Sequence pre-training-based graph neural network for predicting lncRNA-miRNA associations.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• MicroRNAs (miRNAs) regulate gene expression by targeting messenger RNAs (mRNAs).
• Long non-coding RNAs (lncRNAs) can function as competitive endogenous RNAs (ceRNAs), modulating miRNA activity and influencing gene expression.
• Identifying lncRNA-miRNA interactions is vital for understanding gene regulation but experimentally challenging.

Purpose of the Study:

• To propose a novel deep learning framework, sequence pre-training-based graph neural network (SPGNN), for predicting lncRNA-miRNA associations.
• To leverage RNA sequences and existing interaction networks for accurate prediction.

Main Methods:

• Utilized a sequence-to-vector approach for RNA sequence pre-training to generate embeddings.
• Employed a Graph Neural Network (GNN) in the fine-tuning stage to learn from a heterogeneous graph of lncRNA-miRNA interactions.
• Combined k-mer technique and Doc2vec for pre-training with Simple Graph Convolution Network for fine-tuning.

Main Results:

• SPGNN demonstrated superior performance in predicting lncRNA-miRNA associations compared to state-of-the-art methods on an animal dataset.
• Effectiveness of individual components and parameters was validated through ablation studies and hyperparameter analysis.

Conclusions:

• The proposed SPGNN model effectively predicts lncRNA-miRNA associations by integrating sequence information and network topology.
• This approach offers a valuable computational tool for exploring ceRNA mechanisms and their medical implications.