Molecular Docking

The engine that explores the ligand's translational, rotational, and torsional degrees of freedom to find the optimal binding pose. Systematic algorithms (e.g., incremental construction in FlexX) explore all possibilities but are computationally expensive. Stochastic algorithms (e.g., Genetic Algorithms in GOLD, Monte Carlo in Glide) sample the energy landscape randomly, balancing speed and thoroughness. The search must be sufficiently exhaustive to avoid missing the true binding mode while remaining fast enough to screen millions of compounds.

A mathematical approximation of the binding free energy (ΔG) used to rank predicted poses and separate true binders from non-binders. Force-field-based functions (e.g., DOCK, AutoDock) calculate van der Waals and electrostatic interactions. Empirical functions (e.g., ChemScore, GlideScore) sum weighted terms for hydrogen bonds, hydrophobic contacts, and entropic penalties, trained on known protein-ligand complexes. Knowledge-based potentials (e.g., PMF, DrugScore) derive statistical preferences from observed atom-pair distances in structural databases. The accuracy of the scoring function remains the primary bottleneck in docking.

How the protein target is modeled critically impacts docking success. Rigid receptor docking treats the protein as a fixed structure, which is fast but fails to capture conformational changes upon binding. Soft docking allows slight steric overlap by reducing van der Waals radii. Side-chain flexibility permits rotation of selected residue side chains (e.g., in GOLD or AutoDock). Ensemble docking docks ligands into multiple experimentally determined or computationally generated receptor conformations, accounting for full protein flexibility and enabling the discovery of cryptic binding pockets.

Docking is applied in two distinct modes with different requirements. Pose prediction (binding mode prediction) aims to reproduce the experimentally observed binding geometry of a known ligand, evaluated by RMSD (Root Mean Square Deviation) from the crystal structure. Success is typically defined as RMSD < 2.0 Å. Virtual screening aims to enrich a small set of true binders from a large decoy library, evaluated by enrichment metrics. A docking protocol optimized for pose prediction does not necessarily perform well for virtual screening, and vice versa.

A specialized docking mode for irreversible inhibitors that form a covalent bond with a specific nucleophilic residue (e.g., cysteine, serine, lysine). The algorithm must simultaneously optimize the non-covalent binding pose and the geometry of the bond-forming reaction. Tools like CovDock (Schrödinger), DOCKovalent, and AutoDock4 with custom reactive force fields model the attack angle, distance, and transition state of the warhead. This is critical for designing targeted covalent inhibitors (TCIs) for kinases like EGFR and BTK.

A strategy to mitigate the inaccuracies of individual scoring functions by combining predictions from multiple orthogonal functions. Consensus scoring ranks compounds by the intersection of top-ranked poses from different scoring functions (e.g., requiring a compound to score well in both ChemScore and GoldScore). Rescoring involves re-evaluating top docked poses with a more computationally expensive and theoretically rigorous method, such as MM-GBSA (Molecular Mechanics Generalized Born Surface Area) or MM-PBSA, which use implicit solvent models to provide more accurate relative binding energy estimates.

The mathematical heart of docking that estimates binding affinity. Force-field-based functions calculate van der Waals and electrostatic energies, empirical functions sum weighted terms like hydrogen bonds and hydrophobic contacts, and knowledge-based potentials derive from statistical analyses of protein-ligand complexes. Modern consensus scoring combines multiple functions to improve accuracy. The key limitation: accurately approximating entropy and desolvation remains an open challenge.

A deep learning methodology that accelerates structure-based virtual screening of billion-scale libraries. A neural network is trained on docking scores from a small, diverse subset of the library, then predicts scores for the remaining compounds. Only top-predicted molecules proceed to full docking. This iterative active learning loop can reduce computational cost by 100-fold while recovering over 90% of top-scoring hits, making ultra-large virtual screening feasible on standard compute clusters.

Treating the receptor as rigid is the primary simplification in docking. Real proteins undergo side-chain rotations, loop movements, and domain shifts upon binding. Advanced approaches include:

• Ensemble docking: docking against multiple pre-generated receptor conformations
• Induced-fit docking: allowing active site residue side chains to move during ligand placement
• Molecular dynamics-refined docking: using short MD simulations to relax the complex Ignoring flexibility can miss cryptic pockets that only appear upon ligand binding.

Specialized docking for designing irreversible inhibitors that form a covalent bond with a target residue. The algorithm must simultaneously optimize non-covalent pose and bond geometry—the ligand's reactive warhead must be positioned within bonding distance and angle of the nucleophilic amino acid (typically cysteine). Tools like CovDock and DOCKTITE incorporate geometric constraints and custom scoring terms for bond formation energy. Critical for targeted covalent inhibitors in oncology.

The revolution in protein structure prediction has transformed docking. AlphaFold2 and RoseTTAFold now provide high-confidence receptor models for targets lacking experimental structures. However, predicted structures are typically apo-form (unbound), requiring caution: active site side chains may be mispositioned, and binding pockets may be partially collapsed. Best practice involves model quality assessment (pLDDT scores), using multiple predicted conformations, and applying relaxation protocols before docking.

The gold-standard computational method for predicting relative binding free energies between similar ligands. FEP uses alchemical molecular dynamics simulations to gradually transform one ligand into another within the binding site, computing the free energy change along the path. While far more accurate than docking scores (errors ~1 kcal/mol), it is computationally expensive and limited to congeneric series. Often used in lead optimization to rank a small set of analogs after docking has identified the core scaffold.