Pinc : a simple probabilistic AlphaFold interaction score

---

This repository is a read-only mirror of the primary repository on GitLab: https://git.mpi-cbg.de/tothpetroczylab/Pinc

Description

Pinc (Probability of interface native contacts) is a simple interaction confidence metric derived from an AlphaFold model and its predicted aligned error (PAE) matrix. A Pinc score of 0.8 means that approximately 80% of native interface contacts are predicted to be present in the model. The method can be used on both AlphaFold2 and AlphaFold3 (including server) outputs.

Using the Colab notebook

• Open the interactive Colab notebook here or click "Open in Colab" above
• Run the first cell (▶) to enable file upload
• Upload a structure file (PDB/mmCIF) its matching PAE matrix (JSON)
• Execute subsequent cells step by step (▶)
• Download the results as a zipped folder in the final cell

Command-line usage

Clone the repository and compile the program. A C compiler (gcc/clang) is required and is pre-installed on most systems:

git clone https://git.mpi-cbg.de/tothpetroczylab/Pinc.git
cd Pinc
make

Getting started

Calling the program empty will display its usage information:

./Pinc

Description:
    R script to compute the Pinc score from an AlphaFold PAE
    matrix (JSON) and the corresponding structure file (PDB/CIF).

 Usage:
    Rscript Pinc.R <json_file> <structure_file> [options]

 Options:
    None        Pinc score for all unique chain pairs (default)
    --all       Full contact probability matrix
    --pairlist  Residue pair list of non-zero contact probabilities
    --results   Optional path to output folder

 Output format:
    - default:  <structure_file_name>_Pinc.csv
    - all:      <structure_file_name>_contact_probability.json
    - pairlist: <structure_file_name>_pairlist.csv

Usage

The test directory contains example input files you can use to test the below functionalities.

Assuming you have a structure file model_0.cif and an associated JSON data file model_0.json with the PAE matrix, to calculate Pinc scores for all chain pairs you can call the program simply by:

./Pinc model_0.json model_0.cif

The program will generate a model_0_Pinc.csv file with columns chain1, chain2, Pinc1, Pinc2, and Pinc, where chain pairs are non-redundant (for A-B there is no B-A). Pinc1 and Pinc2 are the mean contact probabilities viewed from each chain's own PAE column, and Pinc is their symmetric average.

If you need the full contact probability matrix, add the --all flag:

./Pinc model_0.json model_0.cif --all

In this case the program will additionally generate a model_0_contact_probability.json file with node names token_chain_ids and contact_probability, similar to the AlphaFold3 output. Note that token_chain_ids define the shape of the square matrix that can be reconstructed from the numeric array.

If you need a list of all non-zero probabilities at the token level, call:

./Pinc model_0.json model_0.cif --pairlist

The program will additionally generate a model_0_pairlist.csv file with columns token1, token2, contact_p, and distance columns. The tokens represent residues for proteins and bases for nucleic acids, but can also stand for atoms in ligands or ions. The pair list is non-redundant (for chains A-B there is no B-A), and represent the mean of the two contacts (i->j and j->i). The distance is the centre-of-mass distance between the tokens given in Ångströms.

Note that --all and --pairlist flags can be used together.

If you would like to specify a results folder, use:

./Pinc model_0_full_data.json model_0.cif --results /usr/home/path_to_pinc_folder

Make sure the directory has write access.

Legacy R script

An R implementation (Pinc.R) is retained in the repository for reference. It requires R (≥ 3.5.0) and the jsonlite package. Note that repeated calls make it substantially slower than the C program due to interpreter overhead. Usage is identical to the C version, substituting Rscript Pinc.R for ./Pinc:

Rscript Pinc.R model_0.json model_0.cif

Contact

Please report any technical problems and questions via Issues or alternatively by email: [email protected]

Reference

@article{Pinc,
  title={Pinc: a simple probabilistic AlphaFold interaction score},
  author={Badonyi, Mihaly and Toth-Petroczy, Agnes},
  journal={bioRxiv},
  year={2026},
  publisher={Cold Spring Harbor Laboratory}
}