Tools for a Reproducible Workflow for ChIP-seq QC and Visualisation.

Thomas Carroll

Head of Bioinformatics

MRC Clinical Sciences Centre

Overview

• Higher output more complex hypotheses.
• New challenges for:
  • Quality control
  • Visualisation
  • Dynamics of epigenetic states.
  • Reproducibility and reporting.

Then

• Limited sequence depth.
• Shorter reads.
• Developing protocols.
• Single sample studies.
• Non standardised toolset.
• ChIP-seq != RNA-seq

Now

• Greater depths per lane.
• Established protocols
• Multiplexing technologies - Sample replication.
• Advanced toolsets
• Borrowing of analysis techniques between data-types

High-throughput Sequencing  technologies

Bigger studies require large scale quality control

• ChIP-seq has many sources of noise
  • Inefficient or non-specific antibodies.
  • Artefact signals.
  • Signal complexity.
• Requirement for methods to assess quality  within sample groups and across large experiments.

How to assess what is useful.

• Encode and others have established metrics of ChIP-seq quality.
  • SPP/phantomQualityTools.
  • HTSeqtools.
• Use and interpretation of these metrics varies with ChIP type.
• Such metrics may be overfitted to these big studies and dependent on the standardised processing protocols applied.

Epigenetic's data QC study

• > 500 datasets of ChIP-seq for transcription factor and epigenetic marks.
• Evaluate quality control metrics under different conditions.
• Identify importance/redundancy of QC metrics.
• Establish effect of processing on QC metrics

Carroll et al Impact of artifact removal on ChIP quality metrics in ChIP-seq and ChIP-exo data. Front Genet April 2014.

Blacklisted Regions

• Identified by Peter Park.
• DAC list created by Anshul Kundaje for Encode.
• Species specific regions of consistent ultra high signal 
• Contains repeat regions and genomic expansion.

Blacklisted Regions and the "Phantom Peak"

• Fragment length peak indicates ChIP signal strength.
• "Phantom peak" often identified in cross-correlation analysis at read length.

• Signal from blacklisted regions contribute solely to this peak

Measures of inequality of coverage (SSD)

• Standardised Standard Deviation (SSD) of ChIP-signal.
  • Measure of overall "peakiness".
  • More suited for longer marks
• SSD strongly influenced by blacklisted regions.

Library complexity and duplication levels

• Library complexity often assessed by duplication rate.

• Artefact duplicates often overestimated.

• Duplicates may arise from high efficiency and/or high depth.

ChIP-exo

• ChIP-exo offers higher resolution and greater efficiency.
• Reduced "Blacklisted" signal.
• Higher duplication in peaks.
• Cross-correlation profiles require intrepretation.

ChIPQC package

High-throughput QC of epigenetics data.

• Implemented QC learned from study.
• Designed for work on large experiments with multiple groups.
• Implemented in R and Bioconductor for ease on installation, implementation and integration with reporting

Examples ChIP-QC

• Evaluate +/- read clustering around peaks.

• Contribution of artifact signal.

• Distribution of ChIP signal across genome.

QC from low depth

(MiSeq)

• Speed of Miseq's  run allows for rapid evaluation of ChIP prior to large scale sequencing.
• Direct comparison of Hiseq vs Miseq demonstrates its ability to evaluate ChIP quality

~ 1 Million reads

Visualisation of data

• Essential step is to look at your data.

• Evaluate results in their genomic context

• Identify patterns and artefact regions.

Genome Browsing

• Genome Browsers present a linear representation of genomics data.

• Allow for integration of multiple data-types as well as inclusion of both the user's and public datasets.

• Popular browsers include UCSC (server) and IGV (local).

Integration with genome browsers

• Trackhubs and Sample information files may be created and manually curated.

• For bigger studies, the organisation of tracks in a genome browser is not a trivial task.

• Evaluation of results requires exporting, conversion and potentially uploading to FTPs.

tracktables package.

Generation and  integration of R objects and dynamic HTML reports with the IGV genome browser. 

• Allow for rapid presentation of high-throughput sequencing results in the IGV genome browser.

• Produce experiment focused IGV-linked HTML reports.

• Cast R objects into HTML tables with links to genomic locations.

• Integration with pre-existing reporting tools reinforcing reproducible research.

Building Experiment Reports

Rapid prototyping by visualisation

• Cast R/Bioconductor objects into HTML tables.

• Allows user to evaluate results in genomic context.

Dynamics of epigenetic states

• With the advent of multiplexing technologies
  • Greater number of sample groups
  • Higher replicate numbers
• More complex hypotheses.
  • From simple epigenetic mapping  -> Identifying significant changes in epigenetic states.

Borrowing RNA-seq methods for ChIP-seq data

• Differential affinity/binding of epigenetic marks or transciption factors.

• Methods implemented in RNA-seq now being used in ChIP-seq (DEseq2/limma/EdgeR/Diffbind)

• ChIPQC integrates with Diffbind.                                                        

Visualising dynamics of epigenomic states over genomic interval sets

• RNA-seq visualisation methods don't cover ChIP-seq

• Require tools to investigate shapes of signal.

• Simple tools designed for small sample numbers.

soGGi package

Summarising Over Grouped Genomic Intervals

• Accepts multiple input types (BAMs/bigWigs/Motifs) to generate profiles.

• Leverages GGplot2 to enable intuitive plotting grammar

• Implements normalisation procedures and arithmetic operations across profiles.

Profile types require different techniques

• soGGi allows for three types of plot
  • Point (signal around a genomic location).
  • Meta (signal over and around a genomic region normalised to that genomic regions length).
  • Hybrid (Point plot over edges and Meta plot over central genomic region).

Integrating Arithmetic Operations 

• Profiles can be further normalised and operated on to visualise hypotheses.

• Most common arithmetic operations and transformations can be applied between profiles.

Accounting for noise in ChIP-seq

• Technical variation in ChIP-seq is typically considered very high.

• Signal across artefact regions and duplication rates can be highly variable.

• Such defined sources of noise can inform the models used in differential binding analysis.

pol2Rates package (unpublished)

• Tools to evaluate quantitative changes in Pol2.

• Calculates normalisation factors from Pol2 signal distribution.

• Not available yet but hope for release before end of year

Comparison of Uncorrected vs Corrected

• Correction significantly improves correlation with RNA-seq data and known targets.

A Reproducible Workflow

• Created three packages to handle:
  • Quality control
  • Genome Browser integration.
  • Visualisation of signal structure over genomic regions of interest.
• Integrated with R's rMarkdown (Reporting) and packrat (Version control) packages.
• Together allows for creation of customisable and reproducible reporting for ChIP-seq.

Acknowledgements

MRC Clinical Sciences Centre

• Ziwei Liang
• Gopuraja Dharmalingham
• Sanjay Khadayate
• Yi-Fang Wang
• Matthias Merkenschlager

CRUK Cambridge Institute

• Rafik Salama
• Ines de Santiago
• Rory Sark

References

• Carroll et al Impact of artifact removal on ChIP quality metrics in ChIP-seq and ChIP-exo data. Front Genet. April 2014.
• Bioconductor (http://bioconductor.org/)
  • ChIPQC Bioconductor 2.13 (May 2014)
  • tracktables Bioconductor 2.14  (September 2014)
  • soGGi Bioconductor 3.0 (May 2015)