Interpretation of RNA Universe and Coding Potential Using IntRNA

TL;DR

IntRNA is a deep learning framework that improves the interpretation of RNA's coding potential and classification of RNA types.

Contribution

IntRNA introduces a novel dual-path model and image-like RNA sequence representation for enhanced RNA classification and interpretability.

Findings

IntRNA outperforms existing methods in RNA classification benchmarks.

Long-distance nucleobase pair interactions are critical for determining coding potential.

IntRNA provides interpretable insights into RNA structure and function.

Abstract

The interpretation of RNA universe and coding potential are long‐standing issues in modern RNA studies, and three crucial questions remain unanswered: a) how to detect and interpret the coding potential of RNA, b) how to annotate the sophisticated taxonomy of the sncRNAs, and c) how to successfully distinguish between circular and linear lncRNAs. In this study, a multi‐channel deep learning framework, IntRNA, is thus constructed to interpret RNA universe and coding potential. First, a large number of RNA encoding features are proposed, which dramatically enlarged the available feature space. Second, a method realizing image‐like representation of RNA sequences is developed to describe the intrinsic correlation among the encoding features generated above. Third, a dual‐path model is constructed, which consistently performed the best among existing methods in various benchmarks. IntRNA’s interpretability is also validated by analysis, and all source codes are accessible at: https://idrblab.org/intrna/ and https://github.com/idrblab/intrna.

This study constructs a multi‐channel deep learning framework IntRNA to interpret RNA universe and coding potential. IntRNA consistently performed the best among existing methods in various benchmarks. Moreover, IntRNA’s interpretability is also validated by real‐world analysis, which found some features indicating long‐distance contact between nucleobase pair to be critical in determining coding potential.

Figures