PocketSense Student Guide

What is an active site?

Every protein that does something useful — an enzyme that catalyses a reaction, a receptor that binds a hormone, a transporter that moves a molecule — has a specific region on its surface (or buried inside it) where that job happens. That region is called the active site (for enzymes) or binding pocket (for any protein that binds a ligand).

An active site is typically:

• A concave depression or enclosed cavity on the protein surface
• Lined by a specific set of amino acid residues whose chemistry is tuned for binding or catalysis
• Often conserved across evolution (mutations here are usually destructive)
• Frequently occupied by a cofactor (NAD, FAD, metal ion) or a substrate

Finding the active site from a 3D structure is one of the first and most important steps in understanding a protein's function, designing drugs, or engineering enzymes.

Why is this hard?

You might think: "just find the biggest hole in the protein." But that doesn't work:

1. The exterior surface is full of bumps and grooves. A simple distance check returns the entire outer surface, not just interesting pockets.
2. Some active sites are shallow, not deep. Proteases have open active sites. Metalloproteins have a metal in a shallow depression. Algorithms that only look for deep cavities miss these.
3. Volume alone is a bad ranking metric. The biggest pocket is not always the active site. You need chemical information.
4. Biological context matters. A dimer's active site may be at the interface between two chains.

How PocketSense works

PocketSense uses a two-stage architecture inspired by P2Rank and DeepPocket:

Stage 1: Geometric detection

1. Parse the structure — reads the PDB file, separates protein from ligands/metals, builds the biological assembly if REMARK 350 records exist, and collapses NMR ensembles to a consensus.
2. Build a 3D grid — a cubic grid is placed over the protein's bounding box.
3. Flood-fill from outside — starting from the grid boundary, the algorithm spreads through solvent. Everything reachable is EXTERIOR; everything cut off is INTERIOR — an enclosed cavity. This is the key step that distinguishes real pockets from the outer surface.
4. Burial count — for each solvent point near the surface, counts how many of its 26 neighbours are protein. High burial = deeply enclosed. This catches shallow clefts that flood-fill misses.
5. DBSCAN clustering — groups candidate points into discrete pockets. Handles noise and varying pocket shapes.
6. Residue assignment — finds all protein atoms within 5 Å of any pocket boundary point (not just the centroid, which fails for elongated pockets).

Stage 2: ML reranking

1. Feature extraction — computes 20 physicochemical features per pocket: volume, sphericity, burial, hydrophobicity, polarity, charge, H-bond donors/acceptors, aromatic content, enclosure ratio, ligand/metal proximity.
2. Random Forest scoring — a 200-tree classifier takes these 20 features and outputs a probability (0–1) that the pocket is a true binding site. Trained on protein–ligand complexes from the PDB.
3. Ranking — pockets are re-ranked by ML probability. Cross-validation: ROC-AUC 0.98.

How to interpret results

Reading the druggability score

The druggability score is built from five physical properties, each weighted by how much it matters for drug binding: size (30%) — drug-sized ligands occupy roughly 150–400 ų, so pockets near 500 ų score highest; enclosure (25%) — an enclosed cavity retains the ligand better than an open cleft; depth (20%) — deeper pockets bury more of the ligand; hydrophobic character (15%) — most drugs are greasy, so a moderate hydrophobic fraction (~50%) is ideal; aromatic residues (10%) — offer π-stacking contacts for aromatic scaffolds. A site scoring High on both druggability and ML probability is the best candidate to inspect first.

Limitations

• Membrane proteins: Transmembrane regions may be falsely flagged as pockets.
• Large complexes: Structures > 1500 residues are rejected (server protection).
• Cryptic pockets: Pockets that only open on ligand binding are not detectable from static structures.
• Training data: The model was trained on ~20 structures. Future versions will use scPDB/BioLip.

References

• P2Rank — github.com/rdk/p2rank
• DeepPocket — Aggarwal et al. (2021) ACS JCIM
• PocketPicker — Weisel et al. (2007) Chem. Eur. J.
• POCASA — Yu et al. (2010) Bioinformatics
• PocketAnchor — Li et al. (2023) Cell Systems
• PocketMiner — Meller et al. (2023) Nature Communications
• fpocket — Le Guilloux et al. (2009) BMC Bioinformatics