Rna修飾部位を予測するための深層学習と機械学習戦略に関する研究
Year: 2025
Type: dissertation
Abstract: RNA modifications play critical roles in regulating RNA stability, structure, and function, influencing various biological processes. Among nearly 300 known chemical modifications, dihydrouridine (DHU or D) is commonly found in tRNA, mRNA, and snoRNA, which are closely associated with disease pathogenesis and various biological processes in eukaryotes. 2¢-O-methylation (2-OM or Nm) is another widespread RNA modification observed in various RNA types like tRNA, mRNA, rRNA, miRNA, piRNA, and snRNA, which occurs on the ribose sugar of RNA and contributes to stability and translational control. N4-acetylcytidine (ac4C) is another novel and highly conserved chemical modification observed in both eukaryotic and prokaryotic tRNA, rRNA, and mRNA, that is involved in maintaining translational fidelity and enhancing mRNA stability. Understanding these modifications is essential for elucidating post-transcriptional gene regulation and its potential implications in human health and disease. To comprehend its modification mechanisms and potential epigenetic regulation, it is necessary to accurately identify the modification (DHU, 2-OM, ac4C) sites. Traditional experimental methods for detecting RNA modifications, including DHU, 2-OM, and ac4C, have significantly advanced our understanding of RNA biology. However, these methods are often labor-intensive, technically demanding, time-consuming, and costly, particularly when aiming for single-nucleotide resolution across large transcriptomes. Given these challenges, there is a growing demand for computational prediction models to complement and accelerate the discovery of RNA modification sites. Despite progress, existing computational approaches still face several limitations in prediction performance. For instance, DHU site predictors often suffer from data redundancy due to duplicate samples, limited generalizability to independent datasets, and model overfitting. In the case of 2-OM, many predictors are developed using a single type of RNA (e.g., mRNA or rRNA) or target only specific nucleotide modifications (Am, Gm, Cm, or Um), and are often trained on relatively small datasets. For ac4C, current predictors show low accuracy on independent test datasets. Moreover, several machine learning methods and feature encoding strategies remain unexplored, limiting the full potential of prediction models for these RNA modifications. To address these challenges, in this study, we proposed three cutting-edge predictors named Stack-DHUpred, Meta-2OM, and GAPred-ac4C, which can accurately identify DHU, 2-OM, and ac4C, respectively. In Stack-DHUpred, we systematically evaluated six classifiers across 11 RNA sequence features, resulting in the development of 66 baseline (single-feature) models. These baseline models were then combined using logistic regression in a stacked ensemble framework. The optimal subset of baseline models was selected to construct the final stacked model, named Stack-DHUpred. This model achieved an accuracy greater than 0.77 and an AUC above 0.87 on the independent dataset, outperforming existing tools on both training and independent datasets. Meta-2OM utilized a meta-learning approach that considered eight conventional machine learning algorithms and eighteen different feature encoding algorithms that cover physicochemical, compositional, position-specific, and natural language processing information. The predicted probabilities of 2-OM sites from the 144 baseline models are then combined and trained using logistic regression to generate the optimal prediction. On the independent test set, Meta-2OM achieved an overall accuracy above 0.87 and AUC above 0.93, demonstrating superior performance compared to the existing predictors. GAPred-ac4C was built by combining probability scores from 120 ML and DL-based single-feature models. Using forward feature selection, a genetic algorithm-based meta-model leveraging six top-performing features from CNNBiGRU, CNNBiLSTM, CNNAtt, and LGBM achieved optimal results. GAPred-ac4C presents the AUC values of 0.893 and 0.902 for training and independent datasets, respectively, which outperformed all existing state-of-the-art methods and demonstrated the superiority of the model. Therefore, the proposed approaches significantly improved the prediction performance, and we believe that these can be extended to other sequence-based function prediction problems, including enhancer prediction, peptide therapeutic function prediction, and post-translational modification sites prediction. To facilitate its use, two user-friendly web servers and standalone programs have been developed and are freely available at http://kurata35.bio.kyutech.ac.jp/Stack-DHUpred, http://kurata35.bio.kyutech.ac.jp/Meta-2OM/, https://github.com/kuratahiroyuki/Stack- DHUpred, and https://github.com/kuratahiroyuki/Meta-2OM.
DOI: 10.18997/0002002010
Author: Md Harun-Or-Roshid,
Language: English
Topic: RNA modifications and cancer
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7110859856
Cites:
Cited by:
Related to: