EnANNDeep: An Ensemble-based lncRNA-protein Interaction Prediction Framework with Adaptive k-Nearest Neighbor...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Long non-coding RNA-protein interactions (LPIs) are crucial for biological processes.
• Existing computational methods for LPI prediction often suffer from dataset bias and limited validation strategies.
• Current models fail to effectively predict interactions for novel lncRNAs or proteins.

Purpose of the Study:

• To develop a robust computational framework, EnANNDeep, for accurate LPI prediction.
• To address limitations of existing methods by employing rigorous cross-validation across lncRNAs, proteins, and pairs.
• To systematically identify potential LPIs for a deeper understanding of molecular mechanisms.

Main Methods:

• Integration of multiple feature sources to represent lncRNA-protein pairs.
• Development of an ensemble learning framework combining adaptive k-nearest neighbor, deep neural network, and deep forest models.
• Application of soft voting to combine prediction probabilities from individual models.

Main Results:

• EnANNDeep achieved superior performance over five classical LPI prediction models across multiple cross-validation strategies.
• The model demonstrated high average AUCs (0.8660-0.9166) and AUPRs (0.8545-0.9054).
• Case study identified a potential dense linkage between SNHG10 and protein Q15717.

Conclusions:

• EnANNDeep offers a powerful and reliable approach for LPI prediction, overcoming limitations of previous methods.
• The adaptive k-nearest neighbor component allows for flexible parameter selection.
• Deep learning components effectively capture essential features of lncRNAs and proteins for accurate interaction prediction.