Binding Site Analysis

Binding Site Analysis

The step before docking. Every docking skill in this bundle assumes you already know where the ligand goes and that the site is worth the compute — this is where those two assumptions get checked. fpocket runs in seconds and can save a month of screening against a pocket that was never going to bind anything.

Tool: fpocket, MIT, conda install -c conda-forge fpocket. Alpha-sphere cavity detection by Voronoi tessellation. Checked against: fpocket 4.x output format.

Read [references/fpocket-output.md](references/fpocket-output.md) before parsing a run, [references/druggability.md](references/druggability.md) before calling a site druggable or not, and [references/cryptic-and-allosteric.md](references/cryptic-and-allosteric.md) when the answer is "no pocket" — that one is judgement, not syntax.

The three scripts

| Script | Answers | |---|---| | pocket_report.py | Which cavities are there, and is any of them worth targeting? | | pocket_box.py | Where exactly does the docking box go? | | site_compare.py | Does a pocket appear only when something is bound? |

Score and Druggability Score are different, and pocket1 is not the answer

This is the thing to get right. fpocket reports two numbers per cavity and they measure different things. Score ranks cavities geometrically, and pocket numbering follows it. Druggability Score is a logistic model trained to separate sites with known drug-like ligands from sites without — it is the one that answers "worth a campaign".

They disagree often:

python skills/binding-site-analysis/scripts/pocket_report.py rank --out-dir receptor_out

# fpocket ranks pocket 1 first by Score, but pocket 2 is the most druggable.
pocket  druggability  score  volume  apolar_fraction  verdict     reason
2       0.871         0.31   720.5   0.7143           druggable   resembles sites with known drug-like ligands
1       0.183         0.412  980.4   0.3172           poor        does not resemble a small-molecule binding site

Pocket 1 is larger and scores higher. It is also 68% polar surface, which is a groove rather than a pocket. Volume alone is misleading — the apolar fraction is what distinguishes a site that will bind a small molecule, and this script derives it because fpocket does not.

Thresholds applied: druggability ≥ 0.5 is druggable, 0.2–0.5 marginal; volume receptor.pdb && fpocket -f receptor.pdb

A structural metal or covalent prosthetic group should stay; a substrate analogue should go.

## Producing the box

```bash
python skills/binding-site-analysis/scripts/pocket_box.py from-pocket --out-dir receptor_out \
    --pocket 2 --format vina

center_x = 12.0
center_y = 22.0
center_z = 33.0
size_x = 12.0
size_y = 12.0
size_z = 14.0

That output pastes directly into an AutoDock Vina config. Two sizing rules are built in: 4 Šof padding per side, so the ligand can translate and rotate rather than being pinned; and a warning past 27 000 ų, because Vina spreads a fixed exhaustiveness over the whole volume and a box twice as wide samples eight times as thinly.

from-ligand centres on a crystallographic ligand instead, and that is the better option whenever a holo structure exists — a real bound pose beats a predicted cavity. It lists the candidate HETATM residues when the one you named is not present.

Cryptic sites, or why the apo structure lied

python skills/binding-site-analysis/scripts/site_compare.py match --apo apo_out --holo holo_out

Classifies each cavity as cryptic (in holo, absent in apo), induced fit, stable, closes on binding, or apo only. Superpose the structures first — matching is spatial, and unaligned inputs make every cavity look cryptic. The script says so when nothing matches.

This is not a corner case. The KRAS G12C switch II pocket does not exist in unliganded KRAS; thirty years of "undruggable" rested on structures that could not show it.

Four ways this misleads

1. A low druggability score means "unlike sites we have drugged before", not "impossible". The

training set predates degraders, covalent inhibitors, and most protein–protein interface drugs.

1. You scored one conformation. Apo structures under-report pockets systematically.
2. Structure quality propagates. Missing loops, uncertain rotamers above ~2.5 Å, and

AlphaFold's tendency toward closed apo-like states all change the answer. Check with uniprot-rcsb first.

1. A detector that cannot recover a known site should not be trusted on an unknown one. If a

holo structure exists, verify the top cavity contains the crystallographic ligand.

When the answer is "no druggable pocket"

That is a conclusion about conventional reversible small molecules, not about the target. In rough order of how often they work: covalent inhibition (how KRAS G12C fell), targeted degradation (degraders — a degrader needs a binding site, not an inhibitory pocket), molecular glues, cryptic sites found by mixed-solvent MD, biologics if the target is extracellular, and oligonucleotides to sidestep the protein entirely.

Composing with the rest of the bundle

• uniprot-rcsb → before: check resolution, missing residues, and whether a holo structure exists.
• This skill → autodock-vina: pocket_box.py --format vina writes its config directly.
• This skill → diffdock / boltz: which site to focus on before posing.
• molecular-dynamics → alongside: mixed-solvent simulation to find cryptic pockets.
• chemical-space → after: only worth a giga-scale library once the site is worth it.

Reporting results honestly

Give the druggability score, the volume, and the apolar fraction together — one number is not an assessment. Say which structure and which conformation was analysed, and whether waters and ligands were stripped. If a known ligand exists, say whether the detector recovered its site. Call a predicted cryptic pocket a hypothesis until a fragment soak or thermal shift confirms it.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: K-Dense-AI
• Source: K-Dense-AI/drug-discovery-agent-skills
• License: MIT
• Homepage: www.k-dense.ai

Install and usage instructions live in the source repository linked above.