Allosteric communication underlies ligand-dependent transcriptional responses mediated by nuclear receptors. While studies have elucidated many of the components involved in this process, the energetic architecture within the receptor protein that mediates allostery remains unknown. Using a sequence-based method designed to detect coevolution of amino acids in a protein, termed the statistical coupling analysis (SCA), we identify a network of energetically coupled residues that link the functional surfaces of nuclear receptor ligand binding domains. Functional analysis of these predicted residues demonstrates their participation in an allosteric network that governs the ability of heterodimeric receptors to activate transcription in response to ligand binding by either partner. Interestingly, mutation of a single network residue can discriminate between receptor activation by endocrine, dietary, and synthetic agonists. These results reveal a structural network required for RXR heterodimer allosteric communication and suggest that the specificity of ligand response and permissivity coevolved to enable signal discrimination.