Abstract Motivation Protein–Protein Interactions (PPI) are essentials for most cellular processes and thus, unveiling how proteins interact is a crucial question that can be better understood by identifying which residues are responsible for the interaction. Computational approaches are orders of magnitude cheaper and faster than experimental ones, leading to proliferation of multiple methods aimed to predict which residues belong to the interface of an interaction. Results We present BIPSPI, a new machine learning-based method for the prediction of partner-specific PPI sites. Contrary to most binding site prediction methods, the proposed approach takes into account a pair of interacting proteins rather than a single one in order to predict partner-specific binding sites. BIPSPI has been trained employing sequence-based and structural features from both protein partners of each complex compiled in the Protein–Protein Docking Benchmark version 5.0 and in an additional set independently compiled. Also, a version trained only on sequences has been developed. The performance of our approach has been assessed by a leave-one-out cross-validation over different benchmarks, outperforming state-of-the-art methods. Availability and implementation BIPSPI web server is freely available at http://bipspi.cnb.csic.es. BIPSPI code is available at https://github.com/bioinsilico/BIPSPI. Docker image is available at https://hub.docker.com/r/bioinsilico/bipspi/. Supplementary information Supplementary data are available at Bioinformatics online.