Riboswitches are structured allosteric RNA molecules capable of switching between competing conformations in response to a metabolite binding event, eventually triggering a regulatory response. Computational modelling of these molecules is complicated by complex tertiary contacts, conditioned to the presence of their cognate metabolite. In this work, we show that Restricted Boltzmann machines (RBM), a simple two-layer machine learning model, capture intricate sequence dependencies induced by secondary and tertiary structure, as well as the switching mechanism, resulting in a model that can be successfully used for the design of allosteric RNA. As a case study we consider the aptamer domain of SAM-I riboswitches. To validate the functionality of designed sequences experimentally by SHAPE-MaP, we develop a tailored analysis pipeline adequate for high-throughput probing of diverse homologous sequences. We find that among the probed 84 RBM designed sequences, showing up to 20% divergence from any natural sequence, about 28% (and 47% of the 45 among them having low RBM effective energies), are correctly structured and undergo a structural allosteric in response to SAM. Finally, we show how the flexibility of the molecule to switch conformations is connected to fine energetic features of its structural components.