Data Analysis for High-Throughput Sequencing

1
Data Analysis for High-Throughput Sequencing

• Mark Reimers
• Tobias Guennel
• Department of Biostatistics

2
Unto the Frontiers of Ignorance

• I love the way this workshop starts off with
  things we understand fairly well and works up to
  the cutting edge of things we dont understand at
  all
• - Mike Neale, Oct 14, 2010

3
The New Boyfriend/Girlfriend
4
Where Does HTS Really Make the Difference?

• Sequencing for novel variants
• ChIP-Seq for DNA-binding proteins or less common
  histone marks
• Allele-specific expression
• COMING SOON
• DNA methylation

5
Outline

• Biases in reads
• RNA-Seq
• normalization
• basic tests
• differential splicing
• Finding peaks in ChIP-Seq

6
Technical Biases Sequence Start
The initial bases of reads are highly biased, and
the bias depends on RNA/DNA preparation
7
Sequence Biases K-mers Differ

• (Schroeder et al, PLoS One, 2010) calculated
  proportions of words (k-mers) starting at various
  positions

Expected frequencies if bases random
8
Position of single mismatch in uniquely mapped
tags
Courtesy Jean Danielle Thierry-Mieg
9
Types of mismatches in uniquely mapped tags with
a single mismatch are profoundly asymmetric and
biased
Courtesy Jean Danielle Thierry-Mieg
10
Technical Biases Initiation Sites
COX1
11
Different Platforms Have Different Biases

• (Harismendy et al, Genome Biology, 2009)
  sequenced a section of 4 HapMap individuals on
  Roche 454, on Illumina, and on SOLiD
• 454 had most even coverage

12
Initiation Biases Dwarf Splicing

• Counts of reads along gene APOE in different
  tissues of data from Wold lab. (a) Brain, (b)
  liver, (c) skeletal muscle

13
Variation in Technical Biases

• Sometimes the initial base biases change
  substantially most base proportions change
  together one PC explains 95
• In most preparations the initiation site biases
  change by a few percent
• In a few preparations the initiation site biases
  change by 20-30
• This may have consequences for representation in
  ChIP-Seq assays

14
RNA-Seq Data Analysis
15
Biases in Proportions

• Fragments compete for real-estate on the lane
• If a few dozen genes are highly expressed in one
  tissue, they will competitively inhibit the
  sequencing of other genes, resulting in what
  appears to be lower expression

16
Effects of Competition

• (Robinson Oshlak, Genome Biology, 2010)

17
A Simple Normalization

• Align the medians of the housekeeping genes, or
  the genes that are not expressed at very high
  levels in any sample, across the samples

18
A Simple Model for Counts

• Poisson distribution of counts within a gene with
  mean proportional to Np
• SD of variation equal to square root of Np
• Problem Actual variation of counts between
  replicate samples is significantly higher than
  root Np
• Probably reflecting systematic biases

19
Hacks for Over-Dispersion

• Like l fudge-factor in GWAS
• Use negative binomial model
• There is no relation to meaning of distribution
  numbers of nulls until something happens
• Convenient way to parametrise over-dispersion
• Bioconductor package edgeR estimates parameters
  by Maximum Likelihood

20
Alternate Transcripts Splicing Index

• For each exon, the proportion of transcripts in
  which the exon appears
• Hard to estimate because different exons have
  different representation probabilities
• Use ratios of exons
• Use constitutive exons (if known) as baseline
  for them SI1

from Wang et al, Nature, 2008
21
Detecting Alternate Splicing I

• (Wang et al, Nature, 2008) measured splicing
  index for several tissues

22
Splicing Junction Reads

• Some reads will span two different exons
• Need long enough reads to be able to reliably map
  both sides
• Can use information from one exon to identify
  gene and restrict possibilities for 5 end other
  exon

from Wang et al NAR 2010
23
ChIP-Seq
24
Courtesy Raphael Gottardo
25
A View of ChIP-Seq Data

• Typically reads are quite sparsely distributed
  over the genome
• Controls (i.e. no pull-down by antibody) often
  show smaller peaks at the same locations
• Probably due to open chromatin at promoter

Rozowsky et al Nature Methods, 2009
26
Always Have a Control

• High correlation between peaks in control samples
  and peaks in ChIP sample
• Must subtract estimate of background from control
  tags

From Zhang et al, Genome Biol 2008
27
Locating Binding Sites

• Use the fact that reads on opposite sides of the
  site represent are sequenced in opposite senses

From Zhao et al NAR 2009