Riboswitches are structured allosteric RNA molecules that change conformation in response to a metabolite binding event, triggering a regulatory response. In this work we focus on the de-novo design of riboswitches aptamers, which should bind the metabolite and perform the structural switch as well as the natural ones. To this aim, we use the Restricted Boltzmann machines (RBM), an unsupervised machine learning architecture, to learn, only from sequence data, a generative model of the riboswitch family of interest. We first verify, on three different riboswitches families, that RBM generated sequences correctly capture the conservations and covariations of the natural sequences, which are induced by the constraints from the secondary and tertiary structure as well as from the switching mechanism between open and closed conformations. The RBM model is then used to design novel artificial allosteric SAM-I riboswitch aptamers. To experimentally validate the functionality of the designed molecules, we resort to chemical probing (SHAPE-MaP), and develop a tailored analysis pipeline adequate for high-throughput tests of the switching ability of diverse homologous sequences. We probed a total of 476 RBM designed sequences compared to 201 Natural sequences in two experiments. The RBM designed molecules with high RBM-score, showing between 20% and 40% divergence from any natural sequence, display ≈ 30% success rate of correctly undergoing a structural switch in response to SAM, as compared to the natural sequences having ≈ 50% of success rate.
