Service discovery and recommendation has been considered an effective technique for freeing workflow developers out from time-consuming work of manually selecting suitable software services from a sea of service candidates. Facing complex scientific workflow development, recommending multiple chainable services (service unit of work or UoW in short) instead of individual services shall not only increase workflow composition productivity but also address the data shimming problem. Unfortunately, the problem of UoW recommendation is NP-hard. This study presents a novel workflow UoW recommendation engine called RANGER that comprises two integral components: a RANker network and a UoW merGER. The RANker network is trained to rank its comprising edges against a query. At each step during a workflow composition, the RANker consumes the embeddings of the composition context as input. The RANker network mines past service collaboration networks and ranks chained service pairs as “seed” UoWs. The UoW merGER then tries to further chain close-by seed UoWs in the ranked list into larger UoWs as candidate UoWs. Finally, the top K candidate UoWs are suggested to the developers. Extensive experiments on a real-world dataset have demonstrated that the RANGER technique saves development efforts when composing scientific workflows.