1. Introduction
2. The Development Of Transcriptomics
   • Expressed Sequence Tags
   • Quantitative PCR (qPCR)
   • Microarrays
   • RNA Seq
3. Future Directions

1. What is the transcriptome?
2. Why do we study the transcriptome?

The transcriptome can be defined as:

• In general this can be summarised as the RNA content of a cell
  • Includes messenger RNA (mRNA), non-coding RNA (ncRNA), small RNA
  • Can also include transfer RNA (tRNA) and ribosomal RNA (rRNA), but generally doesn’t

• mRNA is the intermediary step between the genome (DNA) and proteins
• small RNA and ncRNA play significant roles in gene regulation
• Will be different in each individual cell but will share many common and specific features within cell/tissue types
• Is always a point-in-time snapshot of a dynamic process

• Is often focussed around the level of (or changes in) transcriptional activity at each locus:
  • Gene Expression Levels
  • Transcript usage within a locus
  • Allele specific expression - preferential expression of a parent allele
  • Expression quantitative trait loci (eQTL)
• May even discover new transcripts: de novo transcriptome assembly

• Differential Expression (DE) analysis often results in large lists of DE genes/transcripts - what next!? Biological context
  • Pathway Analysis
  • Identification of Transcription Factor binding motifs
• Make inference about the biological processes driving our observations
  • Targets for curing disease
  • Biomarkers for tumour detection etc.

Canonical mRNA Processing

The first attempt at capturing the transcriptome was in 1991

• Sequenced 609 mRNA human brain mRNA sequences
• Used ESTs ~100-800bp
• ESTs were generated by reverse transcribing poly-A selected mRNA

• “Gold-standard” for measurement of transcription levels
• Not a high-throughput technique
  • Targets a single transcript region with specific primers
  • cDNA produced is fluorescently labelled
• Measure Cycle Threshold (C_T) Value
  • Is the PCR cycle that fluorescence passes a given threshold

• A 10-fold dilution series

• Estimate absolute quantity of target
  • Using a dilution series and standard curve
• Estimate relative quantity of target
  • Compare C_T between samples

What might be a key consideration when comparing samples?

• Estimate absolute quantity of target
  • Using a dilution series and standard curve
• Estimate relative quantity of target
  • Compare C_T between samples
  • Housekeeper (HK) genes are used to correct for technical artefacts (inc. pipetting variability): Normalisation

• Relative abundances are often referred to as fold-change (FC)
  • Down regulation is sqeezed between 0 and 1
  • Up regulation is from 1 to \(infty\)
  • We often use log₂ fold-change get a more better scale e.g. 
    • A 2-fold up regulation in abundance \(\log_2 2 = 1\)
    • A 2-fold down regulation in abundance \(\log_2 \frac{1}{2} = -1\)
    • No change in abundance \(\log_2 1 = 0\)
  • This is often abbreviated as logFC
• In qPCR this value is \(\Delta\Delta C_T\) ^{[ ]}

1. First we calculate the difference in \(C_T\) between our gene and a housekeeping gene: \(\Delta C_T = C_{T[\text{gene}]} - C_{T[\text{HK}]}\)
2. Compare the \(\Delta C_T\) between samples: \(\Delta\Delta C_T = - (\Delta C_{T[\text{treat}]} - \Delta C_{T[\text{control}]})\)

• Assuming an amplification efficiency of 2 (i.e. doubling of quantity for every cycle)
  • \(\Delta\Delta C_T = 1 \implies\) logFC of 1

• Marked the birth of modern transcriptomics
• Purely interested in relative abundances
• Can assess expression levels for 1000’s of genes simultaneously

1. Fluorescent labelling during mRNA \(\rightarrow\) cDNA
2. Complimentary probes bind target sequences (hybridisation)
3. Fluorescence detection at each probe

The key assumption is Fluoresence \(\propto\) mRNA abundance

• Probes with known sequences at known locations
  • Probes were ~65mer complimentary cDNA
  • Originally printed in local facilities
• Samples are labelled with Cy3 (Green @ 570nm) or Cy5 (Red @ 670nm)
  • Both are hybridised to array
  • Relative Red/Green intensities were of interest
  • Gave an estimate of logFC within each array

• Became the dominant arrays until RNA seq
• Single channel (colour)

• 1,000,000 25-mer probes

• Probes targetting the 3’ end of transcripts - reduce issues with RNA degradation
• Usually about 11 probes per transcript (aka probeset)

• Exon/Gene arrays targeted the whole transcript - not just the 3’ end
• \(\leq\) 4 probes / exon
• No successful methods for determining alternate transcript usage
• Released in mid-2000s

• Illumina produced a microarray of 50-mer probes using barcoded beads
  • Multiple beads per gene \(\implies\) randomly distributed across the array
  • Dealt with spatial artefacts more effectively

• Still heavily used in methylation analysis (Illumina 450k and EpiC arrays)
• Sample is bi-sulphite treated prior to hybridisation
  • Converts unmethylated C to T (i.e Uracil)
• Probes for unmethylated (C \(\rightarrow\) T) and methylated (C \(\rightarrow\) C) sequences
• We compare proportions of methylated signal
• Target >450,000 genomic CpG dinucleotides

Two Initial Problems to solve

1. Adjust for overall differences in signal
   • Normalisation
2. Removal of Background Signal (non-specific binding + optical noise)
   • Background Correction

• We need the exact same amount of sample on each array
  • If one sample had more material hybridised, all genes would appear upregulated
  • In qPCR we used a ‘housekeeper gene’
• Can be performed in silico by adjusting probe intensities directly
• Quantile Normalisation is the most robust

1. Find the probe with the lowest signal on each array
2. Give each lowest signal’ probe the average of all ‘lowest signal’ probes
3. Find the probe with the second lowest signal on each array
4. Repeat until all probes have been given new values

• Final result is identical intensity distributions on all arrays
• Equivalent to having identical amounts of source material (mRNA)
• The lowest probe on each array will probably target a different gene

• Can be performed without using probe sequence information (Robust Multichip Average; RMA)
• Or by using probe sequence information (GC-RMA)
• Illumina arrays (50-mer probes) are less prone to non-specific binding than Affymetrix arrays (25-mer probes)

• After BG correction (RMA), we have an estimate of true signal for each probe
• \(\geq 11\) probes per gene (or transcript) per array
• A single estimate of expression is obtained by averaging probe signals within each array/sample

• Gives a continuous value: Assume Normality
• We then fit standard linear regression models for each gene
  • Or perform simple \(t\)-tests to obtain a \(p\)-value
  • \(H_0\): No difference in average gene expression levels
  • \(H_A\): Some difference in average gene expression levels

• We will perform 15-25,000 hypothesis tests
  • 25,000 genes where \(H_0\) is true would give ~ 1250 \(p < 0.05\)

• If we set the rejection value to \(\alpha* = \frac{\alpha}{m}\) we control the FWER at the level \(\alpha\)
• To ensure that \(p\)(one or more Type I errors) = 0.05 in our example:
  
  \(\implies\) Reject \(H_0\) if \(p < \frac{0.05}{25000} = 2 \times 10^6\)
• Controlling the FWER (Bonferroni) may often be too strict
  • Leads to a high Type II error rate (false negatives)

• An alternative is to allow a small number of Type I Errors in our results \(\implies\) we have a False Discovery Rate (FDR)
• Instead of controlling the FWER at \(\alpha = 0.05\), we control the FDR at \(\alpha = 0.05\):
  • We allow up to 5% of our list to be Type I Errors
  • Can perform downstream inference assuming that the signal will dominate the noise

• Probe intensities are considered continuous variables
• We can assume normally distributed data

• Microarray:
  • Probes designed with a prior knowledge of target sequences
  • Give a fluorescence intensity \(\propto\) expression level
• RNA-seq:
  • Contains specific sequence information
  • We count amplified fragments from source material
  • Can detect/quantify novel/unknown sequences
  • Can de novo assemble a transcriptome: novel transcripts
  • Possibility of detecting alternate isoform usage

• poly-A enriched (mRNA is extracted)
  • Will contain intact mRNA only (and ncRNA with polyA tail)
• Total RNA libraries (rRNA is reduced)
  • Will include ncRNA, pre-mRNA, partially degraded RNA
• small RNA libraries
  • Targets micro RNA (miRNA), piwi-interacting RNA (piRNA), etc.

1. Clean up raw reads
2. Align to genome or transcriptome
3. Count

Are we interested in activity at a locus:

• Summarise all reads to a gene \(\implies\) Differentially expressed genes

Are we interested in transcript-level information

• How do we count reads now? Reads can align to multiple transcripts
• Do we summarise splice junctions?
• Is alternate isoform usage the same as differential expression?

• Treating each gene as summarised by locus
• Highly analogous to microarray analysis
  • Except this is count data
  • Not Normally distributed

• Counts are traditionally modelled using the Poisson Distribution
  • This models a rate of occurrence per fixed unit
• Poisson \((\lambda)\) has mean = variance = \(\lambda\)
• What if our observed variance is greater than the mean?
  • The model is inappropriate

• Gene counts can be modelled as Negative Binomial data
  • Just like Poisson, except has an ‘dispersion’ parameter for when variance differs from the mean
  • We use this to model greater variability than the Poisson would allow
• Instead of \(T\)-tests:
  • Exact Test (Robinson & Smyth) for comparison of two groups, or;
  • Generalised Linear Models

• Counts are directly impacted by library size
  • Smaller libraries will have lower counts
• In contrast to microarrays, we do not directly adjust our counts
  • Include a term in models
  • dgeList$samples$norm.factors

• Do we need to account for Type I errors? \(\implies\) FDR
• Can we fit moderated \(T\) statistics?
• Can we down-weight samples

• This transforms negative binomial counts into continuous data
• Enables the assumption of Normality
• All microarray techniques are back in play!
  • Moderated \(T\)-tests
  • Weighted Regression

What happens after we have a list of DE genes?

What happens after we have a list of DE genes?

• Do we integrate our data with methylomic analysis?
• Is there co-expression in Topologically Associated Domains?
• Do we look for genes with high correlations \(\implies\) look for common TF binding sites in promoters?

We can test for enrichment of pre-defined gene sets in our data:

• Gene Ontology (GO);
• Kyoto Encyclopaedia of Genes & Genomes (KEGG) pathways
• Reactome pathways

Most common approach is Fisher’s Exact Test

Choosing suitable control/background genes is very important

• Counting alignments become important
• Kallisto aligns to transcriptome using \(k\)-mers
  • Produce pseudo-alignments, or list of possible transcripts a read can align to
  • Deals with multi-mapping problems
• Algorithms for analysis are far from settled!!!

• At the gene/locus level length is constant across treatment groups
• Under DTU, this assumption is no longer valid
  • Algorithms which use this show clear out-performance

An excellent talk/paper is:

Differential analyses for RNA-seq: transcript-level estimates improve gene-level inferences Charlotte Soneson, Michael I. Love, and Mark D. Robinson

• Pac-Bio Long reads will be better at correctly identifying transcripts
• Still too expensive for large scale quantitation

• All previous approaches have worked with a pool of starting cells
• Single Cell approaches are now appearing