Preparing Docking Boxes from P2Rank Predictions

June 3, 2026 · View on GitHub

P2Rank predicts where a ligand is likely to bind, but it does not output a ready-made docking search box. This page shows how to derive a box (a center plus a size) for tools like AutoDock Vina, Glide, or GNINA from P2Rank output.

Quick answer

Box center: the pocket centroid (center_x, center_y, center_z in *_predictions.csv).
Box size: the bounding box of the pocket's points, expanded by a margin of at least half the longest dimension of the ligand you intend to dock.
P2Rank often marks only the hot core of a binding site, so err on the generous side and validate your box on a few complexes with known poses.

Note

All P2Rank coordinates are in Angstrom and in the same frame as the input structure: there is no transformation, so a box built from P2Rank output drops straight into a docking tool that reads the same file.

What P2Rank gives you to build a box

Source	Produced by	Contents useful for a box
`*_predictions.csv`	always	Per-pocket centroid (`center_*`), `residue_ids`, `surf_atom_ids`, `sas_points` count
`visualizations/data/{name}_points.pdb.gz`	always	SAS point coordinates (score in the B-factor column, pocket rank in the residue-sequence column)
`{name}_points.{format}`	`-export_points 1`	Same SAS points as a clean table with a `pocket` column (export-points.md)
`{name}_pocket_grid.{format}`	`-export_pocket_grid 1` (2.6+)	A 3D lattice filling the pocket's empty space: the closest proxy to the dockable cavity (export-pocket-grid.md)
`{name}_pocket_descriptors.{format}`	`-export_pocket_descriptors 1` (2.6+)	`volume`, `radius_of_gyration`, `principal_moments`: useful for sizing and for judging how elongated a pocket is (export-pocket-descriptors.md)

Choosing a basis

Basis	Pros	Cons
Centroid + fixed cube	Simplest; one number to tune	Ignores pocket shape and size; easy to make too small
SAS-point bounding box + margin	Traces the cavity mouth; works on any version	Surface sampling, not the cavity interior
Pocket-grid bounding box (2.6+)	Fills the empty cavity, the most faithful extent	Opt-in; only for `prank predict` pockets (loaded fpocket/etc. pockets have no SAS points)

For most users the SAS-point bounding box is the recommended default. On 2.6+ the pocket grid is a good alternative when you want the extent of the empty cavity rather than its surface.

Recipe: box from SAS points

Export the points with the pocket column, then take the bounding box of the pocket you care about.

prank predict -f protein.pdb -export_points 1 -export_points_format parquet

import pandas as pd

pts = pd.read_parquet("test_output/predict_protein/protein.pdb_points.parquet")
pocket = pts[pts["pocket"] == 1]                 # top-ranked pocket; change as needed

mins = pocket[["x", "y", "z"]].min()
maxs = pocket[["x", "y", "z"]].max()
center = (mins + maxs) / 2
margin = 8.0                                      # >= half the ligand's longest dimension (Angstrom)
size = (maxs - mins) + 2 * margin

print("center_x, center_y, center_z =", center.round(3).tolist())
print("size_x,   size_y,   size_z   =", size.round(3).tolist())

The center_* and size_* values map directly onto AutoDock Vina's center_x/y/z and size_x/y/z (both in Angstrom).

Tip

If you only run plain prank predict (without -export_points), the same coordinates are already in visualizations/data/{name}_points.pdb.gz: the pocket rank is in the residue-sequence column and the score is in the B-factor column. -export_points just gives you a cleaner table to load.

Recipe: box from the pocket grid (2.6+)

The grid covers the pocket's empty space, so its bounding box maps more directly onto the volume a ligand can occupy.

prank predict -f protein.pdb -export_pocket_grid 1 -pocket_grid_format parquet

import pandas as pd

grid = pd.read_parquet("test_output/predict_protein/protein.pdb_pocket_grid.parquet")
pocket = grid[grid["pocket"] == 1]

mins = pocket[["x", "y", "z"]].min()
maxs = pocket[["x", "y", "z"]].max()
center = (mins + maxs) / 2
margin = 4.0                                      # the grid already fills the cavity, so a smaller margin is usually enough
size = (maxs - mins) + 2 * margin

Choosing the margin

A good rule of thumb is a margin of at least half the longest dimension of the ligand(s) you are docking.
P2Rank tends to flag the hotspot core of a site rather than its full extent, so a fixed cube centered on the centroid is easy to undersize for elongated ligands. Using the SAS-point or grid bounding box (which already follows the cavity) plus a margin compensates for this.
Fragment screening can use smaller boxes; larger or more flexible ligands need more room.

Important

There is no universal margin. The reliable way to pick one is empirical: run P2Rank on a handful of complexes with known ligand poses, build boxes with your chosen method and margin, and check that each box actually contains the true pose. Adjust the method or margin and repeat.

Caveats

A cubic box (use max(size_x, size_y, size_z) for all three) is the simplest choice and is what some docking tools expect. An anisotropic box hugs the cavity more tightly but is not supported everywhere.
For multi-pocket docking, loop over the pocket ranks (pocket == 1, pocket == 2, ...) and emit one box per pocket.
Boundary points that fall within the extended shells of two pockets are labeled with the better (lower) rank, so a point belongs to exactly one pocket in the export (see export-points.md).