1. Simon Andrews simon.andrews@babraham.ac.uk @simon_andrews v2.3
RNA-Seq Analysis
Simon Andrews
@simon_andrews
v2.3

2. RNA-Seq Libraries rRNA depleted mRNA Fragment Random prime + RT
2nd strand synthesis (+ U)
A-tailing
Adapter Ligation
(U strand degradation)
Sequencing
NNNN u u u u u A T u A T A T

3. RNA-Seq Analysis QC (Trimming) Mapping Mapped QC Statistical Analysis
Quantitation
Mapped QC

4. QC: Raw Data
Sequence call quality

5. QC: Raw Data
Sequence bias

6. QC: Raw Data
Duplication level

7. Mapping
Exon 1
Exon 2
Exon 3
Genome
Simple mapping within exons
Mapping between exons
Spliced mapping
Can simplify by aligning first to a transcriptome and then translate back to genomic coordinates. Can map unmatched reads to the whole genome.

8. Spliced Mapping Software
Tophat (
Star (

9. TopHat pipeline Reference FastQ files Indexed Genome
Reference GTF Models
Indexed Transcriptome
Reads
Maps to transcriptome?
Yes
Translate coords and report No
Maps to genome?
Yes
Report No
Split map to genome
Yes
Build consensus and report No
Discard

10. Post Mapping QC Mapping statistics
Proportion of reads which are in transcripts
Proportion of reads in transcripts in exons
Strand specificity
Consistency of coverage
SeqMonk (RNA-Seq QC plot)
RNASeqQC (easiest through GenePattern)

11. SeqMonk Mapping QC (good)

12. SeqMonk Mapping QC (bad)

13. Quantitation Splice form 1 Exon 1 Exon 2 Exon 3 Splice form 2 Exon 1
Definitely splice form 1
Definitely splice form 2
Ambiguous

14. Options for handling splice variants
Ignore them – combine exons and analyse at gene level
Simple, powerful, inaccurate in some cases
DE-Seq, SeqMonk
Assign ambiguous reads based on unique ones – quantitate transcripts and optionally merge to gene level
Potentially cleaner more powerful signal
High degree of uncertainty
Cufflinks, bitSeq, RSEM

15. Read counting Simple (exon or transcript) Complex (re-assignment)
HTSeq (htseq-count)
BEDTools (multicov)
featureCounts
SeqMonk (graphical)
Complex (re-assignment)
Cufflinks, bitSeq, RSEM

16. Count Corrections
Size of library
Length of transcript
Other factors

17. RPKM / FPKM / TPM
RPKM (Reads per kilobase of transcript per million reads of library)
Corrects for total library coverage
Corrects for gene length
Comparable between different genes within the same dataset
FPKM (Fragments per kilobase of transcript per million reads of library)
Only relevant for paired end libraries
Pairs are not independent observations
RPKM/2
TPM (transcripts per million)
Normalises to transcript copies instead of reads
Corrects for cases where the average transcript length differs between samples

18. Normalisation

19. Normalisation

20. Filtering Genes
Remove things which are uninteresting or shouldn’t be measured
Reduces noise – easier to achieve significance
Non-coding (miRNA, snoRNA etc) in RNA-Seq
Known mis-spliced forms (exon skipping etc)
Mitochonidrial genes
X/Y chr genes in mixed sex populations
Unknown/Unannotated genes

21. Filtering Mouse mRNAs

22. Visualising Expression
Comparing the same gene in different samples
Normalised log2 RPM values
Comparing different genes in the same sample
Normalised log2 RPKM values

23. Linear
Log2
CD74
Eef1a1
Actb
Lars2
Eef2

24. Differential Expression
Microarrays traditionally used continuous statistical tests (t-test ANOVA etc)
RNA-Seq differs in that it is count based data, so continuous tests fail at low counts
Most differential tests use count based distribution tests, usually based on a negative binomial distribution

25. Negative binomial tests (DE-Seq)
Are the counts we see for gene X in condition 1 consistent with those for gene X in condition 2?
Initially modelled using simple Poisson distribution using mean expression as the only parameter
Doesn’t model real data very well

26. Poisson vs Negative binomial
DESeq

27. Parameters Size factors Variance – required for NB distribution
Estimator of library sampling depth
More stable measure than total coverage
Based on median ratio between conditions
Variance – required for NB distribution
Custom variance distribution fitted to real data
Insufficient observation to allow direct measure
Smooth distribution assumed to allow fitting

28. Dispersion shrinkage Plot observed per gene dispersion
Calculate average dispersion for genes with similar observation
Individual dispersions regressed towards the mean. Weighted by
Distance from mean
Number of observations
Points more than 2SD above the mean are not regressed

29. Other filters Cook’s Distance – Identifies high variance
Effect on mean of removal of one replicate
For n<3 test not performed
For n = 3-6 failures are removed
For n>6 outliers removed to make trimmed mean
Disable with cooksCutoff=FALSE
Hit count optimisation
Low intensity reads are removed
Limits multiple-testing to give max significant hits
Disable with independentFiltering=FALSE

30. Replicates
Compared to arrays, RNA-Seq is a very clean technical measure of expression
Generally don’t run technical replicates
Some statistics can be run on single replicates, but they can only tell you about technical noise (how likely is it that this change is due to a technical issue)
Assessing biological variation requires biological replicates

31. Replicates Traditional statistics require min 3x3
DESeq can operate at 2x2, but this is minimum, not recommended
True number of replicates required will depend on your biology and requirements
4x4 design is fairly common
Always expect at least one sample to fail
Randomise samples during sample prep

32. The problem of power…
Gene A (1kb)
Gene B (8kb)
In a library Gene B is much better observed for the same copy number
Power to detect DE is proportional to length

33. 5x5 Replicates
5,000 out of 22,000 genes (23%) identified as DE using DESeq (p<0.05)

34. Intensity difference test
Different approach to differential expression
Doesn’t aim to find every differentially expressed gene
Conservative test
Guaranteed to never return large numbers of hits

35. Assumptions Noise is related to observation level
Similar to DESeq
Differences between conditions are either
A direct response to stimulus
Noise, either technical or biological
Find points whose differences aren’t explained by general disruption

36. Method

37. Results

38. Exercises Look at raw QC Mapping with tophat
Small test data
Quantitation and visualisation with SeqMonk
Larger replicated data
Differential expression with DESeq
Review in SeqMonk

39. Useful links
FastQC
Tophat
SeqMonk
Cufflinks
DESeq
Bioconductor
RSEM