MicroRNAs are involved in many critical cellular activities through binding to their mRNA targets, for example, in cell proliferation, differentiation, death, growth control, and developmental timing. Prediction of microRNA targets can assist in efficient experimental investigations on the functional roles of these small noncoding RNAs. Their accurate prediction, however, remains a challenge due to the limited understanding of underlying processes in recognizing microRNA targets. In this article, we introduce an algorithm that aims at not only predicting microRNA targets accurately but also assisting in vivo experiments to understand the mechanisms of targeting. The algorithm learns a unique hypothesis for each possible mechanism of microRNA targeting. These hypotheses are utilized to build a superior target predictor and for biologically meaningful partitioning of the data set of microRNA–target duplexes. Experimentally verified features for recognizing targets that incorporated in the algorithm enable the establishment of hypotheses that can be correlated with target recognition mechanisms. Our results and analysis show that our algorithm outperforms state-of-the-art data-driven approaches such as deep learning models and machine learning algorithms and rule-based methods for instance miRanda and RNAhybrid. In addition, feature selection on the partitions, provided by our algorithm, confirms that the partitioning mechanism is closely related to biological mechanisms of microRNA targeting. The resulting data partitions can potentially be used for in vivo experiments to aid in the discovery of the targeting mechanisms.

RNA inverse design is an essential part of many RNA therapeutic strategies. To date, there have been great advances in computationally driven RNA design. The current machine learning approaches can predict the sequence of an RNA given its 3D structure with acceptable accuracy and at tremendous speed. The design and engineering of RNA regulators such as riboswitches, however, is often more difficult, partly due to their inherent conformational switching abilities. Although recent state-of-the-art models do incorporate information about the multiple structures that a sequence can fold into, there is great room for improvement in modeling structural switching. In this work, a relational geometric graph neural network is proposed that explicitly incorporates alternative structures to predict an RNA sequence. Converting the RNA structure into a geometric graph, the proposed model uses edge types to distinguish between the primary structure, secondary structure, and spatial positioning of the nucleotides in representing structures. The results show higher native sequence recovery rates over those of gRNAde across different test sets (eg. 72% vs. 66%) and a benchmark from the literature (60% vs. 57%). Secondary-structure edge types had a more significant impact on the sequence recovery than the spatial edge types as defined in this work. Overall, these results suggest the need for more complex and case-specific characterization of RNA for successful inverse design.

(1) Background: MicroRNAs are non-coding RNA sequences that regulate cellular functions by targeting messenger RNAs and inhibiting protein synthesis. Identifying their target sites is vital to understanding their roles. However, it is challenging due to the high cost and time demands of experimental methods and the high false-positive rates of computational approaches. (2) Methods: We introduce a Multi-Input Neural Network (MINN) algorithm that integrates diverse biologically relevant features, including the microRNA duplex structure, substructures, minimum free energy, and base-pairing probabilities. For each feature derived from a microRNA target-site duplex, we create a corresponding image. These images are processed in parallel by the MINN algorithm, allowing it to learn a comprehensive and precise representation of the underlying biological mechanisms. (3) Results: Our method, on an experimentally validated test set, detects target sites with an AUPRC of 0.9373, Precision of 0.8725, and Recall of 0.8703 and outperforms several commonly used computational methods of microRNA target-site predictions. (4) Conclusions: Incorporating diverse biologically explainable features, such as duplex structure, substructures, their MFEs, and binding probabilities, enables our model to perform well on experimentally validated test data. These features, rather than nucleotide sequences, enhance our model to generalize beyond specific sequence contexts and perform well on sequentially distant samples.

RNA inverse design is an essential part of many RNA therapeutics strategies. To date, there have been great advances in the computationally-driven RNA design. Current machine-learning approaches can predict the sequence of an RNA given its 3D structure with acceptable accuracy and tremendous speed. Design and engineering of RNA regulators such as riboswitches, however, is often more difficult, partly due to their inherent conformational switching abilities. Although recent state-of-the-art models do incorporate information about the multiple structures a sequence can fold into, there is great room for improvement in modeling structural switching. In this work, a relational geometric graph neural network is proposed that explicitly incorporates alternative structures to predict the RNA sequence. The proposed model uses edge types to distinguish primary-, secondary-, and tertiary-structure pairing/positioning of nucleotides for its training. Results show higher native sequence recovery rates over gRNAde across different test sets (eg. 72% vs 66%) and a benchmark from the literature (60% vs 57%). The impact of secondary-structure on prediction accuracy was more significant than that of tertiary-structure dependencies as defined here. The code for the above relational GNN can be found here.