Machine learning a model for RNA structure prediction.

RNA function crucially depends on its structure. Thermodynamic models that are used for secondary structure prediction report a large number of structures in a limited energy range, often failing in identifying the correct native structure unless complemented by auxiliary experimental data. In this work we build an automatically trainable model that is based on a combination of thermodynamic parameters, chemical probing data (Selective 2$^\prime$ Hydroxyl Acylation analyzed via Primer Extension, SHAPE), and co-evolutionary data (Direct Coupling Analysis, DCA). Perturbations are trained on a suitable set of systems for which the native structure is known. A convolutional window is used to include neighboring reactivities in the SHAPE nodes of the network, and regularization terms limit overfitting improving transferability. The most transferable model is chosen with a cross-validation strategy that allows to automatically optimize the relative importance of heterogenous input datasets. The model architecture enlightens the structural information content of SHAPE reactivities and their dependence on local conformational ensembles. By using the selected model, we obtain enhanced populations for reference native structures and more sensitive and precise predicted structures in an independent validation set not seen during training. The flexibility of the approach allows the model to be easily retrained and adapted to incorporate arbitrary experimental information.