1. Microbiome 16S analysis Morgan G. I. Langille, PhD Assistant Professor
Canada Research Chair in Human Microbiomics
Dalhousie University
Halifax, Nova Scotia, Canada
Oct. 19th, 2016

2. Interests PICRUSt 2.0 Multi-omics and machine learning
Integrating knowledge from hundreds of microbiome studies
Dabbling with MinIon sequencing
Exercise and the gut microbiome
Drug metabolism by the gut microbiome
Aging/frailty
Pediatric IBD
Pediatric cancer
Pediatric UTIs
Welder’s lungs
Wild blueberry agriculture
Beer

3. Module #: Title of Module3

4. Acknowledgements Will Hsiao and the Canadian Bioinformatics Workshop
Many slides in this presentation
Gavin Douglas, PhD student
Microbiome Helper Software and Tutorials
Andre Comeau, PhD
16S sequencing and Integrated Microbiome Resource
Various bioinformatic developer groups
QIIME, PEAR, STAMP, etc.

5. Learning Objectives
Understand and Perform marker based microbiome analysis
Analyze 16S rRNA marker gene data
Use 16S rRNA marker gene to profile and compare microbomes
Select suitable parameters for marker gene analysis
Explain the advantages and disadvantages of marker based microbiome analysis

6. Generating data
16S: Samples to DNA

7. rRNAs – the universal phylogenetic markers
Ribosomal RNAs are present in all living organisms
rRNAs play critical roles in protein translation
rRNAs are relatively conserved and thought to be rarely acquired horizontally
Behave like a molecular clock
Useful for phylogenetic analysis
Used to build tree-of-life (placing organisms in a single phylogenetic tree)
16S rRNA most commonly used
Consequtive basepairs
Universal phylogenetic tree based on SSU rRNA sequences
-N. Pace, Science, 1997

8. Other Marker Genes Used
Eukaryotic Organisms (protists, fungi)
18S (
ITS (
Bacteria
CPN60 (
ITS (Martiny, Env Micro 2009)
RecA gene
Viruses
Gp23 for T4-like bacteriophage
RdRp for picornaviruses
Faster evolving markers used for strain-level differentiation

9. A Few Considerations for Selecting Marker Genes
Sufficient resolution to differentiate the different communities you want to study (16S differentiates genera and some species; e.g. HSP65 may be more useful for Mycobacterium)
A (good) reference database is needed for taxonomic assignment
When comparing across studies, need to standardize markers (different V regions of 16S gene), experimental protocols, and bioinformatic pipelines

10. DNA Extraction Many protocols/kits available
Mo Bio Power Soil/Fecal/Food
(now owned by Qiagen)
DIY still popular
Different extraction methods will give different results
HMP and EMP standardize DNA extraction protocol
DNA Extraction can also be done after fractionation to separate out different organisms based on cell size and other characteristics
Gels
Flow Cytometry

11. Contaminations
Lab:
Extraction process can introduce contamination from the lab
Reagents may be contaminated with bacterial DNA
Include Extraction Negative Control in your experiments!
Especially crucial if samples have low DNA yield!
Host / Environment:
Host DNA often ends up in the microbiome sample
More of a problem with shotgun metagenomics
Low ratio of amplicon to other DNA can sometimes cause amplification problems

12. Target Selection and Bias
16S rRNA contains 9 hypervariable regions (V1-V9)
Different V regions have different phylogenetic resolutions – giving rise to slightly different community composition results
Jumpstart Consortium HMP, PLoS One, 2012

13. Variable Region Comparisond
William et al, 2015 mSystems
Comeau et al. in review

14. Target Amplification Adapter barcode primer Comeau et al. in review
PCR Primers designed to target specific region of a genome
Source: Illumina application Note
Comeau et al. in review

15. Stitching improves read quality
Comeau et al. in review

16. Sample Multiplexing
“Multiplexing” means combining samples on the same run
Unique DNA barcodes can be incorporated into your amplicons to differentiate samples
Dual-indexing uses 16 forward x 24 reverse = 384 samples
Amplicon sequences are quite homogeneous
Necessary to combine different amplicon types or spike in PhiX control reads to diversify the sequences
Not as big a problem with recent firmware release (only 5% PhiX needed)

17. Processing data
16S: DNA to OTU tables

18. Available Marker Gene Analysis Platforms
QIIME (
Mothur (

19. Comparison Between QIIME and Mothur
A python interface to glue together many programs
Single program with minimal external dependency
Wrappers for existing programs
Reimplementation of popular algorithms
Large number of dependencies / VM available; work best on single multi-core server with lots of memory
Easy to install and setup; work best on single multi-core server with lots of memory
More scalable
Less scalable
Steeper learning curve but more flexible workflow if you can write your own scripts
Easy to learn but workflow works the best with built-in tools

20. Overall Bioinformatics Workflow
1) Preprocessing: remove primers, demultiplex, stitch reads, quality filter
2) OTU Picking / Representative Sequences
Sequence data (fastq)
3) Taxonomic Assignment
5) Sequence Alignment
Metadata about samples (mapping file)
4) Build OTU Table (BIOM file)
6) Phylogenetic Analysis
Inputs
Processed Data
OTU Table
Phylogenetic Tree
7) Downstream analysis and Visualization – knowledge discovery
Outputs

21. Sample de-multiplexing
1) Preprocessing
Sample de-multiplexing
Reads need to be linked back to the samples they came from using the unique barcodes
De-multiplexing removes the barcodes and the primer sequences and separates reads into individual sample files
Some machines will demultiplex for you (MiSeq)
QIIME has several different scripts to help you prepare your sequence files for analysis

22. Read Stitching
Paired-end reads result in a forward and reverse read from the same sequence

23. Quality Filtering QIIME filtering by:
1) Preprocessing
Quality Filtering
QIIME filtering by:
Maximal number of consecutive low-quality bases ( -r=3 )
Minimum Phred quality score ( -q=3 )
Minimal length of consecutive high-quality bases (as % of total read length) ( -p=0.75 )
Maximal number of ambigous bases (N’s) (-n=0 )
Other quality filtering tools available
Cutadapt (
Trimmomatic (
Sickle (
Sequence quality summary using FASTQC

24. Chimera Sequence Removal
2) OTU Picking
Chimera Sequence Removal
PCR amplification process can generate chimeric sequences (an artificially joined sequence from >1 templates)
Chimeras represent ~1% of the reads
Detection based on the identification of a 3-way alignment of non-overlapping sub-sequences to 2 sequences in the search database
Source:

25. OTU Picking OTUs: formed arbitrarily based on sequence identity
97% of sequence similarity ≈ species
QIIME supports 3 approaches (each with several algorithms)
De novo clustering
Closed-reference
Open-reference
More details about OTU picking
Current recommendation outlined in

26. De novo Clustering Groups sequences based on sequence identity
Pair-wise comparisons needed
Hierarchical clustering commonly used
Requires a lot of disk/memory (hundreds Gb) and time
Suitable if no reference database is available or if a large proportion of novel taxa
Navas-Molina et al 2013, Meth. Enzym

27. De novo clustering
Exhaustive de novo clustering (pair-wise distance matrix followed by hierarchical clustering) is not scalable (e.g sequences requires 1000x1000 comparisons)
Greedy algorithms (e.g. UCLUST, sumaclust) achieves time/space saving by clustering incrementally using a subset of sequences as “centroids” (centroid = seed of a cluster)
Input order of sequences affects centroids picking so input seqs are usually pre-sorted by length or abundance

28. Closed-Reference
Matches sequences to an existing database of reference sequences
Unmatched sequences discarded
Fast and can be parallelized
Suitable if accurate and comprehensive reference is available
Allow taxonomic comparison across different markers and different datasets
Novel organisms missed/discarded
Navas-Molina et al 2013, Meth. Enzym

29. Open-Reference
First matches sequences to reference database (like closed-reference)
Unmatched sequences then go through de-novo clustering
Best of both worlds
Suitable if a mixture of novel sequences and known sequences is expected
Results could be biased by the reference database used
QIIME Recommended approach
Navas-Molina et al 2013, Meth. Enzym

30. No OTU picking! Collapsing sequences at 97% removes information
However, collapsing to 100% identity would allow a single nucleotide error to result in a spurious taxa
Use location of nucleotide differences to distinguish sequencing errors from real mutations

31. Representative Sequence
2) OTU Picking
Representative Sequence
Once sequences have been clustered into OTUs, a representative sequence is picked for each OTU.
This means downstream analyses treats all organisms in one OTU as the same!
Representative sequence can be picked based on
Abundance (default)
Centroid used (open-reference OTU or de novo)
Length
Existing reference sequence (for closed-reference OTU picking)
Random
First sequence

32. 3) Taxonomic Assignment
OTUs don’t have names but humans have the desire to refer to things by names.
I.e. we want to refer to a group of organism as Bacteroides rather than OTU1.
Closed-reference:
Transfer taxonomy of the reference sequence to the query read
Open-reference + de-novo clustering:
Match OTUs to a reference data using “similarity” searching algorithms
It is important to report how the matching is done in your publication and which taxonomy database is used

33. Taxonomic Assignment 3) Taxonomic Assignment Taxonomy DBs
Bacteria; Bacteroidetes/Chlorobi group; Bacteroidetes; Bacteroidia; Bacteroidales; Bacteroidaceae; Bacterroides
Bacteria; Firmicutes; Bacilli; Lactobacillales; Lactobacillaceae; Lactobacillus
Etc.
RDP Classifier
RDP
NCBI BLAST
GreenGenes
rtax
Silva
tax2tree
Specify a matching algorithm
Specify a target taxonomy database
OTUs
Ranked taxonomy names

34. Taxonomy Databases RDP (Cole et al 2009)
3) Taxonomic Assignment
Taxonomy Databases
RDP (Cole et al 2009)
Most similar to NCBI Taxonomy
Has a rapid classification tool (RDP-Classifier)
GreenGenes (DeSantis et al 2006)
Preferred by QIIME
Silva (Quast et al. 2013)
Preferred by Mothur
Choose DB based on existing datasets you want to compare to and community preference

35. OTU Table OTU table is a sample-by-observation matrix
4) Build OTU Table
OTU Table
OTU table is a sample-by-observation matrix
OTUs can have corresponding taxonomy information
Extremely rare OTUs (singletons or frequencies <0.1% of total reads) can be filtered out to improve downstream analysis
These OTUs may be due to sequencing errors and chimeras
Encoded in Biological observation Matrix (BIOM) format (
Sample1
Sample2
Sample3
Sample4
OTU1 10 14 33
OTU2 5 54 2
OTU3…

36. BIOM Format Advantages:
Efficient: Able to handle large amount of data; sparse matrix
Encapsulation: Able to keep metadata (e.g. sample information) and observational data (e.g. OTU counts) in one file
Supported: Have software applications to manipulate the content, check the content to minimize editing errors
Standardization: BIOM is a GSC ( standard; many tools use this file format (e.g. QIIME, MG-RAST, PICRUSt, Mothur, phyloseq, MEGAN, VAMPS, metagenomeSeq, Phinch, RDP Classifier)

37. Rarefaction
As we multiplex samples, the sequencing depths can vary from sample to sample
Many richness and diversity measures and downstream analyses are sensitive to sampling depth
“more reads sequenced, more species/OTUs will be found in a given environment”
So we need to rarefy the samples to the same level of sampling depth for more fair comparison
Alternative approach: variance stabilizing transformations (e.g. DeSeq2)
McMurdie and Holmes, PLOS CB, 2013
Liu, Hsiao et al, Bioinformatics, 2011

38. 5) Sequence Alignment
Sequence Alignment
Sequences must be aligned to infer a phylogenetic tree
Phylogenetic trees can be used for diversity analyses
Traditional alignment programs (clustalw, muscle etc) are slow (Larkin et al 2007, Edgar 2004)
Template-based aligners such as PyNAST (Caporaso et al 2010) and Infernal (Nawrocki, 2009) are faster and more scalable for large number of sequences

39. Phylogenetic Tree Reconstruction
6) Phylogenetic Analysis
Phylogenetic Tree Reconstruction
After sequences are aligned, phylogenetic trees can be constructed from the multiple alignments
Examples:
FastTree (Price et al 2009) – recommended by QIIME
Clearcut (Evans et al 2006)
Clustalw
RAXML (Stamatakis et al 2005)
Muscle

40. Visualization and statistics
What is important?
Visualization and statistics

41. Alpha Diversity Analysis
7) Downstream Analysis
Alpha Diversity Analysis
Alpha-diversity: diversity of organisms in one sample / environment
Richness: # of species/taxa observed or estimated
Evenness: relative abundance of each taxon
α-Diversity: taking both evenness and richness into account
Phylogenetic distance (PD)
Shannon entropy
dozens of different measures in QIIME

42. Beta Diversity Analysis
7) Downstream Analysis
Beta Diversity Analysis
Beta-diversity: differences in diversity across samples or environments
UniFrac (Lozupone et al, AEM, 2005) (phylogenetic)
Bray-Curtis dissimilarity measure (OTU abundance)
Jaccard similarity coefficient (OTU presence/absence)
Large number of other measures available in QIIME (and in Mothur)

43. Beta Diversity (UniFrac)
7) Downstream Analysis
Beta Diversity (UniFrac)
all OTUs are mapped to a phylogenetic tree
summation of branch lengths unique to a sample constitutes the UniFrac score between the samples
shared nodes either ignored (unweighted) or discounted (weighted)

44. Different beta-diversity measurements provide different information
Parks and Beiko 2012, ISMEJ

45. Beta Diversity Analysis
7) Downstream Analysis
Beta Diversity Analysis
First, paired-wise distances/similarity of the samples calculated
Then the matrix (high-dimensional data) is transformed into a lower-dimensional display for visualization
Shows you which samples cluster together in 2D or 3D space
PCoA
Hierarchical clustering
Sample 1
Sample 2
Sample 3 1
0.4
0.8
0.7

46. Principal Coordinate Analysis (PCoA)
7) Downstream Analysis
Principal Coordinate Analysis (PCoA)
When there are multiple samples, the UniFrac scores are calculated for each pair of samples. In order to view the multidimensional data, the distances are projected onto 2 (or more) dimensions using PCoA

47. PCoA 7) Downstream Analysis Each dot represents a sample
Dots are colored based on the attributes provided in the mapping (metadata) file
Source: Navas-Molina et al 2015

48. Is it statistically significant?
You might see separation on PCoA plot, but is it statistically significant?
QIIME has several tests within the compare_categories.py script
Anosim
Adonis
Mantel?

49. Visualization and Statistics
Various tools are available to determine statistically significant taxonomic differences across groups of samples
Excel
SigmaPlot
Past
R (many libraries)
Python (matplotlib)
STAMP

50. STAMP

52. STAMP Plots

53. STAMP
Input
“Profile file”: Table of features (samples by OTUs, samples by functions, etc.)
Features can form a heirarchy (e.g. Phylum, Order, Class, etc) to allow data to be collapsed within the program
“Group file”: Contains different metadata for grouping samples
Can be two groups: (e.g. Healthy vs Sick) or multiple groups (e.g. Water depth at 2M, 4M, and 6M)
Output
PCA, heatmap, box, and bar plots
Tables of significantly different features

54. Machine Learning Fits model (Random Forests) to data
Selects features (species)
Classifies samples (control or exercise)
Samples
Categories
Machine learning fits a computer model, in this case one known as random forests, to the data. It selects the most important features, which for our data are species, that distinguish between data categories. It then classifies samples as being from either exercise or control mice based on those features. It trains on one subset of the data, and then it tests the model on another subset, to see how accurately it can predict the category of a sample based on those selected species. Explain table. As a control for the model, we look at whether the prediction accuracy drops if we randomly select the species that it uses for classification

55. Machine learning identifies differences between exercise and controls
PC2 (15.31%)
PC3 (8.34%)
PC1 (28.56%)
Exercise
Control
Y-axis: Prediction accuracy, or the model’s ability to tell which sample is from which group.
X-axis: We ran the model for all time points.
We looked at accuracy using the 30 most important species that were selected for classification, and we also looked at the accuracy using 30 randomly selected species. We found that accuracy was much better with the selected features (%80) than the random ones (less than %60). You can see that the difference in accuracy gets larger at later time points, indicating that there is a shift in the microbiome. So though we did not see a difference in species richness or shared species across samples, machine learning shows us that there is a difference in microbiome relationships in response to exercise.
So we do see a difference in this method compared to other methods
Future:
narrow down important species
Applying model to other datasets

56. Using Machine Learning
Can run Random Forests within QIIME
Script: supervised_learning.py
More options within either:
R using the caret package
Python using the scikit-learn package

57. Putting it all together
Microbiome Helper

58. Microbiome Helper
Microbiome Helper is an open resource aimed at helping researchers process and analyze their microbiome data
Combines bioinformatic methods from different groups and is updated continuously
Scripts to wrap and integrate existing tools
Available as an Ubuntu Virtualbox
Standard Operating Procedures for 16S and metagenomics
Tutorials with example datasets

59. Microbiome Helper Wiki

60. Microbiome Helper Vbox

61. Questions?

62. Tutorial Microbiome Helper