ChIP-Seq - the long version

Analysis of ChIP-Seq Data
Biological Sequence Analysis
BNFO 691/602 Spring 2014
Mark Reimers
What Are the Questions?
• Where are histone modifications?
• Where do TFs bind to DNA?
• Where do miRNAs or RNABPs bind to 3’
• How different is binding between
Why ChIP-Seq?
• ChIP-Seq is ideal (and is now the standard
method) for mapping locations where
regulatory proteins bind on DNA
– Typically ‘only’ 2,000 - 20,000 active binding
sites with footprint ~200-400 base pairs
• Similarly ChIP-Seq is fairly efficient for
mapping uncommon histone modifications
and for RNA Polymerase occupancy ,
because the genomic regions occupied
are very narrow
Chromatin Immuno-Precipitation
(ChIP) is a
method for
fragments from
DNA near
specific proteins
or specific
From Massie, EMBO Reports, 2008
Chromatin Immuno-precipitation
• Proteins are cross-linked to DNA
by formaldehyde or by UV light
NB proteins are even more linked to
each other than to DNA
• DNA is fragmented
• Antibodies are introduced
NB cross-linking may disrupt epitopes
• Antibodies are pulled out (often on
magnetic beads)
• DNA is released and sequenced
CLIP-Seq – A Related Assay
• Cross-linking immuno-precipitation (CLIP)Seq is used to map locations of RNAbinding proteins on mRNA
• Even miRNA binding can be mapped
indirectly by CLIP-Seq with antibodies
raised to Argonaute – an miRNA
accessory protein
What ChIP-Seq Data Look Like
From Rozowsky et al, Nature Biotech 2009
The Value of Controls:
ChIP vs. Control Reads
Red dots are windows containing
ChIP peaks and black dots are windows
containing control peaks used for
FDR calculation
NB. Non-specific
enrichment depends
on protocol
Need controls for
every batch run
Goals of Analysis
1. Identify genomic regions - ‘peaks’ –
where TF binds or histones are modified
2. Quantify and compare levels of binding or
histone modification between samples
3. Characterize the relationships among
chromatin state and gene expression or
General Characteristics of ChIP-Seq Data
• Fragments are quite large relative to
binding sites of TFs
• ChIP-exo (ChIP followed by exonuclease
treatment) can trim reads to within a
smaller number of bases
• Histone modifications cover broader
regions of DNA than TFs
• Histone modification measures often
undulate following well-positioned
ChIP Reads Pile Up in ‘Peaks’ at TF
Binding Sites on Alternate Strands
ChIP-Seq for Transcription Factors
• Typically several thousand distinct peaks
across the genome
• Not clear how many of lower peaks
represent low-affinity binding sites
From Rozowsky et al, Nature Biotech 2009
ChIP-Seq for Polymerase
• Fine mapping of Pol2 occupancy shows
peaks at 5’ and 3’ ends
From Rahl et al Cell 2010
ChIP-Seq Histone Modifications
• Many histone modifications are over
longer stretches rather than peaks
• May have different profiles
• Not clear how to compare
Issues in Analysis of ChIP-Seq Data
• Many false positive peaks
– How to use controls in data analysis
– How to count reads starting at same locus
• What are appropriate controls?
– Naked DNA, untreated chromatin, IgG
• Some DNA regions are not uniquely
identifiable – ‘mappability’
• How to compare different samples?
– Overlap between peak-finding algorithm
results are often poor
Mapability Issues
• Many TFBS and histone modifications lie
in low-complexity or repeat regions of DNA
• With short reads (under 75 bp), with some
errors, it may not be possible to uniquely
identify (map) the locus of origin of a read
• UCSC provides a set of mapability tracks
– Select Mapping and Sequencing Tracks
– Select Mapability
– 35, 40, 50 & 70-mer mapability (some with
different error allowances)
END for Seq Analysis
ChIP-Seq for Histone Modifications
• Various histone modifications characterize
different regulatory states
Exon Peaks
Intronic Peaks
Intergenic H3K4me3 Peaks
• Peak of
H3K4me3 in
region not
annotated by
• Corresponds to
unknown TAR in
• (annotated by
H3K4me3 vs Gene expression
• 91.5% of
expressed genes
have H3K4me3
H3K4me3 peaks at TSS peaks within 1 kb of the TSS
CLIP-Seq for RNA Binding Proteins
Thomson D W et al. Nucl. Acids Res. 2011;39:6845-6853
© The Author(s) 2011. Published by Oxford University Press.
The ENCODE Project
• Comprehensive characterization of
chromatin state and locations of some TFs
and other DNA-binding proteins (e.g. Pol2)
across various conditions in human cell
lines and in mouse tissues
• So far no normal human tissues, which
likely have rather different epigenetic
Demo: ENCODE Data at UCSC
ChIP-Seq Demo
Peak Calling
Goals of Peak-Calling
• Identify discrete locations where a
particular protein binds
• Often applied more generally (and IMHO
poorly) to identification of short regions
with a particular histone modification
Issues in Peak-Calling
• Background of random genomic reads is
not uniform
– Affected by CG content and other factors
• Most good algorithms try to estimate a
local background
• Local background has peaks too!
Peak-Finding - Simple
• Extend tags; sum overlaps at each base
• Find center of each discrete cluster
• Issues:
– Are clusters discrete?
– How much to extend?
• Fragment size unclear
Peak Finding – Better
• Tags starting on
opposite strands
are likely to start
at opposite ends
of precipitated
• Identifying the
cross-over point
leads to better
Issue: How to Identify a Cluster
• Background varies – often related to local
CG content and chromatin state
• Need statistical test for excess counts
above local background
• Usually done by binning counts into 1kb
bins across genome
Control Reads Show Peaks Also
From Rozowsky et al Nature Biotech 2009
Cause of Variation in Read Density
• In study of FoxA1
binding, even control
reads enriched near
FoxA1 binding site!
• Probably due to open
chromatin near FoxA1
binding site
Density of Control Channel
reads around FoxA1 site
Courtesy Shirley Liu
Peak Finding by MACS
• Smart peak imputation estimate
– Uses read directions
– Empirical estimate of fragment length
• Local frequency estimate
– Using control, if available
– Using wide estimate, otherwise
– Not using sequence
MACS Workflow
• Key innovation is to
estimate fragment
length empirically
• If no control sample
MACS estimates
background from
median of ChIP bin
– No use of CG content
The Value of Controls:
ChIP vs. Control Reads
Red dots are windows containing
ChIP peaks and black dots are windows
containing control peaks used for
FDR calculation
Issue: Fragment Lengths
• Puzzle: Fragments from sonication
expected to be between 200 – 500 bp
• Empirically estimated fragment size ~ 100
• Shirley Liu’s explanation: preferential
fragmentation near TF
Peak Calling Demo
Quantitative Comparison of
ChIP Data

similar documents