Protein-protein interactions represent an essential mechanism, which regulates several cellular functions and is frequently involved in the pathways and networks that link receptors to their ultimate targets. Protein interaction domains can be divided into separate families that are related either to sequence or to ligand-binding properties. One of the most abundant molecular-recognition module is the SH3 domain that generally recognizes proline-rich peptides characterized by the minimal consensus sequence PxxP able to form a left-handed polyproline type II helix. The requirements of SH3 domains for binding to ligands have been examined by numerous approaches and two classes of SH3 domain ligands have been defined (class I and class II), the (R)KxxPxxP and PxxPxK(R) motifs, respectively. Recently, several SH3 domains have been observed to bind unconventional ligands. Atypical SH3 binding motifs are characterized either by the requirement of additional specific residues or by the presence of characteristic sequences instead of the core motif. Here we review literature on the specificity of SH3 domains, with an emphasis on the structural basis of atypical ligand recognition.