1. Differential expression analysis with RNA-Seq
G-OnRamp Beta Users Workshop
Wilson Leung
07/2017

2. Outline Design considerations for RNA-Seq experiments
Interpret FastQC results
Optimize alignment parameters for HISAT
Assess alignment statistics with CollectRnaSeqMetrics
Assemble transcripts with StringTie
Differential expression analysis with DESeq2

3. RNA-Seq overview
Griffith M et al. Informatics for RNA Sequencing: A Web Resource for Analysis on the Cloud. PLoS Comput Biol Aug 6;11(8):e

4. Use the RNA Integrity Number (RIN) to assess the quality of the RNA sample
Electropherograms produced by the Agilent Bioanalyzer
Expects 2:1 ratio between 28S and 18S rRNA
RIN values range from 1 (most degraded) to 10 (least degraded)
RIN = 10
RIN = 1
Fluorescence
Length
28S
18S
Length
Fluorescence
Traditionally measures RNA integrity by the ratio of 28S to 18S rRNA
Increase signal in the Fast region as the RNA degrades
Prefer RIN >= 8 for Illumina RNA-Seq sequencing
Fast region 5S
Marker
Length
Gallego Romero I et al. RNA-Seq: impact of RNA degradation on transcript quantification. BMC Biol May 30;12:42.

5. Common applications of RNA-Seq
Transcriptome profiling
Identify novel transcripts (e.g., gene annotations)
Quantify expression levels
Differential expression
Different developmental stages; treatment versus control
Alternative splicing
Visualization and integration with other datasets
Correlate with epigenomic landscape
Histone modifications, DNA methylation, etc.
Conesa A et al. A survey of best practices for RNA-Seq data analysis. Genome Biol Jan 26;17:13.

6. Using RNA-Seq to identify chimeric transcripts
Often found in cell lines and cancer genomes
Structural changes to transcripts would affect the differential expression analysis
Many DE tools assume that most genes have similar expression levels
Maher CA et al. Chimeric transcript discovery by paired-end transcriptome sequencing. Proc Natl Acad Sci U S A Jul 28;106(30):

7. The optimal RNA-Seq sequencing and analysis protocols depend on the goals of the study

8. Design considerations for RNA-Seq
Experimental design
Number of samples, number of biological and technical replicates
Sequencing design
Spike-in controls, randomization of library prep and sequencing
Quality control
Sequencing quality, mapping bias
Conesa A et al. A survey of best practices for RNA-Seq data analysis. Genome Biol Jan 26;17:13.

9. Biological and technical replicates
Biological replicates
RNA from independent growth of cells and tissues
Account for biological variations
Technical replicates
Different library preparations of the same RNA-Seq sample
Account for batch effects from library preparations
Sample loading, cluster amplifications, etc.
ENCODE long RNA-Seq standards:
Blainey P et al. Points of significance: replication. Nat Methods Sep;11(9):

10. Recommended RNA-Seq sequencing depth based on genome size
Differential expression analysis
(# reads)
Detect rare transcripts /
de novo assembly
Read length
Small
(bacteria / fungi)
5 M
30 – 65 M
50 bp
Intermediate
(D. melanogaster)
10 M
70 – 130 M
50 – 100 bp
Large
(Human)
15 – 25 M
100 – 200 M
> 100 bp
HiSeq 3000 generates 313 million reads per lane

11. How many biological replicates?
As many as possible…
Analysis of 48 biological replicates in two conditions
Requires 20 biological replicates to detect > 85% of all differentially expressed genes
Recommend at least six biological replicates per condition
Twelve biological replicates needed to detect smaller fold changes (≥ 0.3-fold difference in expression)
Three biological replicates per condition can usually detect genes with ≥ 2-fold difference in expression
Three replicates detect only 20-40% of differentially expressed genes
Use edgeR (exact) if there are less than 12 replicates
Use DESeq if there are more than 12 replicates
Schurch NJ et al. How many biological replicates are needed in an RNA-Seq experiment and which differential expression tool should you use? RNA Jun;22(6):

12. Power curves for number of biological replicates in each condition
Web interface for RnaSeqSampleSize:
# Samples in each condition 10 20 30 40 50 60
0.0
0.2
0.4
0.8
0.6
FDR = 0.05
Power
Bottomly – mouse strain comparison – LFC = 0.99, dispersion = 0.035
Statistical power (ability to detect an effect, 1 - Type II error)
Power >= 0.8
Sequence more samples (multiplex) = lower coverage
FDR = 0.01

13. Using Galaxy to perform RNA-Seq analysis
Quality control with FastQC
Trim low quality bases with Trimmomatic
Read mapping with HISAT
Transcriptome assembly with StringTie
Differential expression analysis with DESeq2
Based on RNA-Seq tutorial developed by Mo Heydarian and Mallory Freeberg at Johns Hopkins University:

14. Study design for the differential expression (DE) analysis walkthrough
G1E mouse cell line
ES cells hemizygous for Gata1 knockout
Megakaryocytes (Mk)
Large bone marrow cell
Produces platelets for blood clotting
Goal: Identify transcripts regulated by GATA1
G1E
Differentiation to erythrocytes (red blood cells) or myeloid (bone marrow)
Gata1 is a transcription factor
Part of a larger study that also examined the ChIP-Seq data for multiple transcription factors Mk
Wu W et al. Genome Res Dec; 24(12):

15. Quality control with FastQC
Determine quality encoding of fastq files
Assess quality of sequencing sample
Identify overrepresented sequences
Adapters, potential contamination, etc.

16. FastQC: Per base sequence quality (G1E_R1_f_ds_SRR549355)
van Gurp TP et al. Consistent errors in first strand cDNA due to random hexamer mispriming. PLoS One Dec 30;8(12):e85583.

17. Sequence bias at 5’ end caused by random hexamer priming
Sequence content across all bases %T %C %A %G
Higher G + C content in coding regions
chr19 average: 29% A+T, 21% G+C
Hansen KD et al. Biases in Illumina transcriptome sequencing caused by random hexamer priming. Nucleic Acids Res Jul;38(12):e131.

18. GC distribution over all sequences
Use peaks in the Per sequence GC content panel to identify potential contamination
GC distribution over all sequences
Multiple peaks
GC count per read
Theoretical distribution
FTH1 – storage of iron in soluble and non-toxic state
Peak overlaps with an STS marker
RNA-Seq signal overlaps with STS marker RH94403
G1E_R1_f_ds_SRR549355

19. Investigate overrepresented sequences
G1E Mk
Fth1

20. High RNA-Seq coverage at 5’ UTR of Fth1 overlaps with a STS marker
STS = Sequence-Tagged Sites
Single occurrence in the genome, used for constructing genetic and physical maps
Sequence %A %C %G %T
RH94403
15.3%
45.9%
23.0%
15.8%

21. RNA-Seq read mapping with HISAT
Many alignment parameters available…
Which parameters should be changed?

22. Strand-specific RNA-Seq libraries
Standard libraries do not preserve strand information
Prefer strand-specific RNA-Seq
Transcript reconstruction and quantification
Detect antisense transcripts
Library prep kits available from Illumina:
TruSeq Stranded Total RNA Sample Prep Kit
TruSeq Stranded mRNA Sample Prep Kit
Zhao S et al. Comparison of stranded and non-stranded RNA-Seq transcriptome profiling and investigation of gene overlap. BMC Genomics Sep 3;16:675.

23. Orientations of RNA-Seq reads depend on the paired-end protocol
TruSeq Strand-Specific Total RNA:
First Strand (R/RF)
fr-firststrand (F2R1)
NuGEN Encore:
Second Strand (F/FR)
fr-secondstrand (F1R2)
NuGEN Ovation V2:
FR Unstranded
fr-unstranded
Use the infer_ RSeQC
I = Inward, O= outward, M = matching; S = stranded, U = unstranded; F = forward, R = reverse
F2R1 = second read in forward strand, first read in reverse strand (ISR), RF
See Supplemental Table S5 in
Griffith M et al. PLoS Comput Biol Aug 6;11(8):e

24. Use infer_experiment.py to infer the design of the RNA-Seq experiment
Part of the RSeQC package:
Map RNA-Seq reads using default parameters
Run infer_experiment.py to infer experimental design
Map RNA-Seq reads using the inferred parameters
Available through the Galaxy Tool Shed (rseqc):
Can also try to infer RIN (RNA integrity)

25. Common changes to HISAT alignment parameters
Minimum and maximum intron lengths
Specify strand-specific information
GTF file with known splice sites
Use known gene annotations to guide read mapping if available
Transcriptome assembly reporting

26. Use splice site information during read mapping to improve alignment accuracy
Recommend run STAR and TopHat2 twice:
Round 1 to discover junctions; round 2 use these junctions in read mapping
HISAT by default make use of splice sites found during the alignment process so that it does not have to run twice
(Compare HISATx1, HISAT, and HISATx2)
Kim D et al. HISAT: a fast spliced aligner with low memory requirements. Nat Methods Apr;12(4):

27. RNA-Seq alignment strategy for multiple samples
Map reads from each RNA-Seq sample separately
Use the --novel-splicesite-outfile option to report splice sites identified in each sample
Combine splice junctions from all samples under all conditions into a single splice junction file
Filter splice junctions by score
Retain junctions that appear in multiple biological replicates
Map reads from each RNA-Seq sample using the combined splice junctions file
Use the --novel-splicesite-infile option
For model organisms / human, could use --known-splicesite-infile directly

28. PCR amplification biases in Illumina data
Major contributors to bias:
Fragment size
Base composition
Bridge amplification cycles
Aird D et al. Analyzing and minimizing PCR amplification bias in Illumina sequencing libraries. Genome Biol. 2011;12(2):R18.
Relative abundance
GC content of amplicon
Illumina
qPCR
100 10 1
0.1 20 40 60 80
Bias toward smaller fragments
Problem with GC-rich sequences (secondary structures, higher melting temperature)

29. Identify duplicate reads based on sequence alignments
Assumption:
Rare for different sequenced fragments to have the same start and end positions
The Picard tool MarkDuplicates only considers the start position
RNA-Seq data violates this assumption:
Reads map to a smaller portion of the genome
Probability of reads with the same start position depends on gene length and expression levels
Recommendation:
Mark potential duplicates to verify that all samples have similar levels of “duplication”
Retain duplicate reads in differential expression analyses
Examples: ribosomal, mitochondrial house keeping genes
Tools such as Picard MarkDuplicates only examine the 5’ position (after accounting for clipping)
3’ end typically have lower quality
Williams AG et al. RNA-Seq Data: Challenges in and Recommendations for Experimental Design and Analysis. Curr Protoc Hum Genet Oct 1;83:

30. MarkDuplicatesWithMateCigar does not work with RNA-Seq data (use too much memory for large gaps)
(not used in production pipeline in Broad)
Histogram tries to estimate additional benefits of sequencing (ROI)
DEMO: Mark optical and PCR duplicates using the Picard tool MarkDuplicates

31. Assess RNA-Seq read alignments with CollectRnaSeqMetrics
Requires gene annotations:
Gene annotations in GTF / GFF format
UCSC Genes in refFlat format (from the UCSC Table Browser)
BAM dataset collection
Reference genome (mm10)
CollectRNASeqMetrics – median coverage, 5’/3’ biases, number of reads assigned to correct strand, etc.
Gene annotations in refFlat format
rRNAs in interval list format
Second read transcription strand

32. DEMO: Use CollectRnaSeqMetrics to assess RNA-Seq alignments

33. Identify coverage bias along the transcript
Normalized coverage
Normalized distance along transcript
RNA-Seq coverage vs. transcript position
(G1E_R1)
2.5
2.0
1.5
1.0
0.5
0.0 40 20 60 80
100
cDNA fragmentation
DNase I treatment or sonication
RNA fragmentation
RNA hydrolysis or nebulization
RNA-Seq Read Count
1.0
oligo-dT primed cDNA 5’ 3’
Gene Span (5,099 genes)
Wang Z et al. RNA-Seq: a revolutionary tool for transcriptomics. Nat Rev Genet Jan;10(1):57-63.

34. Two common approaches to RNA-Seq assembly
Reference-based assembly
Map RNA-Seq reads against a reference genome
Examples: TopHat2, HISAT
Assemble transcripts from mapped RNA-Seq reads
Examples: Cufflinks, StringTie
De novo transcriptome assembly
Assemble transcripts from RNA-Seq reads
Examples: Oases, Trinity
More computationally expensive
Merge assemblies produced by different parameters
Advantage of de novo assembly is that it does not require a reference genome

35. Augment mapped RNA-Seq reads with pre-assembled super-reads (SR)
Pertea M et al. StringTie enables improved reconstruction of a transcriptome from RNA-Seq reads. Nat Biotechnol Mar;33(3):290-5.

36. Transcriptome assembly remains an active area of research
Korf I. Genomics: the state of the art in RNA-Seq analysis. Nat Methods Dec;10(12):
Steijger T et al. Assessment of transcript reconstruction methods for RNA-Seq. Nat Methods Dec;10(12):

37. DEMO: Assemble transcripts from mapped RNA-Seq reads using StringTie

38. Metrics for quantifying gene expression levels
RPKM
Reads Per Kilobase per Million mapped reads
Normalize relative to sequencing depth and gene length
FPKM
Similar to RPKM but count DNA fragments instead of reads
Used in paired end RNA-Seq experiments to avoid bias
TPM
Transcripts Per Million
Normalize for gene length, then normalize by sequencing depth
FPKM = Count fragments because both paired-end reads are derived from the same transcript
TPM = number of transcripts that you would see if you sequence 1M transcripts (given the abundance of other transcripts in the sample)
Normalize FPKM so that the total sum is 1M => TPM
Wagner GP et al. Measurement of mRNA abundance using RNA-Seq data: RPKM measure is inconsistent among samples. Theory Biosci Dec;131(4):281-5.

39. Most differential expression analysis tools use read counts as input
RPKM, FPKM, and TPM compare relative gene expression levels within a sample
Most differential expression analysis tools use read counts as input
edgeR uses CPM (TPM) for filtering, TMM normalization for differential expression analysis
DESeq2 uses rlog for clustering

40. Create a transcriptome library with StringTie merge
Combine transcripts from multiple samples into a single transcriptome database
Sample 1
(A)
Sample 2
Sample 3
(B)
Sample 4
Reference annotation
Use the unique exon to try to quantify transcripts instead of genes
Merged assemblies
Pertea M et al. Transcript-level expression analysis of RNA-Seq experiments with HISAT, StringTie and Ballgown. Nat Protoc Sep;11(9):

41. DEMO: Use StringTie merge to construct a transcriptome database for G1E and Mk

42. Differential expression analysis tools
Compare genes expression levels:
DESeq2 (
edgeR (
Compare transcripts expression levels:
Cuffdiff (
Ballgown (
MISO (
Salmon (
Compare both genes and transcripts expression levels:
RSEM + EBSeq (
DESeq2 – MIchael Love started to examine behavior of DESeq2 at the transcript-level:
DESeq was originally designed for differential expression analysis at the gene level:
Recommend using DEXSeq to study differential expression of exons instead of focusing on the isoforms

43. Count the number of reads that overlap with each gene using htseq-count
Three modes of overlap resolution:
union
intersection_strict
intersection_nonempty
htseq-count ignores multi-mapped reads

44. featureCounts is a faster alternative to htseq-count
C versus python…
Liao Y et al. featureCounts: an efficient general purpose program for assigning sequence reads to genomic features. Bioinformatics Apr 1;30(7):

45. Issues with the GTF files produced by the UCSC Table Browser
The gene_id and transcript_id fields in the GTF file have the same values
GTF format stipulates that different isoforms derived from the same gene should have the same gene_id
Two potential solutions:
Download genePred file from the UCSC Genome Browser web site, then use genePredToGtf to create the GTF file
Use the GTF files from Ensembl
Also issue with duplicate entries in the GTF files produced by the UCSC Table Browser
(including the GTF file that is part of the Data Library)

46. DEMO: Calculate the read count for each transcript using featureCounts

47. Use Poisson distribution to model the distribution of read counts
Probability of the number of “events” in a fixed amount of time or space
Events = RNA-Seq reads
Mean = variance = λ
Probability distribution is based on binomial distribution
Model noise within the same condition
Assumptions:
Events are independent
Rate of events are constant

48. Overdispersion in RNA-Seq data
More highly expressed genes show higher variance
log(variance)
log(mean) -5 5 10 15 25 20
Poisson
Overdispersion
Actual
Negative binomial distribution:
Another problem with RNA-Seq data is heteroscedasticity:
More highly expressed genes show more variability
Overdispersion is caused primarily by biological noise
Noise
Low expression: shot noise dominates (solution = sequence deeper)
High expression: biological noise dominates (solution = more replicates)
Zhou YH et al. A powerful and flexible approach to the analysis of RNA sequence count data. Bioinformatics Oct 1;27(19):

49. Use biological replicates to control Type I errors
Use biological replicates to estimate variance within a condition
Identify differentially expressed genes under different conditions -5 5
100
101
102
103
104
105
106
10-1
Mean expression
log2 fold change 5
log2 fold change
If no variation among biological replicates, then expect 0 fold change
Yellow line = Poisson
Purple line = Poisson + local regression (DESeq)
(local regression for gamma family GLM with correction for library size)
Slides available through the Bioconductor web site: -5
100
101
102
103
104
105
Mean expression
Huber W. Differential Expression for RNA-Seq.

50. DEMO: Differential expression analysis using DESeq2

51. Verify results using multiple differential expression analysis tools
Impact of read depth on differential expression analysis
Use the intersection of differentially expressed genes identified by multiple tools in downstream analyses
MAQC = microarray quality control project – human brain reference + universal human reference
K_N = mouse neurosphere cells treated by potassium chloride and norepinephreine
LCL = lymphoblastoid cell lines – Nigeria human samples from HapMap project 5M
10M
20M
30M
Zhang ZH et al. A comparative study of techniques for differential expression analysis on RNA-Seq data. PLoS One Aug 13;9(8):e

52. Tools for analyzing differentially expressed genes
Gene Ontology (GO) terms enrichment:
topGO (
goSTAG (
DAVID (
Pathway analysis:
GAGE (
Reactome (
Sample walkthrough:
From reads to genes to pathways: differential expression analysis of RNA-Seq experiments using Rsubread and the edgeR quasi-likelihood pipeline

53. Additional resources
Analysis of RNA-Seq data: gene-level exploratory analysis and differential expression
Informatics for RNA-Seq: A web resource for analysis on the cloud
UC Davis Bioinformatics Core training course
So you want to do a: RNA-Seq experiment, Differential Gene Expression Analysis
Specific course from UC Davis on RNA-Seq and differential gene expression analysis

54. Summary
Statistical power of RNA-Seq depends on the number of biological replicates and sequencing depth
Assess quality of RNA-Seq reads with FastQC
Assess RNA-Seq alignments with Picard tools
Perform transcriptome assembly with StringTie
Perform differential expression analyses with DESeq2

55. https://flic.kr/p/bhyT8B
Questions?

57. Power curves for number of biological replicates in each condition
Web interface for RnaSeqSampleSize:
Ching T et al. Power analysis and sample size estimation for RNA-Seq differential expression. RNA Nov;20(11):
# Samples in each condition 10 20 30 40 50 60
0.0
0.2
0.4
0.8
0.6
FDR = 0.05
FDR = 0.01
Power
Bottomly – mouse strain comparison – LFC = 0.99, dispersion = 0.035
Statistical power (ability to detect an effect, 1 - Type II error)
Power >= 0.8
Sequence more samples (multiplex) = lower coverage