X-PAIR datasets

This repository contains the processed datasets used to train and evaluate X-PAIR, a multi-task deep-learning model designed to jointly predict protein–protein interactions and residue-level interfaces. The datasets were constructed from publicly available biological resources. Protein–protein interaction data were obtained from STRING v12.0, while non-interacting pairs were generated using the "neighbour exclusion strategy" introduced in SENSE-PPI. As STRING reports same-organism interactions, the interaction dataset contains only intra-species protein pairs. Residue-level interface annotations were obtained from PPI3D and may include both intra-species and inter-species protein interactions. The original data were processed through dedicated filtering and curation pipelines to generate high-quality, non-redundant datasets suitable for model training and evaluation. Four dataset collections are provided: X-fair, a homology-aware benchmark containing both interaction and interface data, designed to reduce information leakage; X-human, containing both interaction and interface data and designed to evaluate cross-species generalisation from human-only training data; X-multispecies, also containing both interaction and interface data and designed to evaluate generalisation using a broader multi-species training set; X-taxonomic-interface, an interface-only benchmark designed to assess residue-level interface prediction across different taxonomic groups. For each dataset collection, training, validation, and test splits are provided. The splits are distributed as tab-separated files containing protein identifiers and the corresponding ground-truth labels, together with FASTA files containing the amino acid sequences associated with the proteins included in each dataset. Additional information is provided in the README.md file included in the dataset folder. Detailed preprocessing, filtering, and dataset construction procedures are described in the X-PAIR manuscript.