Predicting RNA modification with Deep learning models
University of Liverpool · 01.01.2027
DOI: 10.17638/03182394 ↗Kurz gefasst
Fachartikel in der Zeitschrift University of Liverpool, erschienen am 01.01.2027. Verfasst von Yiyou Song.
Abstract (Originalsprache)
RNA modifications or RNA editing refers to the modification or change of an RNA molecule after transcription. RNA modifications influence countless vital cell-level biological processes including rRNA folding, translation, cell structure, pre-mRNA splicing, and RNA degradation. As RNA modification is involved in many vital biological processes, identifying the modification site and modification status of the sites would be helpful for a deeper understanding of these biological processes. There are already many accurate biological methods for identifying modification sites. However, current biological methods share problems including expensiveness, time-consuming, and laborious. To assist biologists, computation models that can predict the site or modification status are created. Although not perfect, computational models are cost-free, labour-free, and instantaneous. This makes them more applicable in fields such as drug design where there are infinitely many possible candidates and is impossible to test every one of them. Computation models include machine learning and deep learning approaches. Machine learning approaches require high domain knowledge and feature selection, while deep learning requires large amounts of data. Although early models are mainly traditional machine learning based, increasingly more models are using deep learning due to its high performance when there are large amounts of data available. More important than their obvious differences in architecture, computational models can be categorized based on their tasks. In addition to earlier models that only predict sites for an organism in general, models can also target detailed conditions such as cancer cell lines or tissues. This thesis focuses on improving the performance and versatility of computation models in this field. Specifically, I broaden the applicability of condition-specific model predictions. I constructed two models: AdaptRM and ExpressRM. AdaptRM is a multitask model that can predict multiple types of RNA modification on multiple organisms upon a low-resolution region. In AdaptRM, we proposed a novel method to solve the dynamic sequence length problem. Specifically, using an adaptive pooling layer in the centre of the CNN model to create a local feature extractor and a global feature extractor. Combining our improved model architecture with multi-task learning. AdaptRM reached SOTA performance on both single-base-resolution and low-resolution datasets. Although AdaptRM reached SOTA performance, similar to all current condition-specific models, they encounter problems when transferring to other datasets. To address this problem, we proposed ExpressRM, the first model that can predict untrained conditions (zero-shot prediction). ExpressRM solves this problem by taking transcriptome-related features (gene expression, relative position, etc) as an additional input. Through this, the model is able to treat a condition as a continuous vector in latent space instead of as a discrete point that is completely unrelated to a different condition. ExpressRM reached superior performance compared with the previous SOTA model AdaptRM even when comparing ExpressRM’s zero-shot prediction upon untrained condition with AdaptRM’s prediction upon trained condition.
Schlagwörter
Zitieren
APA
Song, Y. (2027). Predicting RNA modification with Deep learning models. University of Liverpool. https://doi.org/10.17638/03182394
Vancouver
Song Y. Predicting RNA modification with Deep learning models. University of Liverpool. 2027. doi:10.17638/03182394
BibTeX
@article{Song2027,
author = {Song, Yiyou},
title = {Predicting RNA modification with Deep learning models},
journal = {University of Liverpool},
year = {2027},
doi = {10.17638/03182394}
}
EUVIMED