Considering existing Artificial Intelligence based virus-host protein protein interaction predictors produce decent performance only for specific types of hosts and viruses due to the use of sub-optimal sequence encoding and interaction prediction algorithms, proposed MP-VHPPI makes use of two well known physico-chemical properties based sequence encoding methods namely APAAC and QS order and some optimization strategies to achieve promising performance and generalizeability over 7 benchmark different virus-hosts protein-protein interaction prediction datasets. Unlike existing approaches where all available physico-chemical properties of selected sequence encoders are used, proposed approach makes use of iterative representation learning paradigm where it evaluates the performance potential of each physico-chemical property and finds optimal single or combination of properties to most effectively discretize viral-host protein sequences by capturing comprehensive amino acid order and distribution information. Furthermore, it uses feature agglomeration approach to transform original feature space into more informative feature space. It utilizes the potential of two non-linear classifiers namely Random Forest and Extra Tree to generate 6-Dimensional probabilistic discriminative feature space using separate sequence encodings of APAAC and QS order as well as combined sequence encoding. This discriminative feature space is fed to Support Vector Machine classifier that makes final predictions. MP-VHPPI web interface supports multi-dimensional analysis of viral-host protein sequences, training and optimizing the machine learning models from scratch, using models pre-trained on viral-host proteins belonging to multiple virus and host species to make inference on new viral-host protein sequences, and download interactive artifacts during the lifetime of session.