as hERG channel binding or CYP enzyme inhibition are better assessed with dedicated models, including many ADMET tools, or early assays. This reflects the method's scope, not a docking failure.
A small team should begin on a target with experimental binding data for calibration. Given IC50 values for five to ten known compounds, test whether docking ranks them consistently with experiment. A completely uncorrelated ranking signals a setup problem that must be fixed before trusting the screen. Adding this validation to the first few screens takes time but supports justified confidence in later outputs.
