Cross-linked RNA Secondary Structure Analysis using Network Techniques

Irena Fischer-Hwang, Zhipeng Lu, James Zou, Tsachy WeissmanbioRxiv 2019 · DOI

TL;DR

This paper introduces CRSSANT, an automated computational pipeline that uses network analysis to interpret RNA secondary structure data from the PARIS cross-linking assay. CRSSANT significantly speeds up the analysis of complex RNA structures, which was previously a laborious manual process, by efficiently clustering sequencing reads and predicting potential base-pairing interactions.

What Problem Does This Paper Address?

RNA's ability to form complex 3D structures is crucial for its diverse functions, ranging from carrying genetic information to regulating gene expression. Understanding these structures, particularly the base-pairing patterns that form secondary structures, is key to deciphering RNA function. While biochemical methods like the PARIS (psoralen analysis of RNA interactions and structures) assay have emerged to map these interactions in vivo, analyzing the resulting sequencing data has been a bottleneck. The PARIS assay generates "gapped reads" that represent paired RNA segments, but manually clustering these reads and inferring structures is time-consuming and prone to error. This paper addresses the need for an automated, efficient, and accurate method to process and interpret PARIS data, thereby enabling a more comprehensive understanding of RNA structural dynamics.

Methods

The CRSSANT pipeline integrates network analysis and RNA folding software to automate the interpretation of PARIS data. The core steps involve:

  1. Data Pre-processing: Raw sequencing reads (in SAM format) are converted into a format representing read arms (start and stop positions). This step also handles "chiastic" reads, which arise from unusual ligation events.
  2. Network Representation: Reads are modeled as vertices in a graph, and edges are drawn between reads if they exhibit sufficient overlap in both their left and right arms, based on an overlap threshold (to). This weighted graph captures similarity between reads.
  3. Network Clustering (DG Assignment): The pipeline employs two clustering methods to group similar reads into "duplex groups" (DGs):
  4. Cliques-finding: Identifies fully connected subgraphs (cliques) within the network. This method is strict and can lead to discarding many reads.
  5. Spectral Clustering: A more flexible method that uses eigenvalues and eigenvectors of graph matrices to partition the graph into clusters. This is the default method in CRSSANT due to its efficiency and performance.
  6. DG Filtering and Analysis: Low-quality DGs (e.g., single-read DGs) and duplicate reads are removed. Biologically relevant attributes of the remaining DGs are calculated, such as read count and coverage.
  7. SG Assembly and Analysis: DGs are further processed to form "stem groups" (SGs) by filtering reads based on arm length percentiles. The potential of the sequences within SGs to form stable RNA stems is then tested using the ViennaRNA package. Stems with negative free energy are considered stable.
  8. Output Generation: The pipeline generates various output files (e.g., BED files) that visualize predicted base pairs, stem structures, and other relevant attributes, facilitating downstream analysis.

Datasets

The CRSSANT pipeline was evaluated on several datasets of PARIS reads from human gene sequences, including:

  • RNA component of mitochondrial RNA-processing endoribonuclease (RMRP)
  • Ribonuclease P RNA component H1 (RPPH1)
  • Ribosomal RNA (rRNA) subunits (18S, 5.8S, 28S)
  • Small nuclear ribonucleic acids (snRNAs) (U2, U4, U5, U6, U4atac, U6atac)
  • Small nucleolar RNA U3 (snoRNA U3)

These datasets were chosen because they have existing, manually-curated DG assignments and known secondary structures, allowing for validation of CRSSANT's performance.

Key Insights and Ideas

  • Network analysis is a powerful tool for RNA structure: By representing RNA reads as nodes in a network and defining edges based on read similarity, complex relationships within the data can be effectively captured and analyzed.
  • Automating manual analysis is crucial: The PARIS assay generates rich data, but its manual interpretation is a significant bottleneck. CRSSANT demonstrates that computational automation can unlock the full potential of such data.
  • Spectral clustering offers a good balance: While cliques-finding is conceptually straightforward, spectral clustering proved to be more efficient and less prone to discarding valuable data, making it a suitable default for the pipeline.
  • Default parameters are a starting point, not an endpoint: While CRSSANT provides default parameters, the authors emphasize that user intuition and experimentation are still valuable for fine-tuning outputs and interpreting biological significance, especially for discovering novel structures.
  • PARIS data can reveal novel RNA structures: The pipeline's ability to identify base pairs not present in known structures suggests that the PARIS assay, coupled with CRSSANT, can uncover previously undetected RNA structural motifs and interactions.

Key Results

  • Speed and Efficiency: The CRSSANT pipeline is fast and efficient. DG clustering took less than four minutes for all datasets, and SG generation took less than half a minute.
  • Validation: CRSSANT's DG assignments were validated against manually-produced grouping maps, showing good agreement.
  • Structure Prediction: The pipeline successfully predicted RNA stem structures, with a base pair recall ranging from under one-tenth to nearly three-quarters compared to known base pairs.
  • Novelty: A significant number of novel base pairs were identified by CRSSANT, suggesting its utility in discovering new RNA structural elements.
  • Parameter Optimization: The study systematically tested various clustering parameters, leading to the selection of spectral clustering with specific default values (to=0.5, neig=10, teig=5) that generally performed well across datasets.

Limitations and Future Work

  • User Interpretation Required: Despite automation, users still need to apply their intuition to fine-tune parameters and interpret the biological significance of the generated outputs. The pipeline is designed for customization.
  • Default Parameter Limitations: The wide range of optimal p values (percentile cutoff for arm length) suggests that choosing a single default threshold a priori is difficult. The pipeline outputs multiple sets of results to allow users to select the most relevant ones.
  • Brute-Force Testing: The validation process involved extensive testing, which was necessary due to the limited availability of manually-curated DG assignment sets and known structures.
  • Future Work: The authors suggest future improvements could include:
  • Parameter tuning with additional datasets.
  • Adding the ability to analyze extremely large read datasets with sampling.
  • Implementing methods to further decrease runtime and improve operational efficiency.

Summary updated

Abstract

Abstract Next generation sequencing and biochemical cross-linking methods have been combined into powerful tools to probe RNA secondary structure. One such method, known as PARIS, has been used to produce near base-pair maps of long-range and alternative RNA structures in living cells. However, the procedure for generating these maps typically relies on laborious manual analysis. We developed an automated method for producing RNA secondary structure maps using network analysis techniques. We produced an analysis pipeline, dubbed cross-linked RNA secondary structure analysis using network techniques (CRSSANT), which automates the grouping of gapped RNA sequencing reads produced using the PARIS assay, and tests the validity of secondary structures implied by the groups. We validated the clusters and secondary structures produced by CRSSANT using manually-produced grouping maps and known secondary structures. We implemented CRSSANT in Python using the network analysis package NetworkX and RNA folding software package ViennaRNA. CRSSANT is fast and efficient, and is available as Python source code at https://github.com/ihwang/CRSSANT .
BibTeX
@misc{fischerhwang2019cross,
  author = {Irena Fischer-Hwang and Zhipeng Lu and James Zou and Tsachy Weissman},
  title = {{Cross-linked RNA Secondary Structure Analysis using Network Techniques}},
  year = {2019},
  doi = {10.1101/668491},
  howpublished = {bioRxiv},
}
Figure