AttenRNA: multi-scale deep attentive model with RNA feature variability analysis

TL;DR

AttenRNA is a deep learning model that accurately classifies different RNA types, offering improved performance and generalization across species.

Contribution

Introduces AttenRNA, a multi-scale attentive model for multi-class RNA classification with strong cross-species generalization.

Findings

AttenRNA achieved 89.8% and 89.6% weighted F1 scores on validation and test sets.

The model generalized well to mouse RNA data with 83.89% and 83.38% F1 scores.

UMAP analysis confirmed the model's ability to learn discriminative RNA features.

Abstract

Accurate identification of diverse RNA types, including messenger RNAs (mRNAs), long non-coding RNAs (lncRNAs), and circular RNAs (circRNAs), is essential for understanding their roles in gene regulation, disease progression, and epigenetic modification. Existing studies have primarily focused on binary classification tasks, such as distinguishing lncRNAs from mRNAs or identifying specific circRNAs, often overlooking the complex sequence patterns shared across multiple RNA types. To address this limitation, we developed AttenRNA, a multi-class classification model that integrates multi-scale k-mer embeddings and attention mechanisms to simultaneously differentiate between various RNA classes. AttenRNA achieved high weighted F1 scores of 89.8% and 89.6% on the validation and test sets, respectively, demonstrating strong classification performance and robustness. Dimensionality reduction using Uniform Manifold Approximation and Projection further confirmed the model’s ability to learn discriminative features among RNA types. Additionally, AttenRNA exhibited strong generalization ability on cross-species data, achieving weighted F1 scores of 83.89% and 83.38% on the mouse RNA validation and test sets, respectively. These results suggest that AttenRNA offers a reliable and scalable solution for systematic RNA function analysis.

Figures