TL;DR
This study provides the first comprehensive, genome-wide analysis of long non-coding RNA (lncRNA) stability in human cells, revealing that lncRNAs are generally less stable than mRNAs. The authors identified key factors influencing lncRNA stability, including exon number, RNA secondary structure, and interactions with proteins and miRNAs, and developed a deep learning model to predict lncRNA half-lives.
What Problem Does This Paper Address?
Transcript stability, a crucial determinant of gene expression levels, has been extensively studied for messenger RNAs (mRNAs). However, the factors governing the stability of long non-coding RNAs (lncRNAs) – a class of transcripts with diverse regulatory roles but no protein-coding potential – remain largely unknown. Understanding lncRNA stability is essential for deciphering their functions and regulatory mechanisms, especially given their involvement in various biological processes and diseases. This paper aims to fill this knowledge gap by systematically investigating the genome-wide stability of human lncRNAs and identifying the molecular features that influence their degradation.
Methods
The researchers employed a genome-wide RNA-sequencing (RNA-seq) approach in human lung adenocarcinoma cells (A549). To determine transcript half-lives, they inhibited RNA transcription using actinomycin D and collected samples at multiple time points. This allowed for the construction of RNA half-life datasets. Key analytical methods included:
- RNA-seq and Data Processing: Standard RNA-seq protocols were used, followed by rigorous quality control, read mapping, transcript reconstruction, and expression quantification (FPKM).
- Half-life Calculation: A novel sampling strategy was developed to calculate transcript half-lives by generating multiple time series from replicate experiments and averaging the results. This approach aimed to improve robustness.
- Statistical Analysis: Spearman correlation was used to assess relationships between half-lives and various features. Kolmogorov-Smirnov tests were employed to compare distributions between different transcript populations.
- Bioinformatics Tools: RNAFOLD was used for predicting RNA secondary structures. miRanda was utilized for predicting miRNA-lncRNA interactions. PANTHER was used for Gene Ontology (GO) annotation.
- Deep Learning Regression: To capture potential non-linear relationships between transcript half-lives and influencing factors, a deep learning model (using Keras and TensorFlow) was developed.
Datasets
The study utilized data generated from human lung adenocarcinoma (A549) cells. The primary dataset consists of genome-wide RNA-seq data from multiple time points after transcription inhibition, leading to the construction of RNA half-life datasets for:
- 33,285 lncRNAs (Inc-human)
- 24,710 mRNAs (m-human)
Additionally, the study incorporated publicly available mouse lncRNA and mRNA datasets for comparative analysis.
Key Insights and Ideas
- lncRNAs are generally less stable than mRNAs: The study found that lncRNAs have shorter half-lives on average compared to mRNAs, suggesting a more dynamic regulatory role.
- Exon number matters, but differently for lncRNAs and mRNAs: Unlike mRNAs, where more exons correlate with longer half-lives, human lncRNAs with fewer exons tended to be more stable. This highlights a divergence in stability regulation between these transcript types.
- Cellular location has a limited direct impact on stability: While initial enrichment analyses suggested nucleus-localized transcripts were less stable, direct half-life calculations for common nuclear and cytoplasmic transcripts showed no significant difference, implying that intrinsic transcript properties are more dominant than location.
- miRNA and protein interactions have opposing roles: miRNAs tend to promote lncRNA stability but degrade mRNAs, while protein-RNA interactions generally promote lncRNA stability but reduce mRNA stability. This suggests distinct regulatory pathways for each RNA class.
- Non-linear relationships are crucial: Linear regression models failed to accurately capture the complex interplay between transcript half-lives and influencing factors. A deep learning approach was necessary to reveal these non-linear relationships.
Key Results
- Half-life distributions: Approximately 80% of lncRNAs had half-lives less than 5 hours, while only about 60% of mRNAs fell into this category. The average half-life was 3.96 h for lncRNAs and 6.35 h for mRNAs.
- Exon number impact: Inc-human1 (one exon) had an average half-life of 4.16 h, while Inc-human2 (more than one exon) had 3.85 h, indicating shorter half-lives for single-exon lncRNAs. For mRNAs, m-human1 (one exon) averaged 4.81 h, and m-human2 (more than one exon) averaged 6.38 h, showing longer half-lives for multi-exon mRNAs.
- miRNA-lncRNA interactions: 32 out of 37 analyzed miRNAs were found to significantly promote lncRNA stability.
- miRNA-mRNA interactions: A significant negative correlation was observed between mRNA half-lives and the number of potential miRNA binding sites, indicating mRNA degradation.
- Protein-lncRNA interactions: A positive correlation was found between lncRNA half-lives and protein-binding ability, suggesting promotion of stability.
- Protein-mRNA interactions: A significant negative correlation was observed between mRNA half-lives and protein-binding ability, indicating reduction of stability.
- Deep learning model performance: The deep learning model achieved Spearman correlation coefficients between predicted and real half-lives ranging from 0.727 to 0.807 for human lncRNAs and mRNAs, indicating a reasonable ability to predict stability.
Limitations and Future Work
- Limited mRNA population in m-human1: The study notes that the small number of mRNAs in the m-human1 population (one exon) might limit the reliability of conclusions drawn from it, suggesting a need for more data in this category.
- Need for better prediction models: While the deep learning model showed promise, the authors acknowledge that further efforts are needed to adjust network structures or incorporate new features to achieve a more accurate quantitative prediction model for lncRNA and mRNA half-lives.
- Focus on A549 cell line: The study was conducted in a single cell line (A549). Future work could explore lncRNA stability across different cell types and conditions to assess the generalizability of the findings.
- Investigating Inc-human2 and m-human1: The authors explicitly state that future work will focus on the stability analysis of lncRNAs from Inc-human2 (multi-exon lncRNAs) and mRNAs from m-human1 (single-exon mRNAs), suggesting these areas require further exploration.
Summary updated