Allosteric regulation is a well-established phenomenon defined as a distal conformational or dynamical change of the protein upon allosteric effector binding. Here, we developed a novel approach to delineate allosteric effects in proteins. In this approach, we applied robust machine learning methods, including deep neural network and random forest, on extensive molecular dynamics (MD) simulations to distinguish otherwise similar allosteric states of proteins. Using the PDZ3 domain of PDS-95 as a model protein, we demonstrated that the allosteric effects could be represented as residue-specific properties through two-dimensional property-residue maps, which we refer to as "residue response maps". These maps were constructed through two machine learning methods and could accurately describe how different properties of various residues are affected upon allosteric perturbation on protein. Based on the "residue response maps", we propose allostery as a residue-specific concept, suggesting that all residues could be considered as allosteric residues because each residue "senses" the allosteric events through changing its single or multiple attributes in a quantitatively unique way. The "residue response maps" could be used to fingerprint a protein based on the unique patterns of residue responses upon binding events, providing a novel way to systematically describe the protein allosteric effects of each residue upon perturbation.