p2rank/documentation/rescoring.md

Rescoring Predictions from Other Methods

P2Rank can rescore pocket predictions from other binding site prediction tools, re-ranking their pockets using its own ML model.

Quick Start

prank rescore test_data/fpocket.ds                            # rescore fpocket predictions
prank rescore test_data/pocketeer.ds -o rescore_pocketeer     # rescore pocketeer, output to specific dir
prank eval-rescore test_data/fpocket.ds                       # rescore and evaluate against known ligands
prank fpocket-rescore test_data/basic.ds                      # run fpocket and rescore in one step
prank rescore test_data/pocketeer.ds -c rescore_2024          # use new experimental rescoring model

Commands

Command	Description
prank rescore <dataset.ds>	Rescore predictions and output re-ranked pockets.
prank eval-rescore <dataset.ds>	Rescore and evaluate against known ligands.
prank fpocket-rescore <dataset.ds>	Run Fpocket on proteins, then rescore. Convenience shortcut that can be used as a drop-in replacement for prank predict.

Supported Methods

Method	PREDICTION_METHOD	Prediction column points to	Links
Fpocket	fpocket	Fpocket output file (.pdb/.cif)	GitHub, paper
Pocketeer	pocketeer	pockets.json file	GitHub
PUResNetV2.0	puresnet	Directory with *.pkt.pdb files	GitHub, paper
ConCavity	concavity	*_pocket.pdb grid file	project page, paper
SiteHound	sitehound	*_summary.dat file	paper
DeepSite	deepsite	Results PDB file	paper
MetaPocket2	metapocket2	PDB file with MPT residues	paper
LISE	lise	PDB file with HETATM records	paper
P2Rank	p2rank	*_predictions.csv file	GitHub, paper
SwinSite	swinsite	Per-protein directory with grid<N>_score_<S>.mol2 files	GitHub, paper
Seq2Pocket	seq2pocket	Per-protein directory with <ID>_predictions.txt	GitHub, paper

The last two methods point the prediction column at a per-protein directory rather than a single file. See Rescoring directory-based predictions below for an example.

Dataset File Format

A dataset file (.ds) tells P2Rank which prediction method was used, and lists pairs of prediction output files and their corresponding protein structures.

# Lines starting with # are comments

PARAM.PREDICTION_METHOD=<method>

HEADER: prediction protein

path/to/prediction_output  path/to/protein.pdb

Required elements:

• PARAM.PREDICTION_METHOD -- name of the prediction method (see table above)
• HEADER: line -- defines column order (must include prediction and protein)
• Data rows -- whitespace-separated paths (relative to the .ds file location)

The protein column should point to the structure that was used as input to the prediction tool. For eval-rescore, the protein must contain ligands (to compute evaluation metrics). For plain rescore, ligands are not needed.

The column order in HEADER: is flexible -- prediction protein or protein prediction are both valid.

Examples

Rescoring Fpocket predictions

my_fpocket.ds:

PARAM.PREDICTION_METHOD=fpocket

HEADER: prediction protein

fpocket_output/1abc_out/1abc_out.pdb  structures/1abc.pdb
fpocket_output/2xyz_out/2xyz_out.pdb  structures/2xyz.pdb

prank rescore my_fpocket.ds

Rescoring Pocketeer predictions

my_pocketeer.ds:

PARAM.PREDICTION_METHOD=pocketeer

HEADER: prediction protein

pocketeer_output/1abc/pockets.json  structures/1abc.pdb
pocketeer_output/2xyz/pockets.json  structures/2xyz.cif

prank rescore my_pocketeer.ds

Evaluating rescoring quality

Use eval-rescore with liganated proteins to compare the original ranking against the rescored ranking. This works with any supported method.

my_eval.ds:

PARAM.PREDICTION_METHOD=fpocket

HEADER: prediction protein

fpocket_output/1abc_out/1abc_out.pdb  liganated/1abc.pdb
fpocket_output/2xyz_out/2xyz_out.pdb  liganated/2xyz.pdb

prank eval-rescore my_eval.ds

This outputs evaluation metrics (DCA, DSO success rates, etc.) showing whether rescoring improved pocket ranking.

Rescoring directory-based predictions (SwinSite, Seq2Pocket)

For these methods, the prediction column points to the per-protein output directory (not a single file). The loader picks up the expected files inside: grid*_score_*.mol2 for SwinSite, <ID>_predictions.txt for Seq2Pocket.

my_swinsite.ds:

PARAM.PREDICTION_METHOD=swinsite

HEADER: prediction protein

swinsite_output/1abc  structures/1abc.pdb
swinsite_output/2xyz  structures/2xyz.pdb

prank rescore my_swinsite.ds

The same pattern applies to seq2pocket: point each row at the directory containing its _predictions.txt.

Output

For each protein, two files are generated in the output directory:

File	Contents
{name}_rescored.csv	Re-ranked pockets with new scores
{name}_predictions.csv	Pocket details (scores, centers, residues, surface atoms)

The _rescored.csv contains columns:

Column	Description
name	Pocket name
score	New score assigned by P2Rank
rank	New rank (after rescoring)
old_rank	Original rank from the prediction method

PyMOL visualization files are also generated by default (disable with -visualizations 0).

Parameters

Override parameters on the command line with -param value. A few commonly used parameters:

prank rescore dataset.ds -o output_dir -threads 4 -visualizations 0

Parameter	Default	Description
-o	auto-generated	Explicit output directory (overrides default)
-threads	all CPUs	Number of parallel threads
-visualizations	true	Generate PyMOL visualization files
-fail_fast	false	Stop on first error
-model	default_rescore	ML model to use for rescoring

Experimental Rescoring Model (rescore_2024)

An alternative rescoring model is available via -c rescore_2024. It uses a different feature set that does not depend on B-factor, making it suitable for AlphaFold models, NMR, and cryo-EM structures.

prank rescore fpocket.ds -c rescore_2024
prank fpocket-rescore test.ds -c rescore_2024
prank eval-rescore fpocket.ds -c rescore_2024

This model shows promising results but has not been fully evaluated yet.

Conservation-aware rescoring (rescore_conservation)

A rescoring model that incorporates per-residue sequence conservation scores alongside the standard P2Rank features. Works with any supported prediction method, not just Fpocket.

prank rescore fpocket.ds -c rescore_conservation \
  -conservation_dirs path/to/cons/

Requires HMMER-based .hom conservation files (one per chain, named {baseName}_{chainId}.hom). See conservation.md for the file format and pipeline.