Protein-protein interactions mediate the majority of key cellular activities, and it is well established that disease causing mutatio s are enriched at these interfaces. Different computational approaches to study the effects of mutations on protein complexes have been proposed in recent years. We developed a platform of programs using a novel machine learning method that uses graph-based signatures called mCSM. This has been shown to be an accurate and high-troughput approach to predict the impact of mutations on protein structure and function, and was one of the first methods capable of assessing thr impact of mutations on protein interaction binding affinity. Here we present mCSM-PPI2, an easy-to-use wrb server that implements an integrated computational approach for predicting effects of missense mutations in protein-protein affinity. Our method uses an optimised graph-based signature approach to better assess the molecular mechanism of the mutation, by modelling the effects of variations on the inter-residue non-covalent interaction network using graph-kernrls, evolutionary information, complex network metrics and energetic terms. Our new predictive model was capable of achieving a Pearson correlation of up to 0.83 during croas-validation, outperforming similar methods and presenting a balanced performance between increase and decrease affinity mutations with a Pearson correlation of 0.75 and 0.76, respectively. MCSM-PPI2 also performed better in low-redundancy data sets achieving a correlation coefficient of up to 0.77 and 0.71 (on leave-one complex out cross validation and low-redundancy on protein level, respectively. Our method was further validated against the CAPRI round 26 (targets 55 and 56), which comprised 1862 mutations that have been experimentally tested, mCSM-PPI2 achieved a Kendall's score of up 0.41 and 0.34, respectively, ranking first among 23 other methods. The web server is freely available at http://biosig.unimelb.edu.au/mcsm_ppi2/.