Accurate molecular crystal structure prediction is a fundamental goal in academic as well as industrial condensed matter research. Some of the biggest challenges standing in the way are polymorphism and cost of the search in a vast phase space with many degrees of freedom. Here we tackle these challenges in the difficult case of Glycine, the most polymorphic aminoacid. Despite being repeatedly studied Glycine still hosts little-known phase transitions and unsolved new phases. In this talk we demonstrate how recent progress in Density Functional Theory (as implemented in Quantum ESPRESSO) as well as in evolutionary algorithms (as implemented in USPEX) for crystal structure prediction enabled a leap in predictive power thanks to increased accuracy . We also report the experimental confirmation of our prediction for the elusive Zeta-phase of Glycine and the new questions that arise with it . Finally we propose a phase space search strategy based on machine learning methods (as implemented in TensorFlow) that promises an order of magnitude lower computational cost without compromising accuracy, opening wider pathways for crystal structure prediction in academia and industry. References:  C.H.Pham, E. Kucukbenli, S. de Gironcoli, arXiv:1605.00733  C.L. Bull, G. Flowitt-Hillab, S. de Gironcoli, E. Kucukbenli, S. Parsons, C.H. Pham, H. Y. Playford, M.G. Tucker (submitted to IUCrJ)  E. Kucukbenli, R. Lot (in preparation).