In the Biodiversity World project we are building a GRID to support scientific biodiversity-related research. The requirements associated with such a GRID are somewhat different from other GRIDs, and this has influenced the architecture that we have developed. In this paper we outline these requirements, most notably the need to interoperate over a diverse set of legacy databases and applications in an environment that supports effective resource discovery and use of these resources in complex workflows. Our architecture provides an invocation model that is usable over a wide range of resource types and underlying GRID middleware. However, there is a trade-off between the flexibility provided by our architecture and its performance. We discuss how this affects the inclusion of computationally intensive applications and applications that are highly interactive; we also consider the broader issue of interoperation with other GRIDs. © Springer-Verlag Berlin Heidelberg 2005.