A great challenge in the proteomics and structural genomics era is to discover protein structure and function, including the identification of biological partners. Experimental investigation is costly and time-consuming, making computational methods very attractive for predicting protein function. In this work, we used the existing structural information in the SH3 family to first extract all SH3 structural features important for binding and then used this information to select the right templates to homology model most of the Saccharomyces cerevisiae SH3 domains. Second, we classified, based on ligand orientation with respect to the SH3 domain, all SH3 peptide ligands into 29 conformations, of which 18 correspond to variants of canonical type I and type II conformations and 11 correspond to non-canonical conformations. Available SH3 templates were expanded by chimera construction to cover some sequence variability and loop conformations. Using the 29 ligand conformations and the homology models, we modelled all possible complexes. Using these complexes and in silico mutagenesis scanning, we constructed position-specific ligand binding matrices. Using these matrices, we determined which sequences will be favorable for every SH3 domain and then validated them with available experimental data. Our work also allowed us to identify key residues that determine loop conformation in SH3 domains, which could be used to model human SH3 domains and do target prediction. The success of this methodology opens the way for sequence-based, genome-wide prediction of protein-protein interactions given enough structural coverage.