Binding Pocket Detection

A druggable pocket must possess sufficient depth and enclosure to sterically complement a small molecule. Shallow, flat, or solvent-exposed grooves typically lack the necessary shape complementarity to achieve high-affinity binding. Detection algorithms compute metrics such as the enclosure score—the fraction of a pocket's surface area shielded from bulk solvent—and the depth index, which measures the distance from the pocket's centroid to the nearest bulk solvent boundary. Deep, well-enclosed pockets with a depth greater than 7–10 Å and an enclosure score above 0.5 are strongly correlated with druggability.

• Pocket depth: Minimum 7–10 Å for viable drug targets
• Enclosure score: Fraction of pocket surface occluded from solvent
• Solvent-accessible surface area (SASA): Typically 250–500 Ų for drug-like ligands
• Shape complementarity: Steric fit between pocket contours and ligand topology

The hydrophobic character of a binding pocket is a primary driver of ligand binding affinity. Non-polar desolvation—the displacement of ordered water molecules from hydrophobic surfaces upon ligand binding—contributes a significant entropic gain. Druggable pockets exhibit a high proportion of apolar surface area, typically exceeding 60% of the total pocket surface. Computational tools like Fpocket and SiteMap quantify the hydrophobic ratio by classifying pocket atoms as hydrophobic (aliphatic and aromatic carbons) or polar (hydrogen bond donors and acceptors).

• Apolar surface ratio: >60% hydrophobic for highly druggable pockets
• Entropic desolvation: Release of ordered water molecules drives binding
• Hot spots: Clusters of hydrophobic residues (Leu, Ile, Val, Phe, Trp)
• Contrast with polar patches: Balanced by strategic hydrogen bond sites

While hydrophobicity drives binding, a pocket devoid of polar functionality lacks specificity. Druggable pockets contain a balanced distribution of hydrogen bond donors (backbone NH, Ser/Thr/Tyr OH, Lys/Arg side chains) and acceptors (backbone C=O, Asp/Glu carboxylates, His imidazole) positioned to form directional interactions with ligand functional groups. Detection algorithms map the spatial distribution of these features and compute a polar atom density—typically 2–5 hydrogen bond sites within a 5 Å radius of the pocket center. An optimal pocket provides 3–5 hydrogen bonding opportunities without excessive polarity that would impose a high desolvation penalty.

• Donor/acceptor ratio: Balanced distribution for specificity
• Polar atom density: 2–5 H-bond sites within 5 Å of pocket centroid
• Buried polar groups: Unsatisfied donors/acceptors indicate ligandable sites
• Water-mediated interactions: Structural water molecules bridging protein-ligand contacts

The pocket volume must be commensurate with the size of a drug-like small molecule. Most orally bioavailable drugs occupy a molecular volume between 160 and 480 ų, corresponding to a molecular weight of 200–500 Da. Druggable pockets typically exhibit volumes in the range of 200–800 ų, with an optimal range of 300–500 ų for accommodating lead-like compounds. Detection algorithms such as LIGSITE, PASS, and DeepSite compute pocket volume using grid-based or alpha-shape methods. Excessively large pockets (>1000 ų) may indicate protein-protein interaction interfaces that are challenging to target with small molecules, while volumes below 150 ų are generally too small to accommodate drug-like ligands.

• Optimal volume: 300–500 ų for lead-like compounds
• Drug-like volume range: 160–480 ų (Lipinski-compliant molecules)
• Volume calculation: Grid-based (LIGSITE) or alpha-shape (Fpocket) methods
• Pocket flexibility: Induced-fit effects can expand apparent volume upon binding

Modern pocket detection integrates multiple physicochemical features into a unified druggability score using machine learning classifiers. Algorithms such as DrugPred, PockDrug, and DeepDrug train on datasets of known druggable and non-druggable pockets (derived from the PDBbind and sc-PDB databases) to predict the likelihood that a cavity can bind a drug-like molecule with high affinity. These classifiers input features including pocket volume, hydrophobicity, enclosure, hydrogen bond density, and residue composition, outputting a probability score (0–1). A score above 0.5 typically indicates a druggable pocket, while scores above 0.8 suggest a highly tractable target suitable for lead optimization.

• DrugPred: Random forest classifier trained on PDBbind pocket features
• PockDrug: Combines pocket geometry with residue-level physicochemical properties
• DeepDrug: Deep neural network using 3D convolutional filters on pocket voxel grids
• Druggability threshold: Score >0.5 indicates druggable; >0.8 highly druggable
• Training data: sc-PDB (druggable) and non-redundant PDB (non-druggable) sets

Druggable pockets often correspond to evolutionarily conserved regions within a protein family, particularly at functional sites such as enzyme active sites or allosteric regulatory clefts. Conservation analysis using ConSurf or Rate4Site assigns an evolutionary conservation score to each residue based on multiple sequence alignments. Highly conserved pocket residues indicate functional importance and structural constraints that make the pocket a reliable drug target. Conversely, pockets in hypervariable regions may lead to rapid resistance mutations. Detection algorithms can weight pocket druggability by the conservation ratio—the fraction of pocket-lining residues with high conservation scores—to prioritize targets less prone to mutational escape.

• ConSurf: Bayesian method for residue conservation scoring
• Functional sites: Active sites and allosteric pockets show high conservation
• Resistance prediction: Low conservation correlates with mutational escape potential
• Conservation ratio: Fraction of pocket residues with conservation score >7 (scale 1–9)
• Orthosteric vs. allosteric: Orthosteric pockets typically more conserved than allosteric

A deep convolutional neural network approach that treats the protein structure as a 3D image, classifying each voxel as part of a binding pocket or not. It learns spatial patterns directly from the Protein Data Bank (PDB) without hand-crafted geometric rules.

• Operates on a 3D grid representation of the protein at 1 Å resolution
• Trained on the sc-PDB dataset of druggable binding sites
• Available through the PlayMolecule web platform for accessibility
• Represents an early and influential application of deep learning to pocket detection

A commercial tool that identifies binding sites by computing interaction energy maps between the protein and a small-molecule probe. It evaluates sites based on enclosure, hydrophilicity, and site score to rank druggability.

• Distinguishes between orthosteric and allosteric pockets
• Provides a SiteScore and Dscore for quantitative druggability assessment
• Integrates with the full Schrödinger drug discovery suite
• Widely used in industrial medicinal chemistry campaigns

A deep learning method that combines a 3D voxelized representation of the protein surface with a ResNet architecture to predict binding pockets. It processes surface patches rather than the full volumetric protein, improving computational efficiency.

• Uses surface-based representation to reduce computational cost vs. full-grid methods
• Achieves competitive performance on the sc-PDB and COACH420 benchmarks
• Outputs per-residue binding propensity scores
• Demonstrates the advantage of surface-centric over volume-centric representations

A 3D U-Net architecture designed for end-to-end binding site detection using a distance transform as the target label. Instead of binary classification, it predicts the continuous distance to the nearest pocket center, providing richer spatial information.

• Uses a signed distance function as the training target for smoother gradients
• Outperforms binary classification approaches on the COACH420 dataset
• Capable of detecting multiple pockets simultaneously in a single forward pass
• Represents a modern, segmentation-based approach to the pocket detection problem