Binding pocket prediction: P2Rank vs fpocket
Maintained by the Talindrew team · Last updated
Binding pocket prediction finds the cavities on a protein structure where a small molecule is likely to bind, so docking can be aimed at them. P2Rank scores points on the protein surface with a machine-learned model trained on known ligand sites and clusters the high-scoring points into ranked pockets; fpocket finds cavities geometrically with Voronoi alpha spheres and ranks them by a druggability score. P2Rank is generally the more accurate ranker of true ligand sites, while fpocket gives detailed geometric descriptors; both run in seconds.
How do P2Rank and fpocket work?
P2Rank places points on the solvent-accessible surface, describes each by its local chemistry and geometry, predicts its ligandability with a random forest, and clusters ligandable points into pockets ranked by score. fpocket fills the protein with alpha spheres from a Voronoi tessellation, clusters them into cavities, and reports volume, hydrophobicity and a druggability score per pocket.
How do I choose which pocket to dock into?
The top-ranked pocket is a starting point, not an answer.
- Prefer a pocket that matches known biology: an active site, an allosteric site from the literature, or where a co-crystallised ligand sits in a homolog.
- Check the structure's confidence around the pocket; on a predicted structure, low pLDDT loops lining a pocket make it unreliable.
- Avoid pockets formed between crystal contacts or at the cut ends of a truncated construct.
- If two pockets look plausible, dock into both and compare.
P2Rank and fpocket compared
| P2Rank | fpocket | |
|---|---|---|
| Method | Machine-learned surface point scoring, clustered | Geometric: Voronoi alpha spheres |
| Ranks by | Predicted ligandability | Druggability score |
| Strength | Accurate ranking of true ligand sites | Rich geometric descriptors per pocket |
| Licence | MIT | MIT |
Frequently asked questions
Does pocket prediction work on AlphaFold or ESMFold structures?
Yes, but predicted side chains and loops around a pocket can be wrong or closed, and models predicted without a ligand often show tighter pockets than the bound state. Weight pockets by local confidence, or use a co-folding model such as Boltz-2 with a ligand present.
References
- Krivák R, Hoksza D. P2Rank: machine learning based tool for rapid and accurate prediction of ligand binding sites from protein structure. J Cheminform 10: 39 (2018). https://doi.org/10.1186/s13321-018-0285-8
- Le Guilloux V, Schmidtke P, Tuffery P. Fpocket: an open source platform for ligand pocket detection. BMC Bioinformatics 10: 168 (2009). https://doi.org/10.1186/1471-2105-10-168
Related
Design your Binding pocket prediction: P2Rank vs fpocket construct in Talindrew.
Free to use. No download required. Works in any modern browser.