1. Canadian Bioinformatics Workshops

2. Module #: Title of Module2

3. Module 6 Metatranscriptomics
John Parkinson
Analysis of Metagenomic Data
June 22-24, 2016

4. Learning Objectives of Module
At the end of this module the student will have an appreciation of the opportunities and challenges of metatranscriptomics by:
Understanding the capabilities of metatranscriptomics
Gaining an appreciation of sample collection and experimental design
Learning important steps in data processing
Being able to process a simple metatranscriptomic dataset

5. Overview What is metatranscriptomics – how does it relate to RNA-Seq?
Experimental design, sample collection and preparation
Processing of reads
Filters
Assembly
Dereplication
Functional annotation
Taxonomic annotation
Statistical analysis
Visualization
Consequtive basepairs

6. Metagenomics and metatranscriptomics reveal function
16S rRNA surveys (“Who is there?”) have been widely applied but yield only limited mechanistic insights – cause or consequence?
Metagenomics (“What can they do?”) have revealed dramatic differences in community composition but with conserved microbiome functions
Metatranscriptomics (“Who is doing what?”) examines microbiome activity
Frank et al Cell Host Microbe 2008 / The human microbiome consortium Nature 2012/ Xiong et al PLoS ONE 2012

7. Metatranscriptomics focuses on community activity
Metatranscriptomics exploits RNA-Seq to determine which genes and pathways are being actively expressed within a community
Genes involved in pathways associated with cell wall biogenesis
Metatranscriptomics can reveal active functions
It can also reveal which taxa are responsible for the active functions
Consequtive basepairs
Relative abundance
Relative abundance and contributing taxa

8. Similar microbiomes can express different functions
Distribution of Taxa (n=4)
Perilipin2 KO mice fed a high fat diet exhibited a similar microbiome as WT mice fed the same diet….
Acetolactate synthatse
…however metabolic pathways are differentially expressed – host genotype impacts microbiome function
Isopropylmalate isomerase
Xiong et al In Preparation

9. Metatranscriptomics through RNA Seq
Species B
Species D
Species A
Species C
Extract RNA
Fragment
and sequence
RNA-Seq is the unbiased sequencing of an RNA sample to yield a digital readout of the relative expression of transcripts within a sample
Typically applied to organisms with a reference (sequenced) genome, microbiome applications face a number of challenges
Gene A1 6 3 1 2 4
Gene A2
Gene A3
Gene A4
Gene B1
Gene B2
Gene B3
Gene C1
Consequtive basepairs
Align reads to known transcripts to obtain relative expression
Gene D4

10. Metatranscriptomics: Challenges
In a typical RNA-Seq experiment applied to a single eukaryotic organism, mRNA is isolated. After fragmentation and sequencing, reads are mapped to a reference genome using standard software such as MAQ and BWA to provide: 1) support that the transcript is expressed; 2) the relative abundance of the transcript; and 3) the presence and abundance of isoforms
For microbiome samples we have the following problems:
Lack of a polyA signal makes it difficult to isolate bacterial mRNA and resulting in (massive) rRNA contamination
Environmental microbiome samples lack reference genomes making it difficult to map reads back to their source transcripts
Consequtive basepairs

11. A typical metatranscriptomic analytical pipeline
1. Obtain RNA
2. Prep for sequencing
3. Generate Reads
4. Remove low quality
5. Remove rRNA reads
6. Remove host reads
7. Assemble
7. Identify bacterial transcript
8. Map to pathways
Gene A
Gene C
9. Sample comparisons
Gene B
Gene D

12. 1. Sample collection and RNA extraction
RNA quality deteriorates rapidly – Method of storage and preparation can impact taxa recovered and has yet to be standardized
Best(?): Process immediately to extract RNA then store at -80
Next best(?): Snap freeze in liquid nitrogen and store at -80
Avoid use of RNALater – it lyses some cells and can interfere with RNA extraction kits (e.g Ambion RiboPure Bacteria Kit / LifeTech Trizol Plus)
Number of biological replicates
“At least 2!”
Power analyses challenging
Cost per sample (20 million reads)
~$300-$400

13. 2. Preparing sample for sequencing
Bacterial mRNA’s lack a polyA tail so how to remove abundant rRNA species?
Once RNA has been extracted, several kits are available to remove rRNA – need 500ng-2.5ug RNA/sample
Ribo-Zero (Illumina) provides
reasonable success
mRNA
Host
rRNA
Adaptor / Low quality
Host mRNAs can also prove challenging – can also be informative!

14. 2. Preparing sample for sequencing
Temperature (4oC > -20oC)
Kits (RiboZero > Mxpress)
Storage Days (Fresh > 1 week)
RNA later (+ > -)
Comparisons of storage and RNA treatments

15. 3. Generating reads
How many reads are “enough”?
While 454 and MiSeq provide long reads useful for annotation HiSeq (or NextSeq) provide sequencing depth and offer possibility of multiplexing
~5 million mRNA reads provide 90-95% of ECs in a microbiome
With kits yielding mRNA read rates of ~25%, this suggests 20 million/sample mRNA

16. 4. Analysing the data Metatranscriptomics is a relatively new field
Assembly methods?
Annotation methods?
Analysis methods?
Processing methods?
Least established
Most established
Metatranscriptomics is a relatively new field
Robust software standards and methods still to be developed
New tools continually created – can swap in as they prove more effective
Celaj et al Microbiome 2014
Jiang et al Microbiome 2016

17. Read processing - filtering
To identify reads derived from mRNA bioinformatics pipelines need to be in place that remove contaminating reads:
Low quality - Trimmomatic
Adaptors – Trimmomatic Host – BWA / BLAT
rRNA – BLAT / Infernal
Assembly methods?
Annotation methods?
Analysis methods?
Processing methods?
Least established
Most established
Trimmomatic uses a sliding window approach from the 5` end to identify low quality regions which are then trimmed from the 3` end. Reads < 36 bp are discarded

18. Read processing - Assembly
contig1
contig2
Assembly methods?
Annotation methods?
Analysis methods?
Processing methods?
Least established
Most established
attagcggcgattttcggcgatcttatcttgatctgggcgcgtatcggtagcgtagcgattcgtagc
Assembly improves annotation accuracy
Trinity appears to provide best performance in terms of reads that can be annotated
Chimera’s, misassembled contigs, can become a problem due to reads derived from orthologs from different species

19. Read processing – functional annotation
Assembly methods?
Annotation methods?
Analysis methods?
Processing methods?
Least established
Most established
Functional annotations rely on the identification of sequence similarity searches using tools such as:
BWA, BLAT and BLAST
While fast, BWA and BLAT rely on near perfect matches to reference genomes
However sequence diversity is huge!
Every time a new strain of S. agalactae was sequenced, new genes are discovered that were not present in existing strains
Diversity at the nucleotide level is >> protein level

20. Read processing – functional annotation
One solution is to work in peptide space and use BLASTX to search protein databases - this is very time consuming and requires cloud/cluster computing
Other solutions include DIAMOND and USEARCH (issues over quality and cost)
Even with BLAST many reads remain unannotated
Can be improved with longer read length

21. Quality of BLASTX matches
Typical match to a 71bp read
E-Value = 39 (NOT e-39)
Species looks about
right though
% of Read Length 55 60 65 70 75 80 85 90 95
100 35
0.0 40 45 50
0.1
0.2
0.4
0.3
0.5
0.7
1.6
0.9
1.3
1.5
2.2
1.7
3.8
2.0
2.4
4.9
1.1
0.8
1.8
2.7
3.3
0.6
1.9
4.0
3.7
4.8
2.9
8.6
12.1
Summary for all matches
% ID of match

22. Read processing – converting mappings to expression
To normalize expression levels to account for differences in gene length, read counts are converted to Reads per kilobase of transcript mapped (RPKM)
Expression is biased for gene length (longer transcripts should have more reads) to normalize, reads are converted to Reads per Kilobase of transcript per million reads mapped
RPKMgeneA = 109 CgeneA / NL
CgeneA = number of reads mapped to geneA
N = total number of reads
L = length of transcript in units of Kb #
RPKM 8 1 24 6 8 12
Several software tools available to do mapping and calculate normalized expression measurements across different samples including Bowtie and Cufflinks

23. Read processing – taxonomic annotation
Previous studies have shown that microbiomes can vary significantly in taxonomic contributions while yielding similar functionality
Assembly methods?
Annotation methods?
Analysis methods?
Processing methods?
Least established
Most established
Knowing which species are providing which functions may not be that important - On the other hand, assigning RNA reads to taxa may reveal critical functions contributed by keystone taxa, can also help in binning for assembly
How do we assign taxonomic information?
The human microbiome consortium 2012

24. Read processing – taxonomic annotation
Alignment based methods such as BLAST and BWA can fail where we lack suitable reference genomes – particularly for short read datasets where assignments may be ambiguous
Compositional methods (e.g. nt frequency, codon bias) offer alternative strategies
NBC
Nearest neighbours methods then try to assign a sequence to the genome with the closest distribution
CLARK
Here a sequences is classified into frequencies of 3-mers
KRAKEN

25. Taxonomic assignment through Gist
mRNA
genomes (*.ffn)
taxonomy database
16S counts
other abundances
classifier construction
environment description
taxonomic assignment
class definitions
taxonomy tags
reshape into strain-specific records (DELIN)
generate class distributions
AA NB
AA CD
AA 1NN
AA GMM
BWA
NT NB
NT CD
NT 1NN
NT GMM
class priors
log normalization
infer protein sequence
FragGeneScan
amino acid sequences
Classification 1
select candidate hits for detailed processing for each read
Classification 2
select best hits
final confidence measurements
(T-test for hit specificity)
final assignment scores
CalcTaxonInclusion
initial confidence measurements with T-test for hit specificity
Gist is a computational pipeline for accurate assignment of reads to individual species
Integrates several methods, but uniquely assigns different weights to methods for each genome
Can also take in expected sequence distributions (e.g. based on 16S rRNA surveys)

26. GIST exploits several orthogonal methods to define genome signatures
Naïve Bayes
average + variance
Nearest Neighbor
exact gene composition
Gaussian mixture
subclasses of genes
Expected Codelta
signature features
19:20389:F:275+18M2D19M M2
BWA
exact alignment hit
FragGeneScan
Protein translation

27. Assigning genome specific weights for each method
Increasing weight of method used to score genome
Methods
A key feature of Gist is defining combinations of weights for each feature / method that best discriminate between different genomes
For example, some genomes may perform better using alignment methods than compositional methods

28. Gist performance against a mouse metatranscriptome
The performance of Gist was compared with two other classifiers - MetaCV and NBC using three mouse metatranscriptome datasets. Each mouse was raised under germ free conditions and seeded with altered Schaedler flora (ASF – comprises 8 taxa). MetaCV is unable to assign ~30-40% of reads, while NBC assigns reads to ~15 main taxa. Gist assigns reads to ~9 main taxa, closely reflecting ASF taxa.

29. mRNA expression is not equivalent to rRNA abundance
Comparisons between rRNA and mRNA read distributions reveal subtle differences in read abundance
Artefact due to biases in rRNA sequencing?
Abundance != expression?

30. Visualizing results
Assembly methods?
Annotation methods?
Analysis methods?
Processing methods?
Least established
Most established
Once reads have been assigned to transcripts, we can explore their function through e.g. functional categories such as Gene Ontology or COG
However these types of analyses tend not to be very informative as the categories are very broad
Transcription
Energy production and metabolism
Lipid transport and metabolism

31. Visualizing results
Beyond focusing on broad functional categories, we can also start to undertake systems based analyses
Genes and proteins do not operate in isolation but form parts of interconnected functional modules
Signalling
networks
Metabolic pathways
By placing a bacterial transcripts within these functional contexts, we can understand how the microbiome functions at a molecular level
Protein complexes

32. Visualizing results
By mapping abundance of reads that map to E. coli homologs, protein interaction networks can be used as a template to provide more mechanistic insights – can we identify modules of genes that are coordinately expressed

33. Visualizing results
How do we identify those systems which are differentially regulated across samples (e.g. systems up-regulated in disease vs. health)?
Development of a statistically robust platform
(1) Gene set enrichment analysis (GSEA) requires discrete groups of functionally related genes
(2) Network approach; identify groups of genes based on functional interactions AND their differential expression
GSEA may define trpA-E
as a single group of
related genes, but group
as a whole may not be
significantly upregulated
A network based approach
may define a group based
on their interactions AND
expression

34. Visualizing results
Metabolic pathways are among the most highly conserved and best characterized systems
MG-RAST and MEGAN are automated metagenomic annotation tools that rely on KEGG
A major problem with KEGG pathway definitions is that the boundaries of pathways are arbitrarily defined and links between pathways (i.e. functional relationships) can be lost

35. Metabolic network of a mouse gut microbiome
Visualizing results
Combining taxonomic and functional information allows us to identify contributions of individual taxa within a microbiome
Metabolic network of a mouse gut microbiome
Piecharts indicate relative contribution of each taxon to an enzymatic activity
The interconversion of fructose-1,6 bisphosphate to glyceraldehyde 3-phosphate is largely mediated by bacteroides

36. Visualizing results
In addition to representing protein interaction or metabolic relationships through network visualization, Cytoscape offers the ability to represent nodes (proteins) as pie charts or donuts, allowing the visualization of taxonomic contributions to discrete functions

37. Statistical considerations
There is no dedicated software or statistical tool for statistical comparisons of metatranscriptomic datasets
Number of biological replicates? (at least two, preferably at least three but these are expensive experiments)
Power analyses?
Differential expression of individual genes
Gene set enrichment analyses
Ultimately metatranscriptomics could be viewed as hypothesis generating requiring subsequent targeted validation

38. Statistical considerations
While there are no dedicated tools for metatranscriptomics analyses, tools used for RNA Seq offer potential
DESeq, EdgeR, ALDEx
Alternatively simply rely on fold change
Challenges include which genes to include (minimum RPKM?)
Differentially expressed genes can be subsequently analysed through Gene Set Enrichment Approaches based on (for example) hypergeometric distributions

39. We are on a Coffee Break & Networking Session