The distance between a research question and a runnable computational experiment is easy to underestimate. In the wet lab, "does this compound bind our target?" has a direct answer: run a binding assay. Computationally, it requires decisions before calculation: which protein structure and binding site, which method, what flexibility model, and which scoring settings? Sound choices separate actionable screens from noisy results.
This note follows computational experiment design from wet-lab question to submitted job, including key decisions and defaults that are usually acceptable.
Step 1: Convert the scientific question into computable form
A question becomes computable when its inputs and outputs have interpretable meanings. "Does compound X bind target Y?" is incomplete because it omits the binding-mode question, expected affinity range, and target-structure assumptions. A computable version is: "Does compound X adopt a geometrically plausible binding pose in the ATP-binding site of kinase Y (PDB: 6ABC), and what is the predicted docking score relative to reference inhibitor Z?"
Do this translation before submitting anything. Define a positive result as a good docking score plus a sensible pose, a negative result as no stable pose or a score worse than the negative control, and an ambiguous result as a score between the positive and negative controls without a clear pose rationale. Without a decision boundary set in advance, teams often receive results they cannot interpret.
Step 2: Select and validate the protein structure
Protein structure choice has the greatest effect on docking quality. Prefer an experimental structure with resolution better than 2.5 angstroms, a holo conformation bound to a chemically similar ligand, and a binding site free of crystal contacts or disorder that could disrupt docking.
When several target structures exist, test the same reference compound against each. This is a useful 30-minute investment. Use the structure that repeatedly produces better scores and more sensible poses for known active compounds. If results are similar, choose the best resolution. Dramatic differences indicate a conformational-flexibility problem, often requiring ensemble docking or IFD before screen results can be interpreted.
Use a stricter standard for homology models. A model based on a template with greater than 50 percent sequence identity and the same domain architecture is generally suitable for docking. A remote template with uncertain domain correspondence is risky because its binding-site geometry may be wrong without experimental validation. If known binders exist, use them to validate the model before a large screen.
Step 3: Define the binding site
Binding-site definition sets the protein surface searched by docking. Define it from a co-crystallized ligand, the most reliable option, from cavity detection such as FPocket or SiteMap, or from literature describing key binding residues.
The docking grid box should cover the whole binding site, leaving a few angstroms of clearance on every side. A typical kinase ATP-binding site generally needs 20 to 22 angstroms per side. Larger allosteric sites or PPI interfaces need larger boxes. A box that is too small can reward compounds that fit entirely inside it, while an oversized box includes surface patches where non-specific binding can inflate scores.
For multiple candidate sites, such as orthosteric and allosteric sites on one target, run separate screens with separate grids. One large grid hides which site each compound is predicted to bind.
Step 4: Prepare the ligand set
Ligand preparation requires 3D conformation generation, protonation assignment, and tautomer enumeration. Do all three.
For standard flexible docking, one reasonably optimized conformer per compound is enough. The algorithm explores torsional space, so the starting conformer need not be perfect. ETKDG (RDKit), or an equivalent method, followed by force-field minimization with MMFF94 or UFF provides a good starting geometry.
At pH 7.4, protonation assignment changes the charge of amines, carboxylic acids, sulfonamides, and other ionizable groups. A tertiary amine protonated at pH 7.4 can have a different docking score and preferred pose from its neutral form. Use a tool that considers each ionizable center's pKa in the full molecule, not the isolated functional group.
Tautomer enumeration matters for heteroaromatic systems with plausible alternatives, including pyridone versus hydroxypyridine and keto versus enol. For a large screen, docking only the most stable tautomer is generally sufficient. For high-priority compounds, checking the alternative is worth the extra run.
Step 5: Choose the docking method
For a first-pass screen against a well-characterized target, begin with rigid-receptor docking. It is fast, validated, and its failure modes are understood. The common open-source choice is AutoDock Vina. Commercial options such as Glide SP and GOLD provide similar performance, with better-optimized scoring for some target classes.
Use induced-fit docking (IFD), which permits limited receptor side-chain and backbone flexibility near the site, for targets with documented flexibility after ligand binding. It also suits compounds unlike the reference ligand used to define the site. The cost is longer runtime, typically 10 to 20 times slower than rigid docking per compound. That is acceptable for 20 to 50 focused compounds, but costly for a 500-compound screen.
Reserve free-energy methods, including FEP and thermodynamic integration, for the final 5 to 10 compounds entering synthesis, not screening. They give the most accurate relative affinity estimates but need substantially more setup and compute time per compound pair.
Step 6: Set analysis criteria before viewing results
Decide beforehand how to use the output. Set the docking score threshold for calling a likely binder, the residues a pose must contact, and the physicochemical filters for top-scoring compounds. Pre-setting these rules avoids post-hoc threshold changes made to include a favored compound.
Predefined criteria also check screen quality. If reference active compounds fail the filters, the setup has a problem. Recover known actives before trusting rankings for unknown compounds.
