To manage expert knowledge in VR interactions, we propose a software architecture based on semantic representation with Conceptual Graphs. We applied it on structural biology analysis. We define three semantic layers. The first layer embeds basic and specific domain knowledge (objects of interest). The second one describes the available tools in the application (visual representations). The third layer contains facts coming from VR interactions. Operations such as merging, projection, specialization and generalization allow us to build pattern conceptual graphs. They are translated using application specific wrapper into command interpretable by our targeted application. One goal is to release experts from the technical aspects of commands, and allow them to use their field terminology as vocal commands, in order to concentrate on the exploration task.