Machine Learning Methods and Evaluation in RNA Modification Site Prediction
TL;DR
A review of methodological evolution of RNA modification site prediction highlights standardized benchmarking, model generalization, and multi-source information integration as key priorities for future research to promote more reliable and practical RNA modification site prediction models.
Abstract
Post-transcriptional RNA modifications influence RNA structure and function and play important roles in gene expression regulation and other cellular processes. Accurate identification of RNA modification sites is therefore essential for understanding their biological functions and diseaseassociated mechanisms. With the rapid development of machine learning methods, RNA modification site prediction has progressed from traditional machine learning to deep learning and pre-trained representation learning. This paper summarizes the methodological evolution of RNA modification site prediction. It discusses recent advances in joint modeling of multiple modification types, integrating multi-source biological information, and analyzing direct RNA sequencing data. In addition, by discussing current challenges in data quality, task definition, and evaluation frameworks, this review highlights standardized benchmarking, model generalization, and multi-source information integration as key priorities for future research to promote more reliable and practical RNA modification site prediction models.