Bio Structural Biology Binding Site Detection

Version Compatibility

Reference examples tested with: fpocket 4.1+, P2Rank 2.4+, biopython 1.83+, numpy 1.26+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• CLI: <tool> --version then <tool> --help to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Binding Site Detection

"Find the pockets / druggable cavities on my apo structure" -> Enumerate surface concavities de novo and rank them by a ligandability or druggability score.

• CLI: fpocket -f protein.pdb (alpha-sphere / Voronoi) or prank predict -f protein.pdb (ML ligandability)

"Where could a ligand bind, and how druggable is each site?" -> Detect candidate pockets, then score each for the likelihood of binding a drug-like molecule.

• CLI: fpocket druggability score (per pocket) or P2Rank probability; CASTp / DoGSiteScorer web servers for surface area and SVM druggability

"Detect a cryptic pocket that is not open in this single snapshot" -> Track pocket opening across a conformational ensemble, not one static file.

• CLI: mdpocket --trajectory_file traj.xtc --trajectory_format xtc -f topology.pdb over an MD trajectory or NMR ensemble

Governing Principle

A geometric cavity is NOT automatically a functional or druggable site. Pure geometry - fpocket alpha-spheres (Le Guilloux 2009 BMC Bioinformatics 10:168), CASTp surface topology (Tian 2018 Nucleic Acids Res 46:W363) - finds ALL concavities on the surface, including crystallization-additive sites (a sulfate, glycerol, or PEG cleft), non-functional surface pockets, and inter-domain crevices. A ranked pocket list is therefore a HYPOTHESIS, not a binding site. The single most reliable site is one that is already occupied: if the structure is holo (a ligand is bound), map the residues around that ligand directly - that is interface-analysis, not de-novo detection.

Druggability and ligandability SCORES (the fpocket druggability score, P2Rank probability, DoGSiteScorer SVM) rank pockets, but the ranking is not function and not a guarantee. These models were trained largely on holo, druggable reference sets (Schmidtke 2010 J Med Chem 53:5858; Volkamer 2012 J Chem Inf Model 52:360), so they systematically UNDER-detect and under-score apo, shallow, allosteric, and protein-protein-interface pockets, and they cannot see a CRYPTIC pocket at all - a cryptic site forms a pocket in the holo structure but is closed in the apo structure (Cimermancic 2016 J Mol Biol 428:709), so it only appears across an MD or conformational ensemble, never in one static apo snapshot.

Detecting on a PREDICTED model (AlphaFold, ESMFold) inherits that model's pocket-conformation unreliability. The model is usually an apo-like ground state that carries no ligand, cofactor, or metal that would shape the real site, and the side-chain rotamers lining a pocket are the LEAST reliable atoms even when the backbone pLDDT is high (see alphafold-predictions) - so a "confident" backbone can present a confidently wrong pocket. Validate and prepare the model first (structure-validation, structure-preparation). Whatever the input, state the method used and remember: ranking is not function, and the top pocket is a starting hypothesis to corroborate. The orthogonal corroboration signals are an independent detector, evolutionary conservation of the lining residues (a functional site is usually conserved, an additive/crystal cleft usually is not), hot-spot or mixed-solvent mapping (FTMap, mixed-solvent MD), a homolog holo structure, and known biology - not the geometric rank alone.

Decision: method by question

The one-line rule: geometry enumerates concavities, scores rank them, an ensemble reveals cryptic ones, and a bound ligand settles it. Reach for the cheapest method that answers the actual question.

Decision: structure state and what a pocket result means

Apo detection answers "what could open here in this state", holo answers "what is actually bound", a predicted model answers "what might a pocket look like in the ground state", and an ensemble answers "what transiently opens". Match the claim to the state.

Detect and rank pockets with fpocket

Goal: Enumerate candidate pockets on an apo structure and rank them by druggability, keeping the result explicitly a hypothesis list.

Approach: Run fpocket, which writes a <stem>_out/ directory beside the input containing a <stem>_info.txt descriptor file and per-pocket pockets/pocketN_atm.pdb files. Parse the info file into per-pocket descriptors and sort by the druggability score. The 0.5 druggability cutoff is a rule-of-thumb triage line, not a law (Schmidtke 2010): above it a drug-like molecule is plausible, below it the pocket is likely undruggable by conventional small molecules, but apo and cryptic sites routinely fall below it while still being real.

import subprocess
from pathlib import Path

def run_fpocket(pdb_path):
    subprocess.run(['fpocket', '-f', pdb_path], check=True)
    stem = Path(pdb_path).stem
    return Path(pdb_path).with_name(f'{stem}_out')  # fpocket writes <stem>_out/ beside the input

def parse_fpocket_info(out_dir):
    info = next(Path(out_dir).glob('*_info.txt'))
    pockets, current = [], None
    for raw in info.read_text().splitlines():
        line = raw.strip()
        if line.startswith('Pocket'):
            current = {'pocket': int(line.split()[1])}
            pockets.append(current)
        elif ':' in line and current is not None:
            key, val = (part.strip() for part in line.split(':', 1))
            current[key] = float(val)
    return pockets

out_dir = run_fpocket('protein.pdb')
pockets = sorted(parse_fpocket_info(out_dir),
                 key=lambda p: p['Druggability Score'], reverse=True)
druggable = [p for p in pockets if p['Druggability Score'] >= 0.5]  # 0.5 = triage, not a law
for p in pockets[:5]:
    print(f"pocket {p['pocket']}: drug={p['Druggability Score']:.2f} score={p['Score']:.2f}")

The pocket Score ranks how likely the cavity binds a small molecule; the Druggability Score (0-1) estimates likelihood of binding a DRUG-LIKE molecule specifically. A high-scoring pocket that coincides with a bound additive in the deposited file is a crystallization site, not a target - cross-check the pocket atoms against any HETATM in the input (structure-navigation).

Score ligandability with P2Rank

Goal: Get a template-free, machine-learned ligandability ranking that does not depend on alpha-sphere geometry alone.

Approach: Run P2Rank, which scores points on the solvent-accessible surface and clusters them into ranked pockets, writing <inputname>_predictions.csv. Parse the calibrated probability per pocket. P2Rank is ML-based and template-free (Krivak 2018 J Cheminform 10:39), so it complements fpocket's pure geometry - agreement between the two on the same top pocket raises confidence; disagreement flags a site to scrutinize.

import csv
import subprocess
from pathlib import Path

def run_p2rank(pdb_path, out_dir='p2rank_out'):
    subprocess.run(['prank', 'predict', '-f', pdb_path, '-o', out_dir], check=True)
    return Path(out_dir) / f'{Path(pdb_path).name}_predictions.csv'  # keeps the input extension

def parse_p2rank(csv_path):
    with open(csv_path) as fh:
        reader = csv.DictReader(fh, skipinitialspace=True)  # P2Rank pads columns with spaces
        rows = [{k.strip(): v.strip() for k, v in row.items()} for row in reader]
    return [{'rank': int(r['rank']), 'score': float(r['score']),
             'probability': float(r['probability'])} for r in rows]

predictions = parse_p2rank(run_p2rank('protein.pdb'))
top = predictions[0]  # rows are already ordered by rank
print(f"top pocket: probability={top['probability']:.2f} score={top['score']:.2f}")

P2Rank ranks pockets even when every probability is modest; a low top probability on an apo or PPI target is expected, not a failure to find a site.

Surface topology and SVM druggability servers

CASTp 3.0 (Tian 2018 Nucleic Acids Res 46:W363, http://sts.bioe.uic.edu/castp/) is a web server that delineates pockets, interior cavities, and channels analytically and reports each one's molecular surface AREA and VOLUME - use it when the question is pocket geometry (size, mouth, buried volume) rather than a druggability call. DoGSiteScorer, on the ProteinsPlus server (Volkamer 2012 J Chem Inf Model 52:360, https://proteins.plus/), predicts pockets by a difference-of-Gaussians grid and returns an SVM druggability score trained on a druggable/undruggable set - a second, independent druggability opinion to compare against fpocket. Both are servers with no Python binding; submit a structure and parse the returned table. Treat a server druggability number the same way as fpocket's: trained on holo/druggable data, so it under-scores apo and cryptic sites.