xgBoost based Interface Prediction of Specific Partner Interactions (BIPSPI) is a method for the prediction of partner-specific protein interfaces from pdb files or input sequences. BIPSPI employs Extreme Gradient Boosting (XGBoost) models trained on residue pairs of the protein complexes and a scoring function that converts pair prediction to interface residue predictions.
BIPSPI+ has been trained on two different datasets. One dataset for hetero-oligomeric interactions and another one for homo-oligomeric interactions. The rational behind is that, althouhgh the physics behind all types of interactions is the same, statistically speaking, the interfaces of homo-complexes tend to be different from the ones of hetero-complexes. Compared to version 1, that was trained only on the complexes included in the Protein-Protein Docking Benchmark version 5, our new training sets have grown one order of magnitude, offering better performace, especially for the case of homo-oligomeric interactions.
Datasets and precomputed results used for training and evaluation are available in the download section.
BIPSPI v1 is no longer maintained, but some downloadables can be found in the download section - Version 1