DeepMiRBP: a hybrid model for predicting microRNA-protein interactions based on transfer learning and cosine similarity

Abstract

Background: Interactions between microRNAs and RNA-binding proteins are crucial for microRNA-mediated gene regulation and sorting. Despite their significance, the molecular mechanisms governing these interactions remain underexplored, apart from sequence motifs identified on microRNAs. To date, only a limited number of microRNA-binding proteins have been confirmed, typically through labor-intensive experimental procedures. Advanced bioinformatics tools are urgently needed to facilitate this research.

Methods: We present DeepMiRBP, a novel hybrid deep learning model specifically designed to predict microRNA-binding proteins by modeling molecular interactions. This innovation approach is the first to target the direct interactions between small RNAs and proteins. DeepMiRBP consists of two main components. The first component employs bidirectional long short-term memory (Bi-LSTM) neural networks to capture sequential dependencies and context within RNA sequences, attention mechanisms to enhance the model's focus on the most relevant features and transfer learning to apply knowledge gained from a large dataset of RNA-protein binding sites to the specific task of predicting microRNA-protein interactions. Cosine similarity is applied to assess RNA similarities. The second component utilizes Convolutional Neural Networks (CNNs) to process the spatial data inherent in protein structures based on Position-Specific Scoring Matrices (PSSM) and contact maps to generate detailed and accurate representations of potential microRNA-binding sites and assess protein similarities.

Results: DeepMiRBP achieved a prediction accuracy of 87.4% during training and 85.4% using testing, with an F score of 0.860. Additionally, we validated our method using three case studies, focusing on microRNAs such as miR-451, -19b, -23a, -21, -223, and -let-7d. DeepMiRBP successfully predicted known miRNA interactions with recently discovered RNA-binding proteins, including AGO, YBX1, and FXR2, identified in various exosomes.

Conclusions: Our proposed DeepMiRBP strategy represents the first of its kind designed for microRNA-protein interaction prediction. Its promising performance underscores the model's potential to uncover novel interactions critical for small RNA sorting and packaging, as well as to infer new RNA transporter proteins. The methodologies and insights from DeepMiRBP offer a scalable template for future small RNA research, from mechanistic discovery to modeling disease-related cell-to-cell communication, emphasizing its adaptability and potential for developing novel small RNA-centric therapeutic interventions and personalized medicine.

Keywords: Deep learning; Interaction prediction; MicroRNAs; RNA binding proteins; RNA sorting.

Fig. 1

The schematic workflow of the DeepMiRBP model. In the first component, the source-domain model is trained based on RNA sequences related to known binding sites of different RBPs(RNA-binding proteins). After this training phase, the learned parameters are transferred to the target domain using a transfer learning approach. The target model is then retrained using sequences of protein-interacting miRNAs as input. A cosine similarity measure is applied to identify and rank RBP sequences from the source domain that are most similar to the given miRNA, resulting in a ranked list of candidate proteins. The candidate proteins identified in the first component undergo further analysis in the second component. Position-Specific Scoring Matrices (PSSM) and contact maps are utilized for each candidate protein to perform a more comprehensive similarity assessment. This step enhances the understanding of miRNA-protein interactions, thereby improving the model’s prediction accuracy

Fig. 2

The detailed architecture of DeepMiRBP in both components for predicting microRNA-protein interactions. a First component architecture: This model focuses on training RNA sequences that bind to RBPs to capture intricate features of RNA-protein interactions. Initially, the model learns from RNA sequences bound by RBPs and transfers this knowledge to the target domain. Here, miRNA sequences serve as input, generating embedding codes. Cosine similarity is then applied to identify RNA sequences most similar to the miRNA sequences. b Second component architecture: In this model, each RBP candidate identified in the first part is processed using PSSM and contact maps. CNN layers and max-pooling are employed to encode these matrices. Subsequently, cosine similarity is calculated to compare RBP candidates with other proteins, resulting in a matrix that identifies proteins with a higher probability of binding to the miRNA sequence

Fig. 3

Confusion matrix for test data in the source domain

Fig. 4

ROC Performance. The ROC curve for predicting RNA-protein binding sites on 31 experiment datasets