LPItabformer: Enhancing generalization in predicting lncRNA-protein interactions via a tabular Transformer

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Long-noncoding RNAs (LncRNAs) are key regulators in biological processes and disease.
• Accurate prediction of lncRNA-protein interactions (LPIs) is essential for elucidating lncRNA functions and pathogenic mechanisms.
• Existing computational LPI prediction methods often exhibit poor generalization due to data leakage and ignored biological properties.

Purpose of the Study:

• To develop a robust computational framework for predicting LPIs with enhanced generalization capabilities.
• To address the generalization flaws observed in current LPI prediction methods.
• To improve the accuracy and reliability of LPI predictions.

Main Methods:

• Proposed LPItabformer, a tabular Transformer framework specifically designed for LPI prediction.
• Incorporated a domain shifts with uncertainty (DSU) module to enhance model generalization.
• Utilized protein clusters during data splitting to identify and mitigate data leakage issues.

Main Results:

• LPItabformer effectively alleviates generalization challenges stemming from data biases and protein binding pattern preferences.
• The model demonstrated superior robustness and generalization performance on human and mouse LPI datasets compared to state-of-the-art methods.
• Successfully validated the capability of LPItabformer in predicting novel LPIs.

Conclusions:

• LPItabformer offers a significant advancement in computational LPI prediction, overcoming limitations of existing methods.
• The framework provides a more reliable tool for researchers investigating lncRNA functions and disease mechanisms.
• The developed model shows promise for discovering new LPIs and advancing our understanding of lncRNA biology.