← Về thư mục
📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / skills / bioskills / bio-structural-biology-binding-site-detection / references / usage-guide.md

Binding Site Detection

Overview

This skill finds putative ligand-binding pockets and druggable cavities de novo on a protein structure that has no bound ligand (an apo structure or a predicted model). It covers geometric pocket enumeration (fpocket alpha-spheres, CASTp surface topology), machine-learned ligandability ranking (P2Rank, DoGSiteScorer), and cryptic-pocket detection over a conformational ensemble (mdpocket). The governing idea is that a geometric cavity is a hypothesis, not automatically a functional or druggable site: pure geometry finds every concavity, including crystallization-additive sites and non-functional clefts, and druggability scores trained on holo/druggable reference sets systematically under-detect apo, shallow, allosteric, and cryptic pockets. When a ligand is already bound, mapping the residues around it (interface-analysis) is more reliable than any de-novo prediction.

Prerequisites

Quick Start

Tell the agent what you want to do: - "Find the binding pockets on my apo structure and rank them by druggability" - "Where could a ligand bind on this protein?" - "Score the druggability of the pockets in this PDB file" - "Detect cryptic pockets from my MD trajectory" - "Is the top pocket on my AlphaFold model reliable enough to dock into?" - "Run both fpocket and P2Rank and compare the top-ranked pocket"

Example Prompts

De-novo pocket detection

"I have an apo crystal structure with no ligand bound. Detect all the surface pockets and tell me which one is most likely to bind a drug-like molecule."

"Run fpocket on this PDB and give me the pockets ranked by druggability score, and flag any pocket that overlaps a crystallization additive."

Ligandability scoring and method choice

"Score the ligandability of the cavities in this structure with a machine-learning method that does not just use geometry, and compare its top pocket to fpocket's."

"This is a protein-protein interaction target and the druggability scores all come out low. Does that mean there is no site, or is the scoring under-detecting a flat/apo pocket?"

Cryptic and predicted-model cases

"The known allosteric pocket is not open in my apo structure. How do I detect a cryptic pocket from a molecular dynamics ensemble?"

"I only have an AlphaFold model. Can I trust the pocket it shows well enough to run docking, and what should I check first?"

What the Agent Will Do

  1. Establish the structure state (holo, apo, predicted model, or ensemble) and route accordingly: if a ligand is already bound, it maps the site with interface-analysis instead of detecting de novo.
  2. For a predicted model, it flags the apo-like ground state and unreliable pocket-lining rotamers and recommends validation/preparation before detection.
  3. Runs fpocket to enumerate pockets by alpha-sphere geometry and parses the <stem>_out/<stem>_info.txt descriptors, ranking by druggability score.
  4. Runs P2Rank for a template-free ML ligandability ranking and parses the _predictions.csv probabilities, comparing the top pocket against fpocket.
  5. Points to CASTp for pocket surface area/volume and DoGSiteScorer for a second SVM druggability opinion when a web-server assessment is wanted.
  6. For a cryptic or transient site, runs mdpocket over an MD trajectory or ensemble and weights pockets by opening frequency.
  7. Reports each pocket as a ranked hypothesis - stating the method and structure state - and hands the chosen site to virtual screening.

Tips