1. Interrogating the transcriptome in all its diversity
Joel H Graber
Thanks- I actually wanted to include the word alternative in this title, but I also wanted to keep the title to two lines. But as I hope to convince you, it’s really the alternatives that make the processing interesting and important. So let’s put it into perspective.

2. Empirical transcript measurements to characterize mRNA processing
Stop codon
PolyA sites
mRNA/cDNA
ESTs
Microarray
probes
mRNA-Seq
mRNA-RACE

3. The most important thing I can tell you about (especially large-scale) transcriptome measurement
EVERY procedural step can leave its mark on the data
True of both bench and computational steps
What looks like interesting processing may not be
Know your assumptions
Be suspicious
Test and torture your data to be confident

4. Small numbers of genes to test: qPCR

5. IGF1 mRNA data indicates at least 15 or more transcript isoforms

6. qPCR Primer Pair Set-up should catch most isoform differences

7. Igf1 transcript variants are differentially expressed as a function of strain, tissue, and nutrition

8. Many genes simultaneously I: microarrays
The fundamental hypothesis of transcriptome measurement: the state of the cell can be ascertained by which transcripts are expressed at which levels

9. Using gene expression microarrays to assess variation in mRNA processing
Our modified hypothesis of transcript measurement
The activity of a cell is a function of which isoforms of which genes are expressed
Expression arrays, though designed for abundance measurement, can also reveal isoform variation
Summarization to one “expression level” is problematic
So the work I’ll show you is based primarily on the efforts of two graduate students, Jesse Salisbury and Priyam Singh.
As it says here at the top, we want to modify these measurements and not just get which genes are active, but also in what form.
We are using expression microarrays and explicitly talking about the affymetrix type of array with multiple probes targeting a transcript. The standard analysis is designed for measurement of abundance, but we can reprocess this data to reveal changes in isoform as well as abundance.
Tied up in this is the fact that for many genes, the summarization to just one number representing the activity of that gene is problematic, and in fact for some genes will miss interesting changes in gene activity.

10. Identifying processing changes with expression arrays: a simple example for illustration
One gene with two isoforms
Differing only in polyA site
Regulatory sites for post-transcriptional control only in extended isoform
Microarray probes hybridize to common and differential regions
Explain the model and specifically point to the probes.
Once again, we use the representation of the thickened region representing the coding sequence, and in this example the coding sequence of these two transcripts is identical, so that only the 3’-UTRs are different, and explicitly they differ only by their polyA site.
Just to make things interesting, we’ll include a couple of regulatory elements, that make the transcript behavior different in the presence of the right trans-acting factors.
So now, we have this gene, with two isoforms shown, and we’re going to do a microarray experiment in two different samples…

11. Standard microarray analysis can mask changes in mRNA processing2 1
Sample 1
Sample 2
So we start with two samples to be measured, and we’re showing the balance of the two isoforms as changing between them- explicitly, in this example, I’ve chosen to keep the total abundance the same.
The samples are collected and put onto the microarray. I’m showing here three replicates of each experiment in the raw data, with black representing sample 1 and orange sample 2. The signals for the 3 upstream, common probes are approximately equal, whereas the signals in the extended region are significantly higher for samples 2. In standard analysis, the signal for the probes are summarized to one value for each sample.
But as this example shows, this is not an adequate representation. So instead of summarizing within each sample, we instead compare each probe individually, taking a ratio for each probe- then we look for changes in this ratio along the length, looking for patterns like we see here where there is a clear difference between the different parts of the gene.
So with this analysis in place we applied it to a number of problems of interest. The one I’m going to tell you about is pro-B-cell lymphomas, which we studied in collaboration with Kevin MIlls
Sample 2
Sample 1
Salisbury J et al, PLoS ONE 2009

12. Statistical Test I: Modified t-test
Compute expression ratio for three or more probes on each side of putative break
Significance is assessed by randomizing probes
Spuriously low variance can cause problems

13. Systematic alternative processing of 3’-UTR can be correlated with functional changes
Science 2008
Cell 2009
Cancer
Research
2009

14. The Ube2a long transcript isoform is lost in all three tumor-types
Ubiquitin conjugating enzyme 2a is a signaling gene involved in response to DNA damage.
The layout of the 3’-end of the transcripts along with the microarray probes is shown at the top. Down below, we see a combined plot of the comparison of the three tumor types, as well as the mature b-cells against the wildtype proB cells. As with the illustration I showed before 0 here means no change from the wildtype pro-b-cells.
As can be seen here, all three tumors have essentially no change or maybe slight decrease in signal, but the extended isoform has a decrease between 3 and 8-fold. In contrast, the mature B-cells have a uniform decrease by approximately 2 fold across the entire transcript.
We also see the complementary pattern with preferential preservation of the long isoform:

15. The Pik3ap1 long transcript isoform is preserved in APC and APN, but not LPC tumors
Here we see phophoinositide kinase 3 activating protein 1. The layout of the information is the same as before with the arrangement of the 3’-end of the gene and probes shown at the top, and the probe level analysis shown below.
Pik3ap1 stands out from Ube2a in two notable ways: first of all the pattern is for elongation, rather than truncation with respect to proBs. Secondly, the pattern for Pik3ap1 is not constant across all tumors, but rather is only significant for the two artemis KO tumors. The Lig4 KO tumors show a uniform decrease in signal across all probes, whereas mature B-cells show essentially no change.
The fact that not all genes behave in a common manner led us to test the idea that the patterns of which genes were changed and by how much could be used to distinguish the tumors:

16. Summary: tumors have systematic and characteristic changes in RNA processing
Our data supports alternate 3’-processing as the source of changes rather than isoform-specific in stability
While truncation dominates, genes with elongation are also observed
APC and LPC tumors share an amplified oncogene (Myc), but differ in signature and prognosis
So to summarize what we have found:
It’s important to understand that the signals we observe could arise due to either alternative polyadenylation or equivalently to isoform specific changes in stability. Our work, as well as others, seems to indicate that changes in polyadenylation are playing a significant role; we found changes in the abundance of one critical transacting protein, as well as evidence of changes in a abundance or processing of the transcripts encoding several other polyA factors.
The role of genes with elongation is an open question that we will continue to evaluate. In the growing model that truncation of the 3’-UTR leads to general increase of protein production, elongation would presumably accordingly reduce protein.
Finally, we also performed our analysis on two sets of cell cultures, both drawn from human ovarian cancer, but differing in their response to cisplatin a common chemotherapeutic. With our analysis, we identified about 150 genes with significant changes, offering the possibility of predicting responsiveness to treatment based on RNA signatures.
One of the more intriguing findings is the differences in both phenotype and RNA processing between the APC and LPC tumors, since these share a common amplified oncogene. This suggests the possibility that the differences in DNA damage could be responsible, given the different roles of Art and Lig4. We are following this up:
Singh P et al, Cancer Research 2009

17. Differential splicing using whole-transcript microarrays
FIRMA: a method for detection of alternative splicing from exon array data
E. Purdom, K. M. Simpson, M. D. Robinson, J. G. Conboy, A. V. Lapuk and T.P. Speed
Bioinformatics (15):
Differential splicing using whole-transcript microarrays
M. D. Robinson and T.P. Speed
BMC Bioinformatics 2009, 10:156

18. FIRMA/FIRMAGene details
Full model:
i: array index
J: gene index
k: probe index
Residual:
Exon arrays
FIRMA score:
Gene arrays
FIRMAGene score:

19. RMA decomposition Heart Brain 75:25 B:H Total normalized expression
Estimated probe effect
RMA decomposition of probe-level Affymetrix data. Panel A shows the background-adjusted and normalized probe-level data for PRRX1, from the Affymetrix mixture dataset (see Methods). The probes are displayed in the order which they map to the human genome (not to scale), and lines join all probe intensities of the same sample. PRRX1 is expressed significantly higher in heart tissue compared to brain. Three replicates of pure heart tissue are shown as red lines; green lines represent pure brain tissue replicates and the blue lines represent a mixture of 75% brain tissue and 25% heart tissue. Panel B shows the estimated relative probe effects. Panel C shows the chip effects (i.e. summarized expression levels) and Panel D shows residuals, using the same colour scheme.
Estimated chip effect
Residual

20. Validation with known muscle-specific exons
FIRMA scores represented by color scale for all 11 of the validated probesets (left of dividing line) as well as 15 additional top scoring probesets (right of dividing line). Two validated genes (UNR and ITGB1) also ranked high enough to be included in the top 15 candidate genes and their labels are colored in green to note this. Note that the color scale is not evenly spaced but rather based on percentiles of all FIRMA scores in all non-filtered probesets and samples.
Purdom, E. et al. Bioinformatics : ; doi: /bioinformatics/btn284
Copyright restrictions may apply.

21. Open questions, unresolved issues of FIRMA approaches
Genetic variability
Probes that change hybridization due to sequence variation
Low expression probes
Overlapping probes
Unresponsive/hyperresponsive probes
Interpretation
IRLS is unsupervised; majority can define “normal”

22. CDFs (Chip definition files) MATTER!!!!!!
The CDF maps probes on the array to putative genes/transcripts
CDFs are explicitly dependent on the quality of the annotations used for association
Which CDF is best can literally depend on the specific gene of interest

23. IGF1 annotations (and genomic extent) depend greatly on the data source

24. Even after you get the CDS right, it’s still good to understand the limitations of your array

25. Exon-gene array differences
4 probes per PSR (probe selection region)
A mix of probable and improbable transcribed regions
Gene
Mostly “probable” regions
~25 probes per targeted transcript

26. Comparison of gene (FIRMAgene) and gene (FIRMA) arrays for MBP
Normalized probe-level data and RMA residuals for MBP. Panels A and B show the residuals for Gene and Exon for RMA fits, respectively. There are 36 probes for Gene and 72 probes for Exon. Both panels show 33 lines, one for each hybridization (11 tissues with 3 biological replicates each). The brain and muscle replicates are shown blue and red lines, respectively.
Brain
Muscle

27. Measuring isoform variation with mRNAseq

29. Analysis of large sets of short sequence reads is a rapidly developing field
Alignment first, assembly later
MAQ
Eland
SHRiMP
BowTie/TopHat
SOAP
Assembly first, alignment later
Trinity
transABySS
Oases

30. High throughput sequence data is aligned in conceptually the same way as BLAST
Better heuristics are necessary
The problem is bigger
Program tweaks are different
Tuned to small read size, large genome
(Rapid) Indexing is still the key
Initial attempts were based on standard hashing, later on Burrows-Wheeler Transform

31. Alignment first processing strategy
Enhanced Read Analysis of Gene Expression (ERANGE) and the allocation of multireads. (a) The main steps in the computational pipeline are
outlined at left, with different aspects of read assignment and weighting diagrammed at right and the corresponding number of gene model reads treated in
muscle shown in parentheses. In each step, the sequence read or reads being assigned by the algorithm are shown as a black rectangle, and their assignment
to one or more gene models is indicated in color. Sequence reads falling outside known or predicted regions are shown in gray. RNAFAR regions (clusters of
reads that do not belong to any gene model in our reference set) are shown as dotted lines. They can either be assigned to neighboring gene models, if they
are within a specified threshold radius (purple), or assigned their own predicted transcript model (green). Multireads (shown as parallelograms) are assigned
fractionally to their different possible locations based on the expression levels of their respective gene models as described in the text. (b) Comparison
of mouse liver expanded RPKM values to publicly available Affymetrix microarray intensities from GEO (GSE6850) for genes called as present by Rosetta
Resolver. Expanded RPKMs include unique reads, spliced reads and RNAFAR candidate exon aggregation, but not multireads. Genes with >30% contribution
of multireads to their final RPKM (Supplementary Fig. 4) are marked in red. (c) Comparison of Affymetrix intensity values with final RPKMs, which includes
multireads. Note that the multiread-affected genes that are below the regression line in b straddle the regression line in c.

32. Standard reporting of short reads: RPKM

33. A principal benefit of mRNAseq: novel exon/isoform discovery
Mortazavi et al

34. Better mapping to splice junctions
Align first to genome
Remove perfect matches from further consideration (Good idea?)
Remainder are aligned to a broadened set of possible splice junctions:
Extract ~25 bases from each exon, join and use as target
Standard, annotated splices
Additional
possible splices

35. RNA-seq analysis: alignment

36. RNA seq analysis II: identifying isoforms

37. RNAseq analysis III: visualization and interpretation

38. Assessing the ability to identify alternative isoforms

39. Trinity assembles reads to transcripts first
Grabherr et al, Nature Biotech 2011

40. Sequences in a de Bruijn graph

41. Open questions/problems
Dealing with the length bias
RPKM does not correctly normalize;
Optimal alignment
Still in development
Paralogs and common motifs are a problem
Depth of coverage for isoform characterization
Capture or focused chemistry helps

42. Summarization to total counts leads to false positives

43. Recent work I: Dealing with systematic bias in RNAseq data
Sources of bias
Fragmentation
“random priming”
Papers of note:
Biases in Illumina transcriptome sequencing caused by random hexamer priming
Hansen et al Nucleic Acids Research Volume38, Issue12 p. e131
Using non-uniform read distribution models to improve isoform expression inference in RNA-Seq
Wu et al, Bioinformatics /bioinformatics/btq696

44. Recent work 2: Focused Sequencing of PolyA sites
Standard mRNAseq does not adequately sample polyA sites
Recent directed studies:
Formation, regulation and evolution of Caenorhabditis elegans 3′UTRs
Jan et al, Nature Volume: 469, Pages: 97–101
Comprehensive polyadenylation site maps in yeast and human reveal pervasive alternative polyadenylation.
Ozsolak et al, cell 143(6): (2010)