This paper presents OptiSource, a novel approach of source selection that reduces the number of data sources accessed during query evaluation in complex large scale distributed data contexts in virtual organizations (VO). In these contexts autonomous organizations share data about a group of domain concepts (e.g. patient, client, gene). The instances of such concepts are constructed from non-disjointed fragments provided by several local data sources. This fact, in addition to the absence of reliable statistics on source contents and the large number of sources, make current proposals unsuitable in terms of response quality and/or response time. OptiSource optimizes data source selection in query evaluation using a combinatorial optimization model to distinguish the sets of sources that maximize benefits and minimize the number of sources to contact to while satisfying resource constraints. The precision and recall of source selection during query planning is highly improved as demonstrated by the tests performed with the OptiSource prototype. Furthermore, tests with the optimization model confirmed that the approach can handle different levels of precision on the benefit prediction.