1. Introduction to RNA-seq
Joel Parker, Ph.D.

2. Why mRNAseq? Measurement of differential expression
There are at least four compelling reasons for choosing mRNA-seq instead of microarray based technologies
Specificity of what is being measured
Reduced technical (batch) bias
Increased dynamic range and log ratio (FC) estimates
More sensitive detection of genes, transcripts, and differential expression
Other reasons
Detection of expressed SNVs
Detection of fusions and other structural variations
No transcriptome definition is needed
No probes need to be designed or manufactured
Cost (will soon be equivalent on a per assay basis with microarray)

4. Why mRNAseq? – Reduced Bias
Cell types separate biologically
CD19 CD8 CD14 CD4

5. Why mRNAseq? – Reduced Processing Bias
Client’s miRNAseq samples sequenced on 4 different machines at 2 different sites at different times over several months with no apparent bias in the top principal components
GAIIx
HS-01 HS-02
HS-IL

8. Library preparation mRNA RNA Capture Enrichment via hybridization
Total RNA
Depletion of rRNA via hybridization
Blood, MT, etc

9. PMID:

10. mRNA

12. Sequencing parameters
Read Length
Trapnell et al., Nature Biotechnology 31,46–53 (2013)
Precision = PPV; Recall = Sensitivity

13. Detection is Dependent on Depth
PMID:

14. Liu et al., Bioinformatics (2014) 30 (3): 301-304.

15. Computational Processing
Technical variation (batch effects) from library preparation and sequencing are small, and the sequencing strategy directs the level of repeatability and detection, especially depth
The raw results of sequencing require significant computational processing
Alignment : Maximizing unambiguous alignments; Alignment of reads that cross exon junctions; Ex: Bowtie, BWA, TopHat, Mapsplice, STAR, . .
Abundance estimation : Gene or transcript; Handling alignments that are ambiguous in the transcriptome; Ex: Sailfish, RSEM, Cufflinks, MISO, Salmon, IsoEM, IsoInfer, Rseq, . . .
Normalization of read counts : Minimizing bias due to variation in number of clusters available; Ex: Total count (RPM), Upper quartile, quantile, density
Different algorithmic and computational strategies, reference genome and transcriptome definition, impact performance much more than SE vs. PE, 50 bp vs. 100 bp.

16. Alignment
BWA, Bowtie
alignment to transcriptome X X X X X X X
Trinity,
Trans-Abyss X X X
Transcriptome
Count alignments

17. Example Concordant GeneV2 V1

18. Example Discordant 1 GeneV2 V1

19. Example Discordant 2 GeneV2 V1

20. Alignment
TopHat, MapSplice, STAR
Trinity,
Trans-Abyss

33. Alignment Comparison
Engstrom et al., Nature Methods 10, (2013)

34. Engstrom et al., Nature Methods 10, 1185-1191 (2013)
Alignment Comparison
Splice Junction Accuracy
Engstrom et al., Nature Methods 10, (2013)

37. Computational Processing
Technical variation (batch effects) from library preparation and sequencing are small, and the sequencing strategy directs the level of repeatability and detection, especially depth
The raw results of sequencing require significant computational processing
Alignment : Maximizing unambiguous alignments; Alignment of reads that cross exon junctions; Ex: Bowtie, BWA, TopHat
Abundance estimation : Gene or transcript; Handling alignments that are ambiguous in the transcriptome; Ex: Sailfish, RSEM, Cufflinks, MISO, IsoEM, IsoInfer, Rseq, . . .
Normalization of read counts : Minimizing bias due to variation in number of clusters available; Ex: Total count (RPM), Upper quartile, quantile, density
Different algorithmic and computational strategies, especially the transcriptome definition, impact performance much more than SE vs. PE, 50 bp vs. 100 bp.

38. Multireads: Reads Mapping to Multiple Genes/Transcripts
HTSeq
<<
PMID:
Wang X, Wu Z, Zhang X. Isoform abundance inference provides a more accurate estimation of gene expression levels in RNA-seq. J Bioinform Comput Biol Dec;8 Suppl 1: PubMed PMID:

39. Multireads: Reads Mapping to Multiple Genes/Transcripts
200
350 1
Long
150
100
300 2
Medium
Multireads 50
200 3
Short
Unique
Relative abundance for these genes, f1, f2, f3 N

40. Approach 1: Ignore Multireads
200
350 1
Long
150
100
300 2
Medium 50
200 3
Short
Relative abundance for these genes, f1, f2, f3
Nagalakshmi et. al. Science. 2008
Marioni, et. al. Genome Research 2008

41. Approach 1: Ignore Multireads
200
350 1
Long
150
100
300 2
Medium 50
200 3
Short
Over-estimates the abundance of genes with unique reads
Under-estimates the abundance of genes with multireads
Not an option at all, if interested in isoform expression N

42. Approach 2: Allocate Fraction of Multireads Using Estimates From Uniques
200
350 1
Long
150
100
300 2
Medium 50
200 3
Short
Relative abundance for these genes, f1, f2, f3
Ali Mortazavi, et. al. Nature Methods 2008
Sailfish, RSEM,Cufflinks N

43. Cufflinks
PMID:

44. RSEM
Li and Dewey, 2011
PMID:
The model consists of N sets of random variables, one per sequenced RNA-Seq fragment. For fragment n, its parent transcript, length, start position, and orientation are represented by the latent variables Gn, Fn, Sn and On respectively. For PE data, the observed variables (shaded circles), are the read lengths ( and ), quality scores ( and ), and sequences ( and ). For SE data, , , and  are unobserved. The primary parameters of the model are given by the vector θ, which represents the prior probabilities of a fragment being derived from each transcript.
θi represents the probability that a fragment is derived from transcript i 
A) PE isoform; B) PE gene; C) SE isoform; D) SE gene

45. Salmon Novelties
Streaming variational Bayes (VB) inference combined with batched VB or EM
Lightweight alignment through maximal exact matches
Transcript / gene abundance inference is abstracted from the alignment step [RSEM also permits this; sam-xlate in

49. Repeatability & Detection by Isoform Database
Ensemb
37612 RefSeq
77608 eaGene
Larger reference transcriptomes result in reduced repeatability (left), but increased detection (right)
Detection - 73% of RefSeq, 66% of UCSC, and 52% of Ensembl

50. Computational Processing
Technical variation (batch effects) from library preparation and sequencing are small, and the sequencing strategy directs the level of repeatability and detection, especially depth
The raw results of sequencing require significant computational processing
Alignment : Maximizing unambiguous alignments; Alignment of reads that cross exon junctions; Ex: Bowtie, BWA, TopHat
Abundance estimation : Gene or transcript; Handling alignments that are ambiguous in the transcriptome; Ex: Sailfish, RSEM, Cufflinks, MISO, IsoEM, IsoInfer, Rseq, . . .
Normalization of read counts : Minimizing bias due to variation in number of clusters available; Ex: Total count (RPM), Upper quartile, quantile, density
Different algorithmic and computational strategies, especially the transcriptome definition, impact performance much more than SE vs. PE, 50 bp vs. 100 bp.

51. Normalization

52. Differential Expression
Bullard JH, Purdom E, Hansen KD, Dudoit S. Evaluation of statistical methods for normalization and differential expression in mRNA-Seq experiments. BMC Bioinformatics Feb 18;11:94. doi: / PubMed PMID: ; PubMed Central PMCID: PMC

53. Statistical power
Why not use a simple t-test on log normalized counts?
DESeq2
t-test

55. Dispersion parameter quantity of interest
raw count for gene i, sample j
normalization factor
quantity of interest
one dispersion per gene
the figure shows random variation with increasing mean for a NB with constant dispersion parameter
K_{ij} \sim \text{NB}(s_{ij} q_{ij}, \alpha_i )
\text{Var}(K_{ij}) = \mu_{ij} + \alpha_i \mu_{ij}^2
variance depends on mean value though

56. Differences across condition
r indexes coefficients: r=1 Intercept, r=2 condition B vs A, etc.
K_{ij} &\sim \text{NB}(s_{ij} q_{ij}, \alpha_i ) \\
\log_2 q_{ij} &= \sum_r x_{jr} \beta_{ir}
xj. = [1,0,...] if sample j is in A, while xj. = [1,1,...] if sample j is in B

57. Controlling for different batches
Using a design formula: ~ batch + condition, adds terms that control for batch differences
If batches are unknown, possible to detect these with other methods: svaseq, RUVSeq
data <- data.frame(counts=rnorm(10*4, rep(c(100,200,300,400),each=10), 50), batch=factor(rep(1:2,each=2*10)), condition=factor(rep(c("A","B","A","B"), each=10)));ggplot(data, aes(x=batch, y=counts, fill=condition)) + geom_boxplot()

58. Complex designs
Treatment effect for enriched samples over baseline, controlling for individual effects: ~indiv + enrich + cond + enrich:cond
indiv enrich cond
1 input ctrl
1 IP ctrl
1 input trt
1 IP trt
2 input ctrl
2 IP ctrl
2 input trt
2 IP trt
...

59. "Wilkinson and Rogers" notation
y ~ ...
'y' modeled by a linear predictor of ...
a + b
'a' and 'b' each put in model
a + b + a:b
'a' and 'b' and their interaction
a*b
equivalent to the above row
0 + a
'a' and no intercept
a + I(a^2)
quadratic function of 'a'
poly(a,2)
orthogonal polynomial of deg 2
ns(a, df=df)
natural cubic spline of degree 'df'
-Formula vignette

60. DESeq2 steps DESeq() results() size factors (sequencing depth)
count matrix (from featureCounts, htseq, tximport, etc.)
size factors (sequencing depth)
dispersion (biological variance)
Wald test or likelihood ratio test
build results table
DESeq()
results()

61. Testing against a threshold
"We get too many DEGs..."
using 'lfcThreshold' in results()
null hypothesis: fold change = 1
null hypothesis: fold change is < 2 or > 1/2
"For well-powered experiments, however, a statistical test against the conventional null hypothesis of zero LFC may report genes with statistically significant changes that are so weak in effect strength that they could be considered irrelevant or distracting."

62. Multiple Test Correction
What is a p-value?
How is it distributed when the null is true?

63. Multiple Test Correction
P-value - probability of the observation when null hypothesis is true
Q-value – false discovery rate at which the gene is called significant
False discovery rate – expected fraction false positives corresponding to a set of results

64. Count model vs linear model
DESeq2 and edgeR similar approach, similar results
very sensitive, may sometimes underestimate FDR
limma + voom uses a linear model, weights determined by variance over mean
strong control of FDR, may be less sensitive for small counts and small sample size
recommend when number of biological replicates per group grows large (e.g. > 20) and small counts not of interest

65. Two paths in RNA-seq analysis
Count matrix
Transformations and
Exploratory Data Analysis (EDA)
clustering, heatmaps, sample-sample distances
vst(), rlog(), plotPCA()
cpm(), plotMDS()
Differential expression
testing, p-values, FDR
DESeq()
results()
glmLRT() topTags()
DESeq2
DESeq2
edgeR
edgeR

66. Regularized logarithm, "rlog"
log2(x + 1)
"rlog"
sample 2
sample 2
sample 1
sample 1
Poisson noise from low counts, when squared
a big contribution to Euclidean distance between samples

67. VST and rlog vs log(x+1)
Essentially provides a similar outcome as filtering at T and/or adding a pseudocount of X, but parameter estimation is data-driven.

69. Practical guides