CytoLNCpred-a computational method for predicting cytoplasm associated long non-coding RNAs in 15 cell-lines

Area of Science:

• Molecular Biology
• Genomics
• Bioinformatics

Background:

• The function of long non-coding RNAs (lncRNAs) is critically dependent on their subcellular localization.
• Existing prediction methods suffer from noisy datasets and exclude lncRNAs with subtle localization differences.

Purpose of the Study:

• To develop and evaluate a robust method for predicting cytoplasm-associated lncRNAs across 15 human cell lines.
• To identify lncRNAs with higher abundance in the cytoplasm compared to the nucleus.

Main Methods:

• Development of machine learning (ML) and deep learning models using traditional features (composition, correlation).
• Utilized embedding features from the large language model DNABERT-2 for ML model development.
• Compared performance using five-fold cross-validation and an independent validation dataset, measuring Area Under the Curve (AUC).

Main Results:

• ML models with composition and correlation features achieved average AUCs of 0.7049 and 0.7089, respectively.
• DNABERT-2 embedding features resulted in an average AUC of 0.665.
• Fine-tuned DNABERT-2 achieved an average AUC of 0.6336, indicating correlation-based features with ML outperform LLM-based approaches for this task.

Conclusions:

• Machine learning models utilizing correlation-based features demonstrate superior performance in predicting differential lncRNA localization compared to LLM-based models.
• The developed cell-line specific models and a web server are publicly available for predicting lncRNA localization.