Nowadays a large majority of software systems are built using various technologies that in turn rely on different languages (e.g. Java, XML, SQL etc.). We call such systems heterogeneous applications (HAs). By contrast, we call software systems that are written in one language homogeneous applications. In HAs the information regarding the structure and the behaviour of the system is spread across various components and languages and the interactions between different application elements could be hidden. In this context applying existing reverse engineering and quality assurance techniques developed for homogeneous applications is not enough. These techniques have been created to measure quality or provide information about one aspect of the system and they cannot grasp the complexity of HAs. In this dissertation we present our approach to support the analysis and evolution of HAs based on: (1) a unified first-class description of HAs and, (2) a meta-model that reifies the concept of horizontal and vertical dependencies between application elements at different levels of abstraction. We implemented our approach in two tools, MooseEE and Carrack. The first is an extension of the Moose platform for software and data analysis and contains our unified meta-model for HAs. The latter is an engine to infer derived dependencies that can support the analysis of associations among the heterogeneous elements composing HA. We validate our approach and tools by case studies on industrial and open-source JEAs which demonstrate how we can handle the complexity of such applications and how we can solve problems deriving from their heterogeneous nature.