Class similarity network for coding and long non-coding RNA classification.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Long non-coding RNAs (lncRNAs) are crucial in biological processes, necessitating accurate identification for functional studies.
• Deep learning, specifically Convolutional Neural Networks (CNNs), has advanced lncRNA identification but often overlooks inter-sample relationships.

Purpose of the Study:

• To develop a novel deep learning model that enhances lncRNA classification by directly considering relationships among samples.
• To improve upon the indirect sample relationship analysis of traditional CNNs in lncRNA identification.

Main Methods:

• Introduction of the Class Similarity Network (CSN), inspired by Siamese Neural Networks (SNNs).
• CSN directly explores relationships between input samples and samples from both the same and different classes.
• Training class-specific parameters to extract high-level features and represent class similarity.

Main Results:

• The Class Similarity Network demonstrates superiority over baseline CNNs in coding RNA and lncRNA classification.
• Achieved state-of-the-art performance on two independent test datasets.
• Validation dataset comparisons confirmed the effectiveness of the CSN approach.

Conclusions:

• The developed Class Similarity Network is effective for coding RNA and lncRNA classification.
• The model achieved high accuracy, precision, and F1-scores on multiple datasets, indicating robust performance.
• CSN offers a significant advancement in deep learning-based lncRNA identification.