GPTomics/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

| Question / situation | Method | Best when | Fails / misleads when | |---|---|---|---| | Enumerate every geometric concavity, fast | fpocket (alpha-sphere/Voronoi) or CASTp (surface topology) | Apo or holo, whole-surface scan, ranked candidate list | Reports additive sites, crystal clefts, and non-functional pockets as pockets; ranking != function | | Rank pockets by ligandability / druggability | P2Rank (ML probability), fpocket druggability score, DoGSiteScorer (SVM) | Prioritizing among many cavities on a folded globular domain | Trained on holo/druggable sets -> under-scores apo, shallow, allosteric, and PPI sites | | Detect a cryptic or transient pocket | mdpocket over an MD trajectory or conformational ensemble | Site opens only on binding / not present in a single apo snapshot | A single static structure - a cryptic site is invisible without conformational sampling | | A ligand is already bound (holo structure) | interface-analysis (residues around the bound ligand) | The real binding site is known - map it, do not re-detect | Running de-novo detection when the answer is already in the file | | Surface area / volume of a defined pocket | CASTp (analytic surface + volume) or fpocket descriptors | Quantifying a known pocket's size | Treating a large computed cavity as evidence of function |

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

| Structure state | What a detected pocket reveals | What it cannot reveal | |---|---|---| | Holo (ligand bound) | The real, occupied site - highest reliability; map it directly | Whether other detected pockets are functional (still hypotheses) | | Apo experimental | Open, ligand-accessible pockets present in THIS conformation | Cryptic sites closed in this snapshot; induced-fit geometry | | Predicted model (AlphaFold/ESMFold) | Candidate fold and gross concavities | Reliable pocket rotamers (least reliable atoms), cofactor/metal-shaped sites; validate first | | MD / NMR ensemble | Transient and cryptic pockets, pocket dynamics/persistence | Which opening is biologically relevant (do not over-read rare openings) |

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.

Cryptic pockets over an ensemble

Goal: Find a site that is closed in the single apo structure but opens on binding.

Approach: A cryptic pocket does not exist in one static snapshot, so run pocket detection over a conformational ENSEMBLE - an MD trajectory, an NMR ensemble, or frames from enhanced sampling that is designed to open cryptic sites (mixed-solvent MD / SWISH, accelerated MD) - and track pocket persistence and opening frequency. Generating that ensemble is the real upstream step; mdpocket (Schmidtke 2011 Bioinformatics 27:3276) then analyzes it, building a pocket-frequency grid across frames, and the -S flag adds a per-snapshot drug score so a druggable pocket that only appears transiently is visible.

import subprocess

def run_mdpocket(topology_pdb, trajectory, fmt='xtc'):
    # -S adds the per-snapshot drug score so a transiently druggable pocket is flagged
    subprocess.run(['mdpocket', '--trajectory_file', trajectory,
                    '--trajectory_format', fmt, '-f', topology_pdb, '-S'], check=True)

Do not over-read a pocket that opens in a handful of frames - a rare, short-lived opening is a weaker hypothesis than a persistent one, and enhanced sampling can manufacture openings that are not physiologically populated. Once a cryptic site is chosen, the bound-state geometry (not the closed apo structure) is what a docking run needs.

Once a site is chosen

A detected, ranked pocket is the INPUT to structure-based drug discovery, not the endpoint. Hand the chosen pocket (its center and lining residues) to chemoinformatics/virtual-screening to dock a library into it, and to chemoinformatics/ml-docking-rescoring to rescore poses - both of those presuppose a known site, which is exactly what this skill produces. Carry the caveat forward: docking into an apo or predicted pocket with unreliable rotamers gives confidently wrong poses unless the pocket is prepared and, ideally, cross-checked against a holo conformation.

Common Errors

| Symptom | Cause | Fix | |---|---|---| | Top-ranked pocket is a surface additive site | Geometry finds all concavities, including where a sulfate/glycerol/PEG sat | Cross-check pocket atoms against HETATM/additives in the input; strip additives first (structure-modification) | | A known binding site is missed entirely | Site is cryptic - closed in the single apo snapshot | Run mdpocket over an MD/NMR ensemble, not one static structure | | Druggability scores all low on a real target | Score trained on holo/druggable sets under-scores apo, shallow, allosteric, PPI sites | Low score is not "no site"; corroborate with an independent method (P2Rank/DoGSiteScorer), evolutionary conservation of the lining residues (ConSurf), hot-spot / mixed-solvent mapping (FTMap), homolog holo structures, and known biology | | fpocket finds no _info.txt | Parsed the wrong directory or fpocket failed silently | fpocket writes <stem>_out/ beside the input; check the run completed and the stem matches | | P2Rank parse gives KeyError on column | The CSV has a space after each comma so header keys carry a leading space ( probability) | Use csv.DictReader(skipinitialspace=True) and strip keys before indexing | | Pockets differ between apo and holo of the same protein | Induced fit - the pocket reshapes on binding | Expected; detect on the state that matches the question, do not average them | | "Confident" AlphaFold pocket docks poorly | Pocket-lining rotamers are the least reliable atoms; model is apo-like | Validate/prepare the model (structure-validation); prefer a holo experimental structure | | Same pocket, different rank from two tools | fpocket geometry vs P2Rank ML weight features differently | Agreement raises confidence; treat disagreement as a flag to inspect, not a tie-break | | Reported "N druggable pockets" as a finding | Ranking treated as function | State it as a ranked hypothesis list; a score is not a validated site | | mdpocket pocket appears in few frames trusted as real | Over-reading a rare, short-lived opening | Weight by opening frequency/persistence; rare openings are weak hypotheses | | Detection on the asymmetric unit misses an interface pocket | Functional pocket sits across a symmetry-generated interface | Detect on the biological assembly (structure-io) when the site may be inter-chain |

Related Skills

• structural-biology/interface-analysis - map residues around an ALREADY-BOUND ligand (holo); the complement of de-novo detection
• structural-biology/structure-validation - judge whether a structure or predicted model is trustworthy before detecting on it
• structural-biology/structure-preparation - clean, protonate, and fix the structure before pocket detection (especially a predicted model)
• structural-biology/structure-modification - strip crystallization additives and resolve altlocs before pocket detection
• structural-biology/alphafold-predictions - the apo-pocket and unreliable-rotamer caveats for detecting on a predicted model
• structural-biology/structure-navigation - select chains and HETATM ligands to cross-check pockets against additives
• structural-biology/structure-io - download the biological assembly (not just the asymmetric unit) for inter-chain pockets
• chemoinformatics/virtual-screening - dock a library into a chosen pocket (presupposes a known site)
• chemoinformatics/ml-docking-rescoring - rescore docked poses in the detected site

References

• Le Guilloux V, Schmidtke P, Tuffery P (2009) Fpocket: an open source platform for ligand pocket detection. BMC Bioinformatics 10:168. (fpocket; Voronoi tessellation and alpha spheres)
• Schmidtke P, Barril X (2010) Understanding and predicting druggability. A high-throughput method for detection of drug binding sites. J Med Chem 53(15):5858-5867. (fpocket druggability score; trained on druggable/undruggable cavities)
• Krivak R, Hoksza D (2018) P2Rank: machine learning based tool for rapid and accurate prediction of ligand binding sites from protein structure. J Cheminform 10:39. (P2Rank; template-free ML ligandability)
• Tian W, Chen C, Lei X, Zhao J, Liang J (2018) CASTp 3.0: computed atlas of surface topography of proteins. Nucleic Acids Res 46(W1):W363-W367. (surface pockets, cavities, channels; area and volume)
• Volkamer A, Kuhn D, Grombacher T, Rippmann F, Rarey M (2012) Combining global and local measures for structure-based druggability predictions. J Chem Inf Model 52(2):360-372. (DoGSiteScorer; SVM druggability)
• Cimermancic P, et al. (2016) CryptoSite: expanding the druggable proteome by characterization and prediction of cryptic binding sites. J Mol Biol 428(4):709-719. (cryptic site = pocket in holo but not apo)
• Schmidtke P, Bidon-Chanal A, Luque FJ, Barril X (2011) MDpocket: open-source cavity detection and characterization on molecular dynamics trajectories. Bioinformatics 27(23):3276-3285. (pocket tracking over an ensemble; cryptic/transient pockets)