Finding lncRNA-Protein Interactions Based on Deep Learning With Dual-Net Neural Architecture

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Identifying long noncoding RNA-protein interactions (LPIs) is crucial for understanding lncRNA functions.
• Existing computational models often suffer from prediction bias due to single-dataset evaluation and fail to predict novel interactions.
• There is a need for improved models to accurately identify LPIs.

Purpose of the Study:

• To develop a robust deep learning framework, LPI-DLDN, for identifying lncRNA-protein interactions.
• To address the limitations of existing models, including prediction bias and inability to uncover novel interactions.
• To enhance the accuracy and reliability of LPI prediction.

Main Methods:

• Collected five LPI datasets and extracted features using Pyfeat and BioTriangle.
• Employed dimension reduction and feature concatenation to create a unified feature vector.
• Designed a deep learning model with a dual-net neural architecture for classifying lncRNA-protein pairs.

Main Results:

• LPI-DLDN demonstrated superior LPI classification performance compared to six state-of-the-art methods across four cross-validations.
• Case studies suggested potential interactions between specific lncRNAs (RP11-439E19.10, RP11-196G18.22) and proteins (Q15717, Q9NUL5).
• The framework effectively integrates diverse biological features and selects optimal subsets based on importance.

Conclusions:

• LPI-DLDN offers a powerful and accurate approach for identifying lncRNA-protein interactions.
• The framework's ability to integrate various biological features and its novel deep learning architecture contribute to its effectiveness.
• LPI-DLDN advances the field by providing a more reliable tool for LPI prediction and discovery.