Classifying discourse relations is a hard task: discourse-annotated data is scarce, especially for languages other than English, and there exist different theoretical frameworks that affect textual spans to be linked and the label set used. Thus, work on transfer between languages is very limited, especially between frameworks, while it could improve our understanding of some theoretical aspects and enhance many applications. In this paper, we propose the first experiments on zero-shot learning for discourse relation classification and investigate several paths in the way source data can be combined, either based on languages, frameworks, or similarity measures. We demonstrate how difficult transfer is for the task at hand, and that the most impactful factor is label set divergence, where the notion of underlying framework possibly conceals crucial disagreements.