X-PAIR

X-PAIR is a multi-task deep learning framework designed to jointly predict protein–protein interactions (PPI) and localize residue-level interaction interfaces. It operates directly on protein sequences, leveraging pretrained protein language model (PLM) embeddings and a cross-attention architecture to model interactions between residue representations.

Given pairs of protein sequences, X-PAIR can be trained in three modes:

• interaction-only: predict whether two proteins interact
• interface-only: predict which residues belong to the interaction interface, assuming the pair is interacting
• multi-task: predict both interaction and residue-level interface labels

The full documentation for the package is available in the X-PAIR documentation.

⚙️ Installation

X-PAIR requires Python 3.10 or newer. The package and its required Python dependencies can be installed from PyPI by running:

pip install xpair

Typical installation time: approximately 1 minute on a standard desktop computer with a standard internet connection.

📌 Usage

This package provides three command-line entry points:

• xpair-train: train an interaction-only, interface-only, or multi-task model
• xpair-test: evaluate a trained checkpoint on labeled test data
• xpair-predict: run inference on unlabeled protein pairs

The package already includes a pretrained version of the model (multitask_xfair.ckpt), which is used by default for prediction unless a different checkpoint is specified. This model is trained in a multi-task setting to jointly predict protein–protein interactions and residue-level interfaces for protein sequences ranging from 50 to 2000 amino acids. Interaction data comes from STRING v12 across four different species (H. sapiens, G. gallus, D. melanogaster and S. cerevisiae), while interface data comes from PPI3D (which uses PDB structural information) without species restriction.

All other pretrained models used in the experiments reported in the paper can be found and downloaded from the pretrained_models directory and on Hugging Face.

More detailed instruction of use can be found in X-PAIR documentation. A minimal demo is provided to verify the installation and test X-PAIR workflows on a small example dataset.

🚀 Quick start

X-PAIR predicts whether two proteins interact and, if so, identifies the residues involved in interaction interface.

The simplest way to use X-PAIR is:

xpair-predict --predict_pairs_path pairs_id.tsv --predict_fasta_path sequences.fasta --predict_output_path ./predictions/my_predictions.csv

The file pairs_id.tsv contains protein ID pairs separated by tabs. Example:

id1	id2
protA	protB
protA	protC

The xpair-predict command generates a CSV file that contains both interaction predictions and residue-level interface predictions for all protein pairs listed in pairs_id.tsv.

📚 Citation

If you find this useful, please cite:

Rescalli, S. and Carbone, A. X-PAIR: an ultrafast multitask framework for proteome-scale reconstruction of PPI networks and partner-specific interfaces from sequence. bioRxiv preprint (2026). https://doi.org/10.64898/2026.07.20.739596