RNA-seq: From experimental design to
gene expression.
Steve Munger
@stevemunger
The Jackson Laboratory
UMaine Computational Methods in Biology/ Genomics
September 29, 2014
Outline
General overview of RNA-seq analysis.
• Introduction to RNA-seq
• The importance of a good experimental design
• Quality control
• Read alignment
• Quantifying isoform and gene expression
• Normalization of expression estimates
Understanding gene expression
Alwine et. al. PNAS 1977
ON/OFF
1.5
DeRisi et. al. Science 1997
1.5
10.5
3
Next Generation Genome Sequencers
Illumina HiSeq and MiSeq
PacBio SMRT
454 GS FLX
Oxford Nanopore
N
RNA-Seq: Sequencing Transcriptomes
RNA
N
ATGCTCA AGCTA
TAGATGCTCA AGCTA
ATGCTCA AGCTAATC
ATGCTCA AGCTA
AGTAGATGCTCA AGCTA
ATGCTCA AGCTA
ATGCTCA AGCTA
ATGCTCA AGCTA
TAGATGCTCAAGCTAATC
CTCA AGCTAATCCTAG
Applications of RNA-Seq Technology
Gene expression analysis
Novel Exon discovery
A/G
AGCTAATCCTAG
TAGATGCTCAA AGCTAATCCTAG
ATGCTCA A AGCTA
TAGATGCTCA A AGCTA
ATGCTCA A AGCTAATC
ATGCTCA A AGCTA
AGTAGATGCTCA A AGCTA
ATGCTCA G AGCTA
ATGCTCA G AGCTA
ATGCTCA G AGCTA
TAGATGCTCA G AGCTAATC
CTCA G AGCTAATCCTAG
ATGCTCA A AGCTA
TAGATGCTCA A AGCTA
ATGCTCA A AGCTAATC
ATGCTCA A AGCTA
AGTAGATGCTCA A AGCTA
ATGCTCA A AGCTA
ATGCTCA A AGCTA
ATGCTCA G AGCTA
TAGATGCTCA G AGCTAATC
CTCA G AGCTAATCCTAG
TAGATGCTCA
N
Allele Specific Gene Expression
RNA Editing
Total RNA
RNA-Seq
mRNA
mRNA after
fragmentation
cDNA
Adaptors ligated
to cDNA
Single/ Paired End
Sequencing
N
Challenges in RNA-Seq Work Flow
Study Design
RNA isolation/ Library Prep
Sequencing Reads (SE or PE)
Aligned Reads
Quantified isoform and gene expression
N
Know your application – Design your
experiment accordingly
• Differential expression of highly expressed and well
annotated genes?
– Smaller sample depth; more biological replicates
– No need for paired end reads; shorter reads (50bp) may be
sufficient.
– Better to have 20 million 50bp reads than 10 million 100bp
reads.
• Looking for novel genes/splicing/isoforms?
– More read depth, paired-end reads from longer fragments.
• Allele specific expression
– “Good” genomes for both strains.
N
Good Experimental Design
One Illumina HiSeq Flowcell = 8 lanes
Multiplex up to 24 samples in a lane
Multiplexing
Replication
Randomization
187 Million SE reads per lane
374 Million PE reads per lane
Cost per lane:
• 100 bases PE: $1800 - $2700
• 50 bases PE: $1400 - $2100
• 50 bases SE: $800 - $1200
N
RNA-Seq Experimental Design: Randomization
Experimental Group 1
Experimental Group 2
Two Illumina Lanes
Random.org
Bad Design
N
RNA-Seq Experimental Design: Randomization
Experimental Group 1
Experimental Group 2
Two Illumina Lanes
Random.org
Bad Design
Good Design
N
RNA-Seq Experimental Design: Randomization
Experimental Group 1
Experimental Group 2
Two Illumina Lanes
Random.org
Bad Design
Good Design
Better Design
N
Challenges in RNA-Seq Work Flow
Study Design
RNA isolation/ Library Prep
Sequencing Reads (SE or PE)
Aligned Reads
Quantified isoform and gene expression
N
Total RNA
mRNA
mRNA after
fragmentation
cDNA
poly-A tail selection
Not actually random
“Not So” random
primers
Size selection step
(gel extraction)
Adaptors ligated
to cDNA
PCR amplification
Adapter dimers
S
Sequence Read: Sanger fastq format
@HISEQ2000_0074:8:1101:7544:2225#TAGCTT/1
TCACCCGTAAGGTAACAAACCGAAAGTATCCAAAGCTAAAAGAAGTGGACGACGTGCTTGGTGGAGCAGCTGCATG
+
CCCFFFFFHHHHDHHJJJJJJJJIJJ?FGIIIJJJJJJIJJJJJJFHIJJJIJHHHFFFFD>AC?B??C?ACCAC>BB<<<>C@CCCACCCDCCIJ
@HISEQ2000_0074:8:1101:7544:2225#TAGCTT/1
The member of a pair
Instrument: run/flowcell id
Flowcell lane and tile number
X-Y Coordinate in flowcell
Phred Score:
Q = -10 log10 P
10 indicates 1 in 10 chance of error
20 indicates 1 in 100,
30 indicates 1 in 1000,
Index Sequence
SN
To Trim or not to Trim?
QUALITY CONTROL
S
Quality Control: Sequence quality per base position
Good data
 Consistent
 High Quality Along the reads
Bad data
 High Variance
 Quality Decrease with Length
S
Per sequence quality distribution
bad data
Y= number of reads
X= Mean sequence quality
Average data
NGS Data Preprocessing
S
Per sequence quality distribution
bad data
Good data
Y= number of reads
X= Mean sequence quality
Average data
NGS Data Preprocessing
S
Quality Control: Sequence Content Across Bases
S
K-mer content
counts the enrichment of every 5-mer
within the sequence library
Bad: If k-mer enrichment >= 10 fold at any
individual base position
NGS Data Preprocessing
K-mer content
Most samples
Duplicated sequences
Good: non-unique sequences make
up less than 20%
Bad: non-unique sequences make
>50%
NGS Data Preprocessing
S
PCR duplicates or high expressed
genes? The case of Albumin.
~80,000x coverage here
S
Tradeoffs to preprocessing data
• Signal/noise -> Preprocessing can remove lowquality “noise”, but the cost is information loss.
– Some uniformly low-quality reads map uniquely to the
genome.
– Trimming reads to remove lower quality ends can
adversely affect alignment, especially if aligning to the
genome and the read spans a splice site.
– Duplicated reads or just highly expressed genes?
– Most aligners can take quality scores into consideration.
– Currently, we do not recommend preprocessing reads
aside from removing uniformly low quality samples.
S
The problem with trimming all SE reads
100bp reads
All reads trimmed to 75 bp
Longer is better for splice junction spanning reads
S
Quality Control: Resources
• FASTX-Toolkit
– http://hannonlab.cshl.edu/fastx_toolkit/
• FastQC
– http://www.bioinformatics.babraham.ac.uk/projects/fastq
c/
NGS Data Preprocessing
S
RNA-Seq Work Flow
Study Design
Total RNA
Sequencing Reads (SE or PE)
Aligned Reads
Quantified isoform and gene expression
KB
Alignment 101
100bp Read
ACATGCTGCGGA
Chr 3
ACATGCTGCGGA
Chr 2
Chr 1
KB
The perfect read: 1 read = 1 unique alignment.
100bp Read
ACATGCTGCGGA
✓
Chr 3
ACATGCTGCGGA
Chr 2
Chr 1
KB
Some reads will align equally well to multiple
locations. “Multireads”
100bp Read
✗
ACATGCTGCGGA
✓
✗
ACATGCTGCGGA
ACATGCTGCGGA
ACATGCTGCGGA
1 read
3 valid alignments
Only 1 alignment is correct
KB
The worst case scenario.
100bp Read
✗
ACATGCTGCGGA
ACATGCTGCGGA
✗
ACATGCTGCGGA
ACATGCTGCGGC
1 read
2 valid alignments
Neither is correct
KB
Individual genetic variation may affect read
alignment.
100bp Read
✗
ATATGCTGCGGA
ACATGCTGCGGA
✗
ACATGCTGCGGC
ACATGCTGCGGA
KB
Aligning Billions of Short Sequence Reads
Gene A
Gene B
Aligners: Bowtie, GSNAP, BWA, MAQ, BLAT
Designed to align the short reads fast, but not accurate
KB
Aligning Sequence Reads
Gene A
•
•
•
•
•
•
•
•
Gene B
Gene family (orthologous/paralogous)
Low-complexity sequence
Alternatively spliced isoforms
Pseudogenes
Polymorphisms
Indels
Structural Variants
Reference sequence Errors
KB
Align to Genome or Transcriptome?
Genome
Transcriptome
KB
Aligning to Reference Genome: Exon First Alignment
Exon 1
Exon 2
Exon 3
KB
Aligning to Reference Genome: Exon First Alignment
Exon 1
Exon 2
Exon 3
KB
Aligning to Reference Genome: Exon First Alignment
Exon 1
Exon 2
Exon 3
KB
Exon First Alignment: Pseudo-gene problem
Exon 1
Exon 2
Exon 3
Processed Pseudo-gene
Exon 1
Exon 2
Exon 3
KB
Exon First Alignment: Pseudo-gene problem
Exon 1
Exon 2
Exon 3
Processed Pseudo-gene
Exon 1
Exon 2
Exon 3
KB
Exon First Alignment: Pseudo-gene problem
Exon 1
Exon 2
Exon 3
Processed Pseudo-gene
Exon 1
Exon 2
Exon 3
KB
Align to Genome or Transcriptome?
Genome
Advantages: Can align novel isoforms.
Disadvantages: Difficult, Spurious alignments, spliced alignment, gene families, pseudo genes
Transcriptome
KB
Reference Transcriptome
Gene 1
Exon 1
Exon 2
Exon 3
Isoform 1
Isoform 2
Isoform 3
Gene 2
Exon 1
Exon 2
Exon 3
Exon 4
Isoform 1
Isoform 2
Isoform 3
KB
Align to Genome or Transcriptome?
Genome
Advantages: Can align novel isoforms.
Disadvantages: Difficult, Spurious alignments, spliced alignment, gene families, pseudo genes
Transcriptome
Advantages: Easy, Focused to the part of the genome that is known to be transcribed.
Disadvantages: Reads that come from novel isoforms may not align at all or may be
misattributed to a known isoform.
KB
Better Approach: Aligning to Transcriptome and Genome
Align to Transcriptome First
Align the remaining reads to Genome next
Advantages: relatively simpler, overcomes the pseudo-gene and
novel isoform problems
RUM, TopHat2, STAR
KB
Conclusions
• There is no perfect aligner. Pick one well-suited to your
application.
– E.g. Want to identify novel exons? Don’t align only to the known
set of isoforms.
•
•
Visually inspect the resulting alignments. Setting a
parameter a little too liberal or conservative can have a
huge effect on alignment.
Consider running the same fastq files through multiple
alignment pipelines specific to each application.
–
–
–
•
Gene expression -> Bowtie to transcriptome
Exon discovery -> RUM or other hybrid mapper
Variant detection -> GSNAP, GATK, Samtools mpileup
If your species has not been sequenced, use a de novo
assembly method. Can also use the genome of a related
species as a scaffold.
KB
Output of most aligners: Bam/Sam file
of reads and genome positions
S
Visualization of alignment data (BAM/SAM)
• Genome browsers – UCSC, IGV, etc.
IGV is your friend.
SNP
Read color = strand
Coverage density plot
Aligned Reads to Gene Abundance
Total RNA
100bp Reads
Aligned Reads
Quantified isoform and gene expression
N
Aligned Reads to Gene Abundance: Challenges
Long
Short
Many approaches to quantify expression abundance
N
Aligned Reads to Gene Abundance: Challenges
200
1
Long
1000 reads
100
2
Medium
50
3
Short
Relative abundance for these genes, f1, f2, f3
N
Aligned Reads to Gene Abundance: Challenges
400
200
1
Long
100
2
Medium
50
3
400
200
Short
Relative abundance for these genes, f1, f2, f3
N
Aligned Reads to Gene Abundance: Challenges
400
200
1
Long
100
2
Medium
50
3
400
200
Short
Relative abundance; f1=0.2, f2=0.4, f3=0.4
N
Multireads: Reads Mapping to Multiple Genes/Transcripts
Long
Medium
N
Gene multireads
Gene A
Gene B
Gene C
Reads that map to multiple genomic locations
N
Isoform multireads
Exon 1
Exon 2
Exon 3
Isoform 1
Isoform 2
Isoform 3
N
Multireads: Reads Mapping to Multiple Genes/Transcripts
350
200
1
Long
150
100
2
Medium
300
Multireads
50
3
200
Short
Unique
Relative abundance for these genes, f1, f2, f3
N
Approach 1: Ignore Multireads
350
200
1
Long
150
100
2
Medium
50
3
300
200
Short
Relative abundance for these genes, f1, f2, f3
N
Nagalakshmi et. al. Science. 2008
Marioni, et. al. Genome Research 2008
Approach 1: Ignore Multireads
350
200
1
Long
150
100
2
Medium
50
3
300
200
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
Approach 2: Allocate Fraction of Multireads Using Estimates
From Uniques
350
200
1
Long
150
100
2
Medium
50
3
300
200
Short
Relative abundance for these genes, f1, f2, f3
Ali Mortazavi, et. al. Nature Methods 2008
N
RSEM,Cufflinks
Multireads: Reads Mapping to Multiple Genes/Transcripts
Long
Medium
N
Conclusions for quantitation
• Isoform-level estimates will become easier as
read length increases.
• EM approaches are currently the best option.
• Bad alignment = bad quantitation
• Pseudogenes are problematic.
• Check your alignments -> adjust method.
Iterative process.
N
A speed bump on the road from raw counts to differential expression.
NORMALIZATION
S
Large pool, small sample problem
• Typical RNA library estimated to contain 2.4 x 1012
molecules. McIntyre et al 2011
• Typical sequencing run = 25 million reads/sample.
• This means that only 0.00001 (1/1000th of a percent)
of RNA molecules are sampled in a given run.
• High abundance transcripts are sampled more
frequently.
Example: Albumin = 13% of all reads in liver RNA-seq
samples.
• Sampling errors affect low-abundance transcripts
most.
S
A finite pool of reads.
S
Sample 1
Alb
Low1
Perfect world:
All transcripts
counted.
Sample 2
Alb
Low1
S
Sample 1
Alb
Real world: More reads
taken up by highly
expressed genes means less
reads available for lowly
expressed genes.
Low1
S
Sample 1
Sample 2
Alb
Alb
Low1
Highly expressed genes that are
differentially expressed can
cause lowly expressed genes
that are not actually
differentially expressed to
appear that way.
Low1
S
Normalization of raw counts
• Wrong way to normalize data
– Normalizing to the total number of mapped reads
(e.g. FPKM). Top 10 highly expressed genes soak
up 20% of reads in the liver. FPKM is widely used,
and problematic.
• Better ways to measure data
– Normalize to upper quartile (75th %) of non-zero
counts, median of scaled counts (DESeq), or the
weighted trimmed mean of the log expression
ratios (EdgeR).
S
Normalizing for gene length: Longer transcripts will produce
relatively more reads when fragmented.
Sample 1
Long Gene
Short Gene
Only necessary if you are comparing the relative expression
levels of two genes in one sample. TPM = Transcripts per
million.
S
Summary
RNA
ATGCTCA AGCTA
TAGATGCTCA AGCTA
ATGCTCA AGCTAATC
ATGCTCA AGCTA
AGTAGATGCTCA AGCTA
ATGCTCA AGCTA
ATGCTCA AGCTA
ATGCTCA AGCTA
TAGATGCTCAAGCTAATC
CTCA AGCTAATCCTAG
As sequences get longer, alignment and isoform
quantitation becomes easier!
RNA
ATGCTCA AGCTA
TAGATGCTCA AGCTA
ATGCTCA AGCTAATC
ATGCTCA AGCTA
AGTAGATGCTCA AGCTA
ATGCTCA AGCTA
ATGCTCA AGCTA
ATGCTCA AGCTA
TAGATGCTCAAGCTAATC
CTCA AGCTAATCCTAG
QC and Analysis Pipeline
Experimental Design
Summary
Resources
Aligner
– Bowtie 2 http://bowtie-bio.sourceforge.net/bowtie2/index.shtml
– GSNAP http://research-pub.gene.com/gmap/
– Tophat http://tophat.cbcb.umd.edu/
Transcript Abundance
– Cufflinks http://cufflinks.cbcb.umd.edu/
– IsoEM http://dna.engr.uconn.edu/?page_id=105
– RSEM http://deweylab.biostat.wisc.edu/rsem/
– Miso http://genes.mit.edu/burgelab/miso/
– Cuffdiff http://cufflinks.cbcb.umd.edu/
– DESeq http://www-huber.embl.de/users/anders/DESeq/
– edgeR http://bioconductor.org/packages/release/bioc/html/edgeR.html
Denovo Assembly
– Trinity http://trinityrnaseq.sourceforge.net/
Acknowledgements
•
•
•
•
•
•
•
•
•
Narayanan Raghupathy, KB Choi
Gary Churchill
Ron Korstanje/ Karen Svenson/ Elissa Chesler
Joel Graber
Anuj Srivastava
Churchill Lab – Dan Gatti
Al Simons and Matt Hibbs
Lisa Somes, Steve Ciciotte, mouse room staff
Mice
Thank you!
Differential Expression
Over-estimation of
Under-estimation of
Too conservative
Too sensitive
(Many false positives)
Expression abundance tends to overdispersed in RNA-seq measurements.
Simple Poisson model would result in under-estimation of variance and
many false positive calls.
Differential Expression
Poisson
A term that addresses
over-dispersion
The majority of DE detection methods
are trying to estimate this guy in their
own way.
(ex) DESeq, edgeR, EBSeq, DSS, NBPSeq, baySeq, etc.
Multiple Testing Correction and False Discovery rate
XKCD Significant
2012 IgNobel prize in
Neuroscience for “finding
Brain activity signal in dead salmon using fMRI”
N
Northern Blot