Prediction of LncRNA-Protein Interactions Based on Kernel Combinations and Graph Convolutional Networks

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Long non-coding RNAs (lncRNAs) and proteins form complexes regulating diverse biological processes.
• Experimental verification of lncRNA-protein interactions (LPI) is resource-intensive.
• Computational prediction of LPIs offers a scalable alternative.

Purpose of the Study:

• To develop an efficient computational framework for predicting lncRNA-protein interactions (LPI).
• To leverage sequence, expression, and functional features for enhanced LPI prediction accuracy.

Main Methods:

• Constructing kernel matrices from lncRNA and protein features (sequence, similarity, expression, Gene Ontology).
• Reconstructing kernel matrices and integrating known LPIs to form graph topology.
• Employing a two-layer Graph Convolutional Network (GCN) for representation learning.
• Ensembling multiple LPI-KCGCN variants for final predictions.

Main Results:

• The LPI-KCGCN framework achieved high performance on both balanced and unbalanced datasets.
• On a dataset with 15.5% positive samples, AUC reached 0.9714 and AUPR reached 0.9216.
• On a highly unbalanced dataset (5% positive samples), AUC was 0.9907 and AUPR was 0.9267, outperforming existing methods.

Conclusions:

• LPI-KCGCN provides a powerful and accurate computational approach for predicting lncRNA-protein interactions.
• The framework demonstrates superior performance, especially on imbalanced biological data.
• This method can accelerate the discovery of functional lncRNA-protein complexes.