1. Previous Lecture: NGS Alignment

2. Spring CHIBI Courses BMI Foundations I: Bioinformatics (BMSC-GA 4456)
Constantin Aliferis
Study classic Bioinformatics/Genomics papers and reproduce data analysis, for advanced informatics students
Integrative Genomic Data Analysis (BMSC-GA 4453)
Jinhua Wang
build competence in quantitative methods for the analysis of high-throughput genomic data
Microbiomics Informatics (BMSC-GA 4440)
Alexander Alekseyenko
analysis of microbial community data generated by sequencing technologies: preprocess raw sequencing data into abundance tables, associate abundance with clinical phenotype and outcomes.
Next Generation Sequencing (BMSC-GA 4452)
Stuart Brown
An overview of Next-Generation sequencing informatics methods for data pre-processing, alignment, variant detection, structural variation, ChIP-seq, RNA-seq, and metagenomics.
Proteomics Informatics (BMSC-GA 4437)
David Fenyo
A practical introduction of proteomics and mass spectrometry workflows, experimental design, and data analysis

3. This Lecture
ChIP-seq & RNA-seq

4. ChIP-seq experimental methods Alignment and data processing
Learning Objectives
ChIP-seq experimental methods
Transcription factors and epigenetics
Alignment and data processing
Finding peaks: MACS algorithm
Annotation
RNA-seq experimental methods
Alignment challenges (splice sites)
TopHat
Counting reads per gene
Normalization
HTSeq-count and Cufflinks
Statistics of differential expression for RNA-seq

5. ChIP-seq Combine sequencing with Chromatin‐Immunoprecipitaion
Select (and identify) fragments of DNA that interact with specific proteins such as:
Transcription factors 
Modified histones
Methylation
RNA Polymerase (survey actively transcribe portions of the genome)
DNA polymerase (investigate DNA replication)
DNA repair enzmes

6. ChIP-chip [Pre-sequencing technology]
Do chromatin IP with YFA (Your Favorite Antibody)
Take IP-purified DNA fragments, label & hybridize to a microarray containing (putative) promoter (or TF binding) sequences from lots of genes
Estimate binding, relate to DNA binding of protein targeted by antibody
limited to well annotated genomes
need to build special microarrays
suffers from hybridization bias
assumes all TF binding sites are known and correctly located on genome

7. ChIP-seq High-throughput sequencing Map sequence tags to genome
Immunoprecipitate
High-throughput sequencing
Release DNA
Map sequence tags to genome

8. Alignment
Place millions of short read sequence ‘tags’ (25-50 bp) on the genome
Finds perfect, 1, and 2 mismatch alignments; no indels (BWA)
Aligns ~80% of PF tags to human/mouse genome
We parse alignment files to get only unique alignments (removes 2%-5% of ‘multi-mapped’ reads)

9. ChIP-seq for TF (SISSRS software)
Jothi, et al. Genome-wide identification of in vivo protein–DNA binding sites from ChIP-Seq data. NAR (2008), 36:

12. Saturation
How many sequence reads are needed to find all of the binding targets in the genome?
Look for plateau
100% = 15,291 peaks
Rozowsky, et al. Nature Biotech. Vol 27-1, Jan 2009.
Pol2 data: 11M reads vs. 12M control reads, peaks found with MACS, data sub-sampled.

13. ChIP-seq Challenges
We want to find the peaks (enriched regions = protein binding sites on genome)
Goals include: accuracy (location of peak on genome), sensitivity, & reproducibility
Challenges: non-random background, PCR artifacts, difficult to estimate false negatives
Very difficult to compare samples to find changes in TF binding (many borderline peaks)

14. Peakfinding
Find enriched regions on the genome (high tag density) = “peaks”
Enriched vs. what?
A statistical approach assumes an evenly distributed or randomly distributed background
Poisson distribution of background is obviously not true
Any threshold is essentially arbitrary

15. Compare to Background Goal is to make ‘fold change’ measurements
What is the appropriate background?
Input DNA (no IP)
IP with non-specific antibody (IgG)
[We mostly use input DNA]
Must first identify “peak region” in sample, then compare tag counts vs BG

16. MACS
Zhang et al. Model-based Analysis of ChIP-Seq (MACS). Genome Biol (2008) vol. 9 (9) pp. R137
Open source Unix software (Python !)
MACS improves the spatial resolution of binding sites through combining the information of sequencing tag position and orientation by using empirical models for the length of the sequenced ChIP fragments
(slides + and – strand reads toward center of fragment)
MACS uses a dynamic Poisson distribution (local background count in the control) to effectively capture local biases in the genome sequence, allowing for more sensitive and robust prediction
Uses control to calculate “random” peaks, sets FDR rate.
Feng J, Liu T, Zhang Y. Using MACS to identify peaks from ChIP-Seq data. Curr Protoc Bioinformatics Jun;Chapter 2:Unit 2.14.

17. BED format
BED format defines a genomic interval as positions on a reference genome.
An interval can be a anything with a location: gene, exon, binding site, region of low complexity, etc.
MACS outputs ChIP-seq peaks in BED format
BED files can also specify color, width, some other formatting.
chromosome start end
chr
chr
chr
chr
chr
chr
chr
chr
track name="ItemRGBDemo" description="Item RGB demonstration" itemRgb="On"
chr Pos ,0,0
chr Pos ,0,0
chr Pos ,0,0
chr Pos ,0,0
chr Neg ,0,255
chr Neg ,0,255
chr Neg ,0,255
chr Pos ,0,0

20. Remove Duplicates
In some ChIP-seq samples, PCR amplification of IP-enriched DNA creates artifacts (highly duplicated fragments)
Huge differences depending on target of antibody and amount of IP DNA collected.
“Complexity” of the library

21. PCR ‘stacks’
Always in F-R pairs, ~200 bp apart

22. % of Duplicates varies

23. Different IP Targets
Huge difference between Transcription Factors and Histone modification as targets of IP TF
sequence-specific binding motifs
few thousand sites
binding region ~50bp
oriented tags
yes/no binding
promoters or enhancers
Histone Mods
not sequence-specific
tens to hundreds of thousands of sites
large binding region (~2kb)
tags not oriented
signal may be scaled
associated w/ almost all transcribed genes

25. Mikkelsen, Lander, et al. Genome-wide maps of chromatin state in pluripotent and lineage-committed cells. Nature (2007) 448:

26. H3K4me3

28. Normalization How to compare lanes with different numbers of reads?
Will bias fold-change calculations
Simple method – set all counts in ‘peak’ regions as per million reads
This does not work well for >2x differences in read counts.

30. Evaluation Peaks near promoters of known genes (TSS)
Generally a high %
As parameters become less stringent, more peaks are found, % near TSS declines
Estimate false positive rate
Pure statistical (Poisson or Monte Carlo)
Compare 2 bg sampels (QuEST)
Reverse sample & bg (MACS)
Can’t estimate false negative rate – don’t know ‘true’ number of binding sites

31. Evaluation Overlap with ChIP-chip data Reproducibility Synthetic data
What is an overlap?
What % overlap is good?
Reproducibility
Need to define (we use overlap of 1 bp)
Very important for biological conclusions
Essential for comparisons of diff. conditions
Must have replicate samples!!
Trade off: reproducibility vs. sensitivity
Synthetic data
Allows calculation of sensitivity & specificity
How similar to real data? (All synth has bias)

32. Composite image of sequence reads at promoters of all RefSeq genes.
Histone modification (H3K4) ChIP-seq
Composite image of sequence reads at promoters of all RefSeq genes.

33. The Use of Next Generation Sequencing to Study Transcriptomes:
RNA-seq

34. RNA-seq Measures the Transcriptome
Takes advantage of the rapidly dropping cost of Next-Generation DNA sequencing
Measures gene expression in true genome-wide fashion (all the RNA)
Also enables detection of mutations (SNPs), alternative splicing, allele specific expression, and fusion genes
More accurate and better dynamic range than Microarray
Can be used to detect miRNA, ncRNA, and other non-coding RNA

35. RNA-seq Measures the Transcriptome
Takes advantage of the rapidly dropping cost of Next-Generation DNA sequencing
Measures gene expression in true genome-wide fashion (all the RNA)
Also enables detection of mutations (SNPs), alternative splicing, allele specific expression, and fusion genes
More accurate and better dynamic range than Microarray
Can be used to detect miRNA, ncRNA, and other non-coding RNA

36. RNA-seq vs. qPCR

37. Depth of Coverage
With the Illumina HiSeq producing >200 million reads per sample, what depth of coverage is needed for RNA-seq?
Can we multiplex several samples per lane and save $$ on sequencing?
For expression profiling (and detection of differentially expressed genes), probably yes, 2-4 samples per lane is practical

38. 100 million reads, 81% of genes FPKM ≥ 0.05
Each additional 100 million reads detects ~3% more genes
Toung, et al. Genome Res June; 21(6): 991–998..

39. Illumina mRNA Sequencing
Random primer PCR
Poly-A selection
Fragment & size-select

40. Sample prep can create 3’ or 5’ bias
(strand oriented protocol)
no bias (low coverage at ends
of transcript)
3’ bias
(poly-A selection)

41. Detect Small RNAs – depends on sample prep method

42. RNA-seq informatics workflow: genome mapping splice junction fragments
(predict novel junctions/exons)
counts
normalize
differential expression
gene lists
Oshlack et al. Genome Biology 2010, 11:220

43. RNA-seq Alignment Challenges
Using RNA-seq for gene expression requires counting sequence reads per gene
Must map reads to genes – but this is a more difficult problem than mapping reads to a reference genome
Introns create big gaps in alignment
Small reads mean many short overlaps at one end or the other of intron gaps
What to do with reads that map to introns or outside exon boundaries?
What about overlapping genes?

44. TopHat
RNA-seq can be used to directly detect alternatively spliced mRNAs.

45. Map reads to exons & junctions

46. TopHat Trapnell C et al. Bioinformatics 2009;25:1105-1111
The seed and extend alignment used to match reads to possible splice sites.
The seed and extend alignment used to match reads to possible splice sites. For each possible splice site, a seed is formed by combining a small amount of sequence upstream of the donor and downstream of the acceptor. This seed, shown in dark gray, is used to query the index of reads that were not initially mapped by Bowtie. Any read containing the seed is checked for a complete alignment to the exons on either side of the possible splice. In the light gray portion of the alignment, TopHat allows a user-specified number of mismatches. Because reads typically contain low-quality base calls on their 3′ ends, TopHat only examines the first 28 bp on the 5′ end of each read by default.
Trapnell C et al. Bioinformatics 2009;25:

47. Real data generally support existing annotation
Data from Costa lab

48. RNA-seq informatics Filter out rRNA, tRNA, mitoRNA Align to genome
Find splice junction fragments (join exon boundaries)
Differential expression
Alternatively spliced transcripts
Novel genes/exons
Sequence variants (SNPs, indels, translocations)
Allele-specific expression

49. Count Reads per gene Need a reference genome with exon information
How to count partial alignments, novel splices etc?
Simple or complex model?
Simple: HTSeq-count
Complex: Cufflinks
Normalization methods affect the count very dramatically

50. HTSeq-count A simple Python tool.
Relies entirely on an accurate annotation of genes and exons in GFF file.

51. Cufflinks Isoform Models

52. Differential Expression
ADM
Data from Costa Lab

53. Normalization
Differential Expression (DE) requires comparison of 2 or more RNA-seq samples.
Number of reads (coverage) will not be exactly the same for each sample
Problem: Need to scale RNA counts per gene to total sample coverage
Solution – divide counts per million reads
Problem: Longer genes have more reads, gives better chance to detect DE
Solution – divide counts by gene length
Result = RPKM
(Reads Per KB per Million)

54. Better Normalization RPKM assumes:
Total amount of RNA per cell is constant
Most genes do not change expression
RPKM is invalid if there are a few very highly expressed genes that have dramatic change in expression (dominate the pool of reads)
Better to use “Upper Quartile” (75th percentile) or “Quantile” normalization
Different normalization methods give different results (different DE genes & different p-value rankings)

55. Statistics of DE
mRNA levels are variable in cells/tissues/organisms over time/treatment/tissue etc.
Like microarrays, need replicates to separate biological variability from experimental variability
If there is high experimental variability, then variance within replicates will be high, statistical significance for DE will be difficult to find.
Best methods to discover DE are coupled with sophisticated approaches to normalization
Best to ignore very low expressing genes: RPKM<1

56. Popular DE Statistical methods
Cufflinks-Cuffdiff
part of TopHat software suite – easy to use
Uses FPKM normalization
complex model for counting reads among splice variants
can be set to ignore novel variants
Estimates variance in log fold change for each gene using permutations
finds the most DE genes, high false positive rate
edgeR
requires raw count data, does its own normalization
Estimates standard deviation (dispersion) with a weighted combination of individual gene (gene-wise) and global measures
Statistical model is Negative Binomial distribution (has a dispersion parameter)
Fisher’s Exact test (for 2-sample), or generalized linear model (complex design)
acceptable tradeoff of sensitivity and specificity
Many others: DESeq, SAMseq, baySeq.
Many rather inconclusive benchmarking studies

57. DE genes by different methods

58. Differentially expressed genes
70% of DE genes validated by qPCR
Data from Meruelo Lab

59. Alternative Splicing
Data from Costa Lab

60. Good SNP
data from Zavadil lab

61. Novel Genomes
RNA-seq can be used to annotate genomes – gene discovery, exon mapping.
data from Desplan lab

62. ChIP-seq experimental methods Alignment and data processing
Summary
ChIP-seq experimental methods
Transcription factors and epigenetics
Alignment and data processing
Finding peaks: MACS algorithm
Annotation
RNA-seq experimental methods
Alignment challenges (splice sites)
TopHat
Counting reads per gene
Normalization
HTSeq-count and Cufflinks
Statistics of differential expression for RNA-seq

63. Next Lecture: Signal Processing