The behavioral space is close to infinity, but it can be limited by decomposing behavior into a combination of finite behavioral attributes (just like atoms in molecular structure). This work introduces a general solution, the behavioral molecular structure (BMS) model that characterizes behaviors at the atomic level by transforming the behavioral expression into different combinations of behavioral attributes. A behavioral attribute (like a single atom) is influenced by other behaviors, and perturbations to the behavioral molecular structure can lead to huge differences between behaviors. Correspondingly, this work proposes a GNN-based approach for embedding behavioral molecular structures to obtain highly expressive attribute representations. To demonstrate the effectiveness of behavioral structure, the rationality of behavior molecular structure is examined through theoretical analysis of its expressive power. Furthermore, empirical validation is conducted by deploying this approach in tasks involving fundamental behavioral functions, including detection, prediction, and generation.