Molecular docking predicts how a small molecule (ligand) binds to a protein by computationally searching for the optimal position, orientation, and conformation of the ligand within the protein's binding site. Docking programs (AutoDock, Glide, GOLD) use search algorithms (genetic algorithms, Monte Carlo sampling, systematic enumeration) to explore the conformational space and scoring functions to estimate binding affinity. Docking is widely used for virtual screening (identifying potential hits from large compound libraries) and binding mode prediction (understanding how a known ligand interacts with the target). Key challenges include protein flexibility (most docking treats the protein as rigid), accurate scoring (current functions poorly rank binding affinities), and the entropic contribution of solvent displacement.
The interaction between a protein and a small molecule — a drug, a metabolite, a signaling molecule — is fundamentally a problem of molecular recognition: how does the ligand find the right binding site, adopt the right orientation, and form the right combination of interactions to achieve high affinity and selectivity? Molecular docking attempts to predict this recognition computationally, and its successes and limitations reveal the physical principles governing molecular binding.
A docking calculation has two components: a search algorithm that explores the ligand's possible positions, orientations, and conformations within the binding site, and a scoring function that evaluates each candidate pose. Search algorithms (genetic algorithms, Monte Carlo sampling, fragment-based growth methods) must efficiently explore a vast conformational space — the ligand has 3 translational, 3 rotational, and multiple torsional degrees of freedom. Scoring functions estimate the binding energy from the properties of the docked pose: shape complementarity (how well the ligand fills the pocket), hydrogen bonds (number and geometry), electrostatic interactions (charge complementarity), hydrophobic contacts (desolvation of nonpolar surfaces), and strain energy (the energetic cost of the ligand adopting its bound conformation).
Docking is remarkably good at pose prediction — placing the ligand in approximately the correct orientation and position. For drug-like ligands binding to well-defined pockets, docking reproduces crystallographic binding modes (RMSD < 2 Angstroms) in 70-80% of cases. Docking is much worse at affinity prediction — ranking ligands by how tightly they bind. The scoring functions are too approximate to capture the subtle energetic differences (often < 1 kcal/mol) between tight and weak binders. Key missing elements include the entropic cost of reducing the ligand's conformational freedom upon binding, the energy of displacing ordered water molecules from the binding site, and the protein's conformational response to binding (induced fit).
Virtual screening applies docking at scale: millions of compounds from commercial libraries or virtual chemical spaces are docked to a target, and the top-scoring compounds are purchased and tested experimentally. The enrichment (improved hit rate compared to random screening) typically justifies the computational investment, making docking a standard first step in drug discovery campaigns. More accurate but computationally expensive methods — free energy perturbation (FEP), molecular dynamics with enhanced sampling, and machine learning models trained on structural and activity data — are used for lead optimization, where quantitative affinity prediction matters more. The hierarchy from fast-but-approximate (docking) to slow-but-accurate (FEP) mirrors the drug discovery funnel from broad screening to focused optimization.