Deciphering 3'UTR Mediated Gene Regulation Using Interpretable Deep Representation Learning.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Messenger RNA 3' untranslated regions (3'UTRs) contain critical cis-regulatory elements influenced by functional and evolutionary constraints.
• These constraints resemble linguistic grammars, suggesting natural language processing (NLP) techniques can model RNA sequences.

Purpose of the Study:

• To introduce 3UTRBERT, an attention-based language model applying Bidirectional Encoder Representations from Transformers (BERT) to 3'UTR analysis.
• To evaluate 3UTRBERT's efficacy in downstream tasks including RBP binding site identification, m6A RNA modification site detection, and RNA sub-cellular localization prediction.

Main Methods:

• Pre-training 3UTRBERT on aggregated human mRNA 3'UTR sequences in a task-agnostic manner.
• Fine-tuning the pre-trained model for specific sequence labeling tasks.
• Utilizing the self-attention mechanism for visualizing sequence element relationships.

Main Results:

• 3UTRBERT demonstrated superior performance compared to contemporary methods across various 3'UTR analysis tasks.
• The model effectively identified regulatory regions by visualizing semantic relationships within sequence elements.
• Benchmark results confirmed the model's capability in predicting RBP binding sites, m6A sites, and RNA localization.

Conclusions:

• 3UTRBERT serves as a foundational tool for analyzing sequence labeling tasks in 3'UTR research.
• The model enhances the decipherability of post-transcriptional regulatory mechanisms through advanced NLP techniques.
• 3UTRBERT's self-attention mechanism provides valuable insights into RNA sequence function and regulation.