Drug-Target Interaction (DTI)

The structural and chemical shape match between a drug and its binding pocket. This is governed by steric fit, where the ligand's 3D conformation occupies the target cavity without atomic clashes, and electrostatic complementarity, where positive and negative charge distributions align. Hydrogen bond donors on the drug must pair with acceptors on the protein, and vice versa. Hydrophobic patches on the ligand seek out non-polar residues in the pocket to maximize van der Waals contacts while displacing ordered water molecules for an entropic gain.

The quantitative strength of the interaction, expressed as the dissociation constant (Kd) or inhibition constant (Ki). A lower Kd indicates tighter binding. Affinity is a thermodynamic parameter derived from the Gibbs free energy change (ΔG = -RT ln Kd), which is the sum of enthalpic contributions (hydrogen bonds, ionic interactions, van der Waals forces) and entropic contributions (desolvation, conformational restriction). High-affinity drugs often achieve Kd values in the nanomolar (nM) to picomolar (pM) range.

Specificity is the absolute discrimination of a drug for a single molecular target. Selectivity is the relative preference for one target over others, often expressed as a fold-difference in affinity. A drug's therapeutic window is defined by its selectivity profile. Off-target binding to related proteins—such as other kinases in the kinome or related GPCRs—can cause toxicity. Polypharmacology is the intentional design of a drug to hit multiple targets for complex diseases like cancer or neuropsychiatric disorders.

The temporal dynamics of the interaction, defined by the association rate constant (kon) and dissociation rate constant (koff). The ratio koff/kon equals Kd. A drug's residence time (1/koff) measures how long the complex remains intact. Drugs with long residence times can maintain pharmacological effect even after plasma clearance. Slow-binding inhibitors exhibit time-dependent inhibition, often due to conformational changes in the target's binding pocket upon initial encounter.

Most drugs bind via non-covalent interactions (hydrogen bonds, ionic bonds, hydrophobic effects) and are fully reversible. Covalent inhibitors form a permanent chemical bond with a specific nucleophilic residue, typically a cysteine, lysine, or serine. These drugs exhibit a two-step mechanism: initial non-covalent recognition followed by bond formation. Covalent binding permanently inactivates the target, requiring new protein synthesis for recovery. Covalent docking algorithms must model both the non-covalent pose and the bond-forming step.

Both drug and target are flexible entities. Conformational sampling explores the low-energy 3D shapes a ligand can adopt. The induced-fit model describes how a protein's binding pocket reshapes upon ligand binding. Alternatively, conformational selection posits that the protein pre-exists in multiple states, and the drug stabilizes one. Understanding these dynamics is critical for docking accuracy. Rigid-receptor docking often fails when the apo structure differs significantly from the holo conformation.

The quantitative measure of interaction strength between a drug and its target, typically expressed as Kd (dissociation constant), Ki (inhibition constant), or IC50. Lower values indicate stronger binding. Modern deep learning models like DeepDTA and GraphDTA directly predict these continuous values from raw sequence and structure data, bypassing traditional free energy calculations.

A structure-based computational method that predicts the preferred orientation of a ligand when bound to a target protein. Key paradigms include:

• Rigid-body docking: Both protein and ligand are treated as fixed
• Induced-fit docking: Protein side chains adapt to the ligand
• Covalent docking: Models permanent bond formation with residues like cysteine Recent geometric deep learning models like EquiBind and DiffDock perform blind docking in a single forward pass without exhaustive conformational sampling.

A mathematical function that approximates the binding free energy (ΔG) of a protein-ligand pose. Scoring functions fall into three classes:

• Force-field based: Compute van der Waals and electrostatic energies
• Empirical: Weighted sum of terms like hydrogen bonds and hydrophobic contacts
• Knowledge-based: Statistical potentials derived from known protein-ligand complexes Machine learning scoring functions, such as RF-Score and OnionNet, now outperform classical functions by learning non-linear interaction patterns.

The design or serendipitous property of a drug interacting with multiple distinct targets simultaneously. This can produce therapeutic synergy (e.g., multi-kinase inhibitors) or dangerous off-target effects. Computational methods like proteochemometric modeling and target fishing use matrix factorization and deep learning to predict the full interaction profile of a compound across the entire proteome.

The computational filtering of millions to billions of compounds to identify those most likely to bind a target. Two primary approaches exist:

• Structure-based (SBVS): Docks compounds into a known 3D target structure
• Ligand-based (LBVS): Searches for compounds similar to known actives using pharmacophore or shape queries Performance is measured by the enrichment factor, which quantifies how effectively active compounds are concentrated in the top-ranked fraction of a screened library.

A rigorous alchemical free energy method that computes relative binding free energy (ΔΔG) between two related ligands through a non-physical thermodynamic cycle. FEP requires extensive molecular dynamics sampling but achieves chemical accuracy (~1 kcal/mol). It is the gold standard for lead optimization, though neural network potentials are now accelerating these calculations by replacing expensive quantum mechanical energy evaluations.