1. The Next, Next Generation of Sequencing -
From Semiconductor to Single Molecule
Justin H. Johnson
Director of Bioinformatics
EdgeBio
Washington DC, USA

2. Agenda
• Who We Are
• NGS at 30K
• The Challenges
– Even Before We Get to the Platforms
– When We Get to the Platforms

3. Who We Are

4. Life Tech
Service
Provider

5. Contract Research Division
• Five SOLiD4 sequencing platforms
• One Life Techologies 5500XL
• Two Ion Torrent PGMs
• Bioinformatics consulting on Illumina, 454, and PacBio
• Automation thru Caliper Sciclone & Biomek FX
• Commercial partnerships with companies such as CLCBio,
DNANexus and Genologics
• MD/PhD & Masters Level Scientists and Bioinformaticians
• IT Infrastructure of >100 CPUs and >100TB storage

6. Edge BioServ
Scientific Advisory Board
Elaine Mardis, Ph.D. Steven Salzberg, Ph.D.
Co-Director, Genome Sequencing Center Director, Center for Bioinformatics and
Washington University School of Medicine Computational Biology
University of Maryland
Sam Levy, Ph.D.
Director of Genome Sciences Gabor Marth, Ph.D.
Scripps Translational Science Institute Professor of Bioinformatics
Scripps Genomic Medicine Boston College
Michael Zody, M.S.
Chief Technologist Elliott Margulies, Ph.D.
Broad Institute Investigator
Genome Informatics Section
Ken Dewar, Ph.D. National Human Genome Research Institute
Assistant Professor National Institutes of Health
McGill University and Genome Quebec

7. NGS @ 30K Feet

8. Machines and Vendors

9. Obligatory NGS Exponential Growth Slide
Nature Biotechnology Volume 26 Number10 October2008

10. Ultra High Throughput + Lower Cost = Broader Applications
RNA-Seq/
Whole Transcriptome Epigenome
- mRNA Expression & Discovery - Transcriptionally Active Sites
- Alternative Splicing - Protein-DNA Interactions
- Allele-Specific Expression - Methylation Analysis
- microRNA Expression &
Discovery
Genome
- De Novo
- Resequencing/ Mutation Metagenome
Discovery & Profiling - Microbial Diversity
- Exome Sequencing - Heterogeneous Samples
- Copy Number Variation
- Ancient DNA

11. Challenges

12. Challenges
Technical Expertise

13. Experimental Design Considerations
 Sequencing Platform in Use
 Choice of Library Construction
 Depth of coverage
 Re$ources
 Number of Replicates
 Number of Samples and Control
 Etc…

14. Challenges
Flexibility w/ Standards

15. Flexibility with Standards and Scale
• Then (CE) – The Norm
– 10 Machines, 30 – 360 Days, 1 Project
• Now (Illumina/SOLiD/454) – Scale
– 1 machine, 14 Days, 30 Projects
• Now (Ion Torrent) - Flexibility
– 1 machine, 1 Day, 1 Project.
• Standardization of analysis (Details Later)

16. Challenges
Sample Preparation

17. Sample Sourcing for RNA Projects
– Blood: Large quantities of sample available, but
with limited utility in transcriptome analysis
– Tissue: Needle biopsy most common, but sample
quantity very low
– Surgical section: Larger quantities available, but
limited utility; need laser capture microdissection
to provide useful results, sample quantity very low
– FFPE Slides: Very useful in clinical research but
amount of sample and quality low.

18. Unamplified vs Amplified
• Prostate Cancer Cell Line (Vcap) from CPDR
– Well characterized
– Differential Expression upon the addition of
androgens.
– Compared transcriptome from a single pool of
RNA
• Unamplified, ribosomally depleted (Ribominus™)
• Amplified, no ribosomal depletion required
• Two Pipelines for analysis

19. Amplification Gives Different Results
• Gene Expression in Unstimulated Cells
14,075

20. Spearman’s Correlation from 2
Pipelines
Pipeline A Unamplified Amplified
Androgen + - + -
+ … 0.930 0.904 0.892
Unamplified
- … … 0.896 0.900
+ … … … 0.928
Amplified
- … … … …
Pipeline B Unamplified Amplified
Androgen + - + -
+ … 0.853 0.757 0.701
Unamplified
- … … 0.720 0.712
+ … … … 0.848
Amplified
- … … … …

21. Challenges
Sample Analysis

22. Exome Seq Ultimately About Variants
• Coverage
• Project Design
– Cohorts
– Cancer
• Algorithms a Solved Problem?
– Single open source pipelines
– Single commercial pipelines
– Proprietary internal algorithms.
– A mixture?

23. Ultimately Comes to Variation
• Coverage
• Project Design
– Cohorts
– Cancer
• Algorithms Solved Problem?
– Single open source pipelines
– Single commercial pipelines
– Proprietary internal algorithms.
– A mixture?

24. Digging Deep with an Exome
Genetic variation in an individual human exome.
Ng PC, Levy S, Huang J, Stockwell TB, Walenz BP, Li K, Axelrod N, Busam DA, Strausberg RL, Venter JC.
PLoS Genet. 2008 Aug 15;4(8):e1000160.

25. Venter Genome - Algorithms
• PLOS genetics 2008 vol 4 issue 8 e10000160
• ~21K SNP in exons (29MB Targeted)
• 36,206 expected SNPs for 50MB Kit
% Difference Homozygous TP TN FP FN Sensitivity Pos.pred.val
B 1% 0% -39% -1% 1% 4%
A 31% 0% 88% -41% 31% -6%
C -32% 0% -49% 42% -32% 2%
% Difference Heterozygous TP TN FP FN Sensitivity Pos.pred.val
B 0% 0% 16% 0% 0% -9%
A -15% 0% -44% 21% -15% 16%
C 15% 0% 28% -20% 15% -7%

26. 3 Tools and Associated SNP Counts
• Software A
– 45,551
• Software B
– 29,814
• Software C
– 40,964

27. Software B v. Software A
B A
29,814 45,511
8,564 21,250 24,261
Union: 54,075
Intersection: 21,250
Not to Scale

28. Software B v. Software C
B C
29,814 40,964
6,358 23,456 17,508
Union: 47,322
Intersection 23,456

29. Software A v. Software C
A C
45,511 40,964
14,738 30,773 10,191
Union: 55,702
Intersection: 30,773

30. B A
29,814 45,511
4,750 1,608 13,130
19,642
3,814 11,131
6,377
Union: 60,452
Intersection: 19,642
Voting Scheme (2/3): 36,195
C
40,964

31. Challenges
Platforms

32. The weight in…
Yield/Day Read Error Rates Read Lengths
Illumina MiSeq 2.0 Gb 1.3% (V4) 150
Ion Torrent PGM 0.5 Gb (316 Chip) 1.5% (316 Chip) 120 - 240
PacBio RS 3.0 Gb 2-15% 430-2900
• Illumina and PacBio numbers from Vendor Sequencing
• Ion Torrent from EdgeBio Sequencing

33. Illumina MiSeq
Mid-Range Length, Accurate Reads, Large Throughput
• All Resequencing
• All De novo Applications
• Transcriptome
• Methylation

34. Ion Torrent PGM
Long, Mostly Accurate Reads in 2.5 Hours
• Microbial & Viral Resequencing
• Microbial & Viral De novo Applications
• Eukaryotic Amplicon Sequencing
• Metagenomics
– WGS
– 16S Surveys

35. Pac Bio RS
Ultra Long, Less Accurate Reads & Rapid Sequence
• Microbial & Viral De novo Applications
• Structural Variation / Haplotyping

36. Ion Torrent PGM
Mean Read Total # A20 Mean Read
Name Total # Reads Length Longest Read (Mbp) Q20 Mb Length
HG19-01 2,660,176 139 203 369.91 124.00 74
HG19-02 2,321,405 121 202 281.43 116.43 75
HG19-03 2,471,922 134 203 331.54 124.17 77
Microbe (37% GC) 2,869,789 122 202 350.23 160.48 82
Microbe (30% GC) 2,866,851 122 202 350.16 141.31 81

37. Ion Torrent PGM
Percent of
# Aligned / % Aligned / Aligned
Total # # N50 Largest Consensus
Name Assembled Assembled Genome
Reads Contigs Contig Contig Accuracy
Reads Reads Covered
(AQ40)
DH10B Mapping 1,384,863 1,334,138 96.34% 90 107,749 326,368 99.51% 99.97%
DH10B Denovo 1,384,863 1,335,604 96.44% 216 42,499 146,899 99.53% 1.73%
On Similar Illumina Data Set
• Normalizing for coverage and removing Paired Ends
• N50 of 94926 and Largest Contig of 236274
• Removing normalization improved numbers

38. Why the Difference?
Quality?

39. Quality?
Q-Q plots of the DH10B Ion Torrent 316 chip data expected vs empirical quality
before recalibration (left) and after recalibration (right).

40. Quality
Q-Q plots of DH10B MiSeq data expected vs empirical quality before recalibration
(left) and after recalibration (right).

41. Empirical Quality

42. Empirical Quality - Long Reads

43. Then Why?
• De Bruijn Graphs adversely affected by more
frequent INDEL characteristics of Ion Torrent
• Higher Average Quality reads are less
abundant in Ion Torrent

44. Does this matter in Resequencing?
• Depends on the tools used!
– If you understand error profile, you can correct for it…
• Ran Simulated DH10B mutation experiment
1. make mutated e. coli reference (fakemut)
2. align data to mutated reference (clc, tmap, or other mappers)
3. calculate per base coverage on the BAM file
(genomeCoverageBed)
4. run samtools/mpileup/vcffilter (or CLC SNP/INDELcalling) to
call variants -run various settings to compare variant calling
5. Calculate false positives, true positives, and false negatives
6. Calculate number of variants missed due to low coverage
7. Calculate PPV and corrected sensitivity
8. Graph PPV and corrected sensitivity

45. Resequencing
• Ion claims substitution issues with MiSeq 1
• Illumina claims INDEL issues with Ion 2
1. http://www.iontorrent.com/lib/images/PDFs/co23743_pgm_app_note.pdf
2. http://www.illumina.com/Documents/products/appnotes/appnote_miseq_ecoli.pdf

46. Resequencing
Variants Specificity Sensitivity PPV
Identified
Ion/TMAP/SamTools 460 100% 76.957% 97.676%
(Mod)
Ion/TMAP/SamTools 459 99.895% 91.939% 6.014% (~6500
(Default) False Negatives)
MiSeq/Eland/SamTools 220 99.99996% 95.91% 99.06%
(Default – SNPs ONLY)
MiSeq/CLC/SamTools 459 95.464% 99.998% 83.871 (~65 False
(Default) Negatives)
MiSeq SubSampled on DH10B Ion SubSampled on DH10B
(TMAP/Samtools): (TMAP/Samtools):
9 total variants identified 16 total variants identified
8 SNPs and 1 INDEL 0 SNPs and 16 INDEL

47. Ion Data
PPV and Sensitivity of Samtools Analyses
100.000%
80.000%
60.000% Total PPV
SNPs PPV
INDELs PPV
Total Corrected Sensitivity
40.000%
SNPs Corrected Sensitivity
INDELs Corrected Sensitivity
20.000%
0.000%
Default Q4, h100, o20, Q14, h75, o20, Q7, h50, o10, Q14, h50, o10, Variant Calling Q14, h50, o10,
Samtools e27, m1, H1 e21, m4, H2 e17, m4, H1 e17, m4, H1 e17, m4, H2

48. MiSeq Data
PPV and Sensitivity of Samtools Analyses of MiSeq Data
100.000%
80.000%
60.000% Total PPV
SNPs PPV
INDELs PPV
Total Corrected Sensitivity
40.000%
SNPs Corrected Sensitivity
INDELs Corrected Sensitivity
20.000%
0.000%
MiSeq CLC with Default Samtools MiSeq CLC Map with Variant Analysis MiSeq TMAP Map with Variant
Analysis

49. Resequencing Conclusion
Using appropriate aligners and
variant callers we show both
platforms have high accuracy,
each with strengths and
weaknesses…

50. What About PacBio?
• We have less experience with PacBio
• We (EdgeBio) thinks PacBio may have a niche,
but given large initial investment, waiting.
• Many conferences and posters – only results
seen are for de novo sequencing and finishing
(Broad).
• Will be here all week and would love to hear
why you love it.

51. Take This Home
• There are many challenges before we even get
to picking a platform
– Technical Expertise
– Standards in Prep and Analysis
With Great NGS Power Comes Great Responsibility

52. Acknowledgements
• CPDR (Center for Prostate Disease Research) Collaboration
– Shyh-Han Tan, Ph.D.
EdgeBio Sequencing EdgeBio IFX
Joy Adigun John Seed
Elyse Nagle Anjana Varadarajan
Jennifer Sheffield David Jenkins
Rossio Kersey Phil Dagosto
Ryan Mease Quang Tri Nguyen

53. Questions
Twitter: @Bioinfo
jjohnson@edgebio.com