ChIP-Seq Bioinformatics

For ChIP-Seq Workshop at Newcastle University, 11.05.16. Slides focussed on quality issues with ChIP-Seq samples, not on mapping and peak calling.

Transcript

1. ChIP-Seq Bioinformatics Simon Cockell Bioinformatics Support Unit bsu.ncl.ac.uk @sjcockell
2. ChIP-Seq Workflow • Cross-link proteins to DNA • Usually in-cell • Isolate DNA and shear • Sonication for ‘random’ shearing • Immunoprecipitate protein of interest • Reverse cross-linking • Purify DNA • Sequence From: 10.1093/bib/bbq068
3. Peak calling • Read position randomly selected from genome is Poisson process* • Deviations from Poisson define peak regions • Reads on either strand will flank the true binding position * Not really, due to multiple biases – but close enough… From: 10.1038/nmeth.1246
4. Different types of peak • Peaks can be broad, sharp or mixed • Different approaches to calling for each • Broad peaks present a particular challenge Mixed Sharp Broad Broad Broad Control From: 10/1093/bib/bbw023
5. Issues with ChIP-Seq experiments • Many sources of problems – most not ‘fixable’ with bioinformatics • Major problems include: • Variable mappability of genome • Poor signal:noise ratio (S/N) from ChIP • Issues with low library complexity • Inappropriate or insufficient controls (and replicates)
6. Mappability • Typical to use uniquely mappable reads • Causes obvious problems with regions of low mappability (repeats etc) • If these regions are of interest, a specific strategy needs to be taken • Paired reads, long reads, include multi-mapped in analysis? • ENCODE black-list regions of particular problems • Available for a few organisms • https://sites.google.com/site/anshulkundaje/projects/blacklists
7. Signal:Noise • Depends on experimental factors • Strength/specificity of Ig binding • Difficult to determine until after sequencing • Evaluated by number and strength of peaks: FRiP (fraction of reads in peaks): Npeak /Nnonred NSC (normalised strand coefficient)  =  Cfrag/ Cmin RSC (relative strand correlation) = (Cfrag − Cmin )/(Cread − Cmin ) • C = correlation (Pearson) between mapped read densities of positive and negative strands (y-axis) with shifting one strand (x- axis) • The ENCODE consortium recommends an NSC  ≥ 1.05 and an RSC  ≥ 0.8 for typical TFs (sharp mode)
8. Signal:Noise – cross-correlation From: 10.3389/fgene.2014.00075
9. Library Complexity • Low library complexity often occurs when samples are prepared from a small amount of starting materials • Measured by the non-redundant fraction (NRF) NRF = Nnonred /Nall • Nnonred = reads mapped to same genomic coordinates T times or fewer • In practice, T is usually 1 • The ENCODE consortium endorses an NRF  > 0.8 for 10 million reads (T  =  1)
10. Controls and Replicates • Control is usually experimental without IP step • Known as input • Other controls often requested by reviewers (so consider up front) • RNAi knock-down of IP target • Non-binding control (IgG, GFP) • Also note absolute requirement by some journals for n>=2 • Agreement between replicates another key QC test • Test with Jaccard statistic (raw reads) • Or correspondence of peak profiles
11. Sequencing depth considerations • Depends on a number of experimental factors • Difficult to determine a priori • Can use post hoc saturation analysis • Systematically drop out data and look for plateau point of peaks detected • Especially for histone modifications, there doesn’t appear to be a practical saturation point • More reads == more peaks • RULES OF THUMB • For sharp peaks, 10m uniquely mappable reads per replicate • For broad peaks. 40-50m reads as ’practical minimum’
12. Best peak callers? • No real answer to this • Very difficult to assess • What is ‘truth’ • No bench-marked or synthetic datasets • Poisson model: SICER, CCAT • Negative Binomial: CisGenome, BayesPeak • Zero-inflated Negative Binomial: MOSAiCS, ZINBA • Local Poisson: MACS • Expectation Maximisation: GPS, PICS • +ve & -ve Strand Equivalence: SISSRs, Peakzilla, Q • Use replicate, differential info: PePr, JAMM
13. Best peak callers? “The appropriate method depends on the species, sample conditions and target proteins. Even though there is no clear consensus on which is best, the latest and widely used programs may be satisfactory for our needs.” From: 10/1093/bib/bbw023
14. Reliability of results • Methods for assessing are sparse • Even for TF with canonical motif, condition- specific binding independent of motif can and does occur • The irreproducible discovery rate (IDR) assesses the rank consistency of common peaks between two replicates • One ’failed’ sample will become bottleneck • So bad replicates must be discarded • Makes experiments prone to batch effects • Poor quality samples basically unrecoverable