Modeling, coding and running large and multi-scale scientific applications require a high level abstraction of modeling and an extensive use of heterogeneous computing resources. This thesis investigates a methodology to design, implement and run high performance scientific applications on widely distributed infrastructures. First, two workflows categories are defined for scientific applications: embarrassingly-parallel and communication-intensive. For the embarrassingly-parallel category, a programming model was proposed for applications with a dynamic number of jobs and a non-deterministic convergence condition. For the communication-intensive category, a Multiscale Modeling and Simulation Framework (MMSF) was introduced to model, develop and run multiscale applications across distributed high performance infrastructures. The two methodologies have been then applied on real scientific applications. Distributed computations have been performed and evaluated across coupled computing clusters from different infrastructures: EGI, Universities and cloud resources. The two methodologies are attractive and enable scientists to change the way of doing computational sciences.