2OMe-LM: predicting 2'-O-methylation sites in human RNA using a pre-trained RNA language model.

Area of Science:

• Computational Biology
• Bioinformatics
• Molecular Biology

Background:

• 2'-O-methylation (2OMe) is a critical RNA modification influencing gene expression and disease.
• Experimental identification of 2OMe sites is laborious and expensive.
• RNA pre-trained language models offer powerful tools for RNA bioinformatics.

Purpose of the Study:

• To develop an efficient computational method for predicting 2OMe sites in RNA.
• To leverage recent advancements in RNA pre-trained language models for 2OMe site prediction.
• To address the gap in applying deep learning models for 2OMe site identification.

Main Methods:

• A novel deep learning framework, 2OMe-LM, was proposed.
• Integrated RNA sequence features from RNA pre-trained language models and word2vec.
• Employed fully connected layers, bidirectional LSTM, and a feature fusion module.
• Incorporated an attention block for prediction interpretability.

Main Results:

• 2OMe-LM significantly outperformed existing state-of-the-art predictors.
• Features from RNA pre-trained language models were identified as critical for prediction accuracy.
• Motif analysis indicated potential for discovering 2OMe-related motifs.

Conclusions:

• 2OMe-LM provides a highly effective deep learning approach for predicting RNA 2OMe sites.
• The study highlights the importance of RNA pre-trained language models in this domain.
• The developed framework offers a valuable tool for RNA modification research.