Protein - protein interactions (PPIs) are important since most of biological activities are constructed of a network of a series of PPIs. Therefore specifically targeting proteins which are involved in key biological activities can lead to controlling of entire system, which is the reason why proteins have been considered as the major drug target. Studies on PPIs have been done for a long time by widely used experimental methods such as yeast two hybrid method, SPOT array synthesis, or X-ray crystallography, but those methods are expensive and time-consuming procedures. Consequently, prediction of PPIs and selecting a probable set of protein - protein pairs beforehand is highly desirable. Previous methods basically focus their study in predicting result which has answers. But algorithms which are made and trained based on known data are often failed when novel data set is introduced, i.e. overfitting. Therefore in this research our focus lies on developing method which is flexible to various unknown data. In this study SH3 domain is used for prediction, since the consensus binding motif of SH3 is classified and only small part, 10 amino acids - decapeptides, of binding protein is most important in interaction, therefore problem can be simplified. First, homology modeling of SH3 domains of unknown sequence is done by searching the most similar sequence from PDB and 3 dimensional complex structures which are formed by combining various peptide sequences are generated with MODELLER. A series of molecular dynamics (MD) simulations were performed to minimize intra and inter-chain clashes and to stabilize peptide backbone structures. Then interaction energies between SH3 domain and peptides were calculated based on the complex structures generated by MD simulations. Finally, list of 200 top ranking peptide sequences are used to generate position weighted matrix (PWM). We could conclude with certain aspect that our method can predict PPIs of unknown sequence.