Multiple RNA Methylation Prediction Using Mixture of Experts
Journal of Intelligent Decision Making and Information Science · 15.09.2026 · 3 (19s) · 1759-1785
DOI: 10.17762/jidmis.v3.6370 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Intelligent Decision Making and Information Science, online erschienen am 15.09.2026. Verfasst von Sajid Shah. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Post-transcriptional RNA alterations play crucial roles in regulating various biological processes, such as RNA stability, splicing, translation, immunological tolerance, and degradation. Accurate identification of those sites, such as m5C, m6A, m1A, m5U, m6Am, m7G, Ψ, I, Am, Cm, Gm, and Um is therefore extremely important for understanding their regulatory mechanisms. However, experimental identification (wet-labe mapping) of RNA methylations is time-consuming and expensive; therefore, computational methods got attention of researchers but most of them handle single modification type. Existing computational techniques face various limitations, such as handling high-dimensional RNA sequencing data without proper feature engineering, particularly utilizing RNA structure representation. Deep neural networks have proved their worth in various domains. When it comes to RNA sequence data, convolutional neural networks (CNNs) largely capture local motif patterns but struggle with long-range dependencies, whereas Transformer-based models can learn sequential information, preserving long-term dependencies, but may not incorporate structural information. To address these issues, we propose a mixture-of-experts (MoE) deep learning system that integrates complementary models of RNA biology. The proposed architecture comprises six specialized experts: (i) two CNN-based experts designed to capture local sequence motifs and short-range dependencies, (ii) a Transformer-based expert that models long-range contextual interactions across entire RNA sequences, (iii) an RNA structure–based expert that explicitly incorporates secondary structural information, (iv) a k-mer expert (dealing with string or primary structure of RNA), and (v) an RNA language model. All training, tuning and model selection use only training data. The motif-level, sequence-level, and structure-level features were utilized within an end-to-end framework to increase the accuracy as well as generalizability of the proposed model. The model is evaluated on independently held-out testing datasets. On these test sets our model outperformed stat-of-the-art (SOTA) techniques in 10/12 types. In the remaining 2 types our results are closed to tie. For A-to-I, we have added 1kb genomic context which raised the AUROC from 65.5 to 74.0%. If we add this last setup for A-to-I, then the wining ratio of our model increased to 11/12.
Zitieren
APA
Shah, S. (2026). Multiple RNA Methylation Prediction Using Mixture of Experts. Journal of Intelligent Decision Making and Information Science, 3(19s), 1759-1785. https://doi.org/10.17762/jidmis.v3.6370
Vancouver
Shah S. Multiple RNA Methylation Prediction Using Mixture of Experts. Journal of Intelligent Decision Making and Information Science. 2026;3(19s):1759-1785. doi:10.17762/jidmis.v3.6370
BibTeX
@article{Shah2026,
author = {Shah, Sajid},
title = {Multiple RNA Methylation Prediction Using Mixture of Experts},
journal = {Journal of Intelligent Decision Making and Information Science},
year = {2026},
volume = {3},
number = {19s},
pages = {1759--1785},
doi = {10.17762/jidmis.v3.6370}
}
EUVIMED