TL;DR
N6-methyladenosine (m6A) is a crucial RNA modification, and numerous databases aim to catalog m6A data. However, this paper reveals significant variability in how these databases call m6A peaks due to differences in data processing pipelines, leading to potential inaccuracies. The authors provide a guide for end-users to navigate these complexities and suggest improvements for future m6A databases.
What Problem Does This Paper Address?
N6-methyladenosine (m6A) is the most abundant internal modification in eukaryotic messenger RNA (mRNA), playing a vital role in various cellular processes like RNA stability, splicing, and translation. To understand its implications in human health and disease, researchers rely on large-scale datasets and databases that collate m6A profiles. However, the methods used to generate and process this data are complex and can lead to inconsistencies. Specifically, different m6A profiling techniques (e.g., antibody-based vs. crosslinking-assisted) and varying bioinformatic pipelines for peak calling can result in substantial discrepancies in the reported m6A sites and peaks. This variability, often not explicitly detailed by database creators, can lead to inaccurate interpretations of m6A-associated biology by end-users who may not be aware of these underlying issues. This paper aims to highlight these complexities and provide guidance for researchers using m6A databases.
Methods
The study employs a comparative approach to evaluate existing m6A databases. Key methods include:
- Peak Calling Comparison: The authors re-analyzed publicly available m6A sequencing data (from NOMO-1 and H1299 cell lines) using different database pipelines to assess the variability in peak calling. This involved standard bioinformatics steps like read trimming, alignment, and filtering, followed by peak calling using algorithms such as MACS2 and exomePeak.
- Protocol Comparison (Early vs. Refined): They directly compared m6A data generated by an "early" m6A-seq protocol (requiring high RNA input) and a "refined" RIP-seq protocol (requiring low RNA input) in the A549 cell line. This comparison aimed to identify potential inaccuracies introduced by older protocols.
- Validation with High-Confidence Sites: m6A peaks identified by the early and refined protocols were validated against a database of high-confidence m6A sites derived from miCLIP experiments.
- iM6A Model for Probability Scoring: An intelligent m6A (iM6A) model, a deep learning approach, was used to calculate the probability of a called peak representing a true m6A site. This helped to assess the reliability of peaks identified by different protocols.
- SNP Analysis: The paper also examined how single nucleotide polymorphisms (SNPs) associated with m6A sites are reported in databases, focusing on their potential impact on m6A modification and disease pathogenicity.
Datasets
The study utilized publicly available datasets for its analysis:
- NOMO-1 m6A-seq data: (GSE87190) used for comparing peak calling across multiple databases.
- H1299 meRIP-seq data: (GSE76367) used for replicating the peak calling comparison.
- A549 cell line data: Generated using both an early m6A-seq protocol (GSE76367) and a refined RIP-seq protocol (GSE116002) for direct comparison of protocol sensitivity and accuracy.
- m6A-Atlas database: Used for validating m6A peaks against high-confidence sites.
The paper does not introduce a new benchmark dataset but rather uses existing public data to evaluate current databases.
Key Insights and Ideas
- "Garbage In, Garbage Out" for m6A Databases: The core message is that the way m6A data is processed significantly impacts the final output, and many databases don't adequately disclose these details. This means users might be drawing conclusions from potentially unreliable data.
- Early Protocols are Problematic: The older m6A-seq protocols are less sensitive and prone to false positives compared to newer refined RIP-seq methods. Relying on data from these older methods without careful filtering can lead to misinterpretations.
- Peak Calling Pipelines Matter More Than Algorithms: Surprisingly, the study found that the choice of database pipeline (which includes filtering and processing steps) had a greater impact on peak call concordance than the specific peak-calling algorithm used.
- SNPs and m6A: A Complex Relationship: Databases often report SNPs near m6A sites, but it's unclear if these SNPs directly affect m6A. Furthermore, when SNPs cause amino acid changes, it's difficult to attribute disease pathogenicity solely to altered m6A.
Key Results
- High Variability in Peak Calling: Comparing peak calls from the same accession across different databases revealed substantial variability. For instance, the number of reported peaks could vary by nearly threefold between databases.
- Low Reproducibility: The overlap between peak calls from different database pipelines was generally low, with an average of only 24.2% of peak calls being highly reproducible across pipelines.
- Refined Protocols are More Sensitive: In the A549 cell line, the refined RIP-seq protocol detected three times more m6A peaks than the early m6A-seq protocol, with nearly 78% of RP peaks not detected by EP.
- iM6A Scores Highlight Protocol Differences: The iM6A model showed that peaks identified by the refined protocol generally had higher m6A deposition probabilities, suggesting greater accuracy.
- SNP Reporting Issues: Databases often include SNPs located far from m6A motifs, and they don't distinguish between synonymous and non-synonymous SNPs, complicating the interpretation of their functional impact.
Limitations and Future Work
Limitations:
- Limited Availability of Base-Resolution Data: While base-resolution methods are improving accuracy, they are still underrepresented in current databases due to the difficulty in generating such data.
- Lack of Customization: Most databases do not allow users to adjust the stringency of peak calling parameters, limiting their ability to tailor analyses.
- Inclusion of Potentially Unreliable Data: Databases often include data from older, less reliable m6A profiling methods without sufficient filtering.
Future Work:
- Improved Database Transparency: Future databases should explicitly disclose the m6A profiling methods and processing pipelines used for each sample.
- User-Adjustable Stringency: Allowing users to modify peak calling stringency would enhance data reliability.
- Re-analysis with Modern Tools: Datasets should be re-analyzed using superior tools like TRESS to account for biological variance and improve accuracy.
- Incorporation of iM6A Model: Integrating the iM6A model into databases would provide a quantitative measure of m6A site probability.
- Smarter SNP Reporting: Databases reporting m6A-associated SNPs should filter for synonymous vs. non-synonymous SNPs and focus on those within confirmed DRACH motifs.
- Development of Base-Resolution Databases: As base-resolution techniques mature, databases should prioritize incorporating this higher-resolution data.
Summary updated