Interpretable Deep Learning Model Reveals Subsequences of Various Functions for Long Non-Coding RNA Identification.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Long non-coding RNAs (lncRNAs) are vital in biological processes and disease, but many transcripts remain unannotated.
• Experimental classification of these transcripts is costly and time-consuming.
• Existing computational tools for lncRNA identification lack predictive feature explanations.

Purpose of the Study:

• To develop Xlnc1DCNN, a computational tool for distinguishing lncRNAs from protein-coding transcripts (PCTs).
• To provide interpretable predictions by explaining the features contributing to classification.
• To enhance the accuracy and efficiency of lncRNA identification.

Main Methods:

• Utilized a one-dimensional convolutional neural network (1D CNN) architecture.
• Trained and evaluated the model on human transcript datasets.
• Incorporated a mechanism for generating prediction explanations.

Main Results:

• Xlnc1DCNN demonstrated superior performance over state-of-the-art tools, achieving high accuracy and F1-score on a human test set.
• Explanation analysis revealed distinct sequence features differentiating lncRNAs (e.g., lack of conserved regions, short functional patterns) from PCTs (e.g., conserved protein domains).
• Identified potential inconsistencies in public database annotations, including lncRNAs with protein domains or intrinsically disordered regions.

Conclusions:

• Xlnc1DCNN offers an accurate and interpretable method for classifying lncRNAs and PCTs.
• The tool's explanations provide insights into the molecular characteristics distinguishing these transcript types.
• Xlnc1DCNN aids in refining transcript annotation and understanding lncRNA function, addressing challenges posed by database inconsistencies.