Over the past decade, the fourth paradigm of data-intensive science rapidly became a major driving concept of multiple application domains encompassing and generating large-scale devices such as light sources and cutting edge telescopes. The success of data-intensive projects subsequently triggered the next generation of machine learning approaches. These new artificial intelligent systems clearly represent a paradigm shift from data processing pipelines towards the fifth paradigm of composite cognitive applications requiring the integration of Big Data processing platforms and HPC technologies. The paper addresses the existing impedance mismatch between data-intensive and compute-intensive ecosystems by presenting the Spark-MPI approach based on the MPI Exascale Process Management Interface (PMIx). The approach is demonstrated within the context of hybrid MPI/GPU ptychographic image reconstruction pipelines and distributed deep learning applications.