In order to share data, creating a data transform is a mandatory step since disparate data sources have heterogeneities in data representation. Manual transform creation is extremely tedious and error-prone even though we have a repository of usable transforms. Recent work has focused on finding a data transform as part of data mapping or schema matching. However, the work is mostly about a structural data mapping or applying a single data transform between schemas. We consider a \textit{transform composition} which reuses existing transforms for constructing users' desired transforms. Users' desired transforms and existing transforms are described in our transform model represented in RDF triples. Our transform model includes not only semantics of input/output fields but also behavior of a transform. We model a transform composition problem as a graph search problem. Also, we describe overall architecture of our system which automatically discovers users' desired transforms.