Machine learning a model for RNA structure prediction

Abstract

RNA function crucially depends on its structure. Thermodynamic models currently used for secondary structure prediction rely on computing the partition function of folding ensembles, and can thus estimate minimum free-energy structures and ensemble populations. These models sometimes fail in identifying native structures unless complemented by auxiliary experimental data. Here, we build a set of models that combine thermodynamic parameters, chemical probing data (DMS and SHAPE) and co-evolutionary data (direct coupling analysis) through a network that outputs perturbations to the ensemble free energy. Perturbations are trained to increase the ensemble populations of a representative set of known native RNA structures. In the chemical probing nodes of the network, a convolutional window combines neighboring reactivities, enlightening their structural information content and the contribution of local conformational ensembles. Regularization is used to limit overfitting and improve transferability. The most transferable model is selected through a cross-validation strategy that estimates the performance of models on systems on which they are not trained. With the selected model we obtain increased ensemble populations for native structures and more accurate predictions in an independent validation set. The flexibility of the approach allows the model to be easily retrained and adapted to incorporate arbitrary experimental information.

Review of machine learning methods for RNA secondary structure prediction

“…Similarly, another paper [ 137 ] reported that the F score of ContextFold also lowered by 24% when testing on a set of structurally dissimilar RNAs to the training set. Although most DL-based methods take many precautions to alleviate overfitting by many techniques (such as using regularization [ 100 ], enlarging dataset [ 97 ], adding constraints [ 99 ], or combining Turner’s nearest neighbor free energy parameters), the concerns about overfitting remain.…”

“…In addition to the encoded RNA sequences being used as the input, other information can also be incorporated into the DL model. Calonaci and colleagues [ 100 ] trained an ensemble model based on a combination of SHAPE data, co-evolutionary data (DCA), and RNA sequence data. Their model consists of a convolutional neural network (CNN) subnetwork and an MLP subnetwork to predict penalties based on SHAPE and DCA data, respectively, with an RNAfold [ 17 ] module to generate structures using RNA sequences and penalties.…”

ExpertRNA: A new framework for RNA structure prediction

“…Given the impact of structure on RNA functionality, the accurate computational prediction of the secondary and tertiary structure of RNA is an ongoing area of great interest in the computational biology community (Calonaci et al, 2020;Cruz et al, 2012;Wayment-Steele et al, 2020).…”

“…Data driven approaches have also started to become popular that use machine learning to evaluate structures rather than FE or kinetic models. These include ContraFold, DMfold, and structure prediction with neural networks (Do et al, 2006;Wang et al, 2019;Calonaci et al, 2020). Furthermore, in the attempt to achieve advantages of model or data driven approaches, methods have been proposed that attempt to aggregate multiple information sources to get more accurate secondary structure prediction.…”

“…Furthermore, in the attempt to achieve advantages of model or data driven approaches, methods have been proposed that attempt to aggregate multiple information sources to get more accurate secondary structure prediction. Within the data driven category, some examples of information sources are the experimentally determined SHAPE data (Lucks et al, 2011;Low and Weeks, 2010), and evolutionary covariation information (Calonaci et al, 2020). On the model driven side, statistical ensemble approaches are used to boost the solutions obtained by the different FE driven folding algorithms.…”