1. Informatics challenges for next-generation sequence analysis
Gabor T. Marth
Boston College Biology Department
PSB 2008
January

2. Read length and throughput
Illumina/Solexa, AB/SOLiD short-read sequencers
1Gb
(1-4 Gb in bp reads)
bases per machine run
100 Mb
454 pyrosequencer
( Mb in bp reads)
10 Mb
ABI capillary sequencer
1Mb
read length
10 bp
100 bp
1,000 bp

3. Current and future application areas
Genome re-sequencing: somatic mutation detection, organismal SNP discovery, mutational profiling, structural variation discovery
reference genome
SNP
DEL
De novo genome sequencing
Short-read sequencing will be (at least) an alternative to micro-arrays for:
DNA-protein interaction analysis (CHiP-Seq)
novel transcript discovery
quantification of gene expression
epigenetic analysis (methylation profiling)

4. Fundamental informatics challenges (I)
1. Interpreting machine readouts – base calling, base error estimation
2. Dealing with non-uniqueness in the genome: resequenceability
3. Alignment of billions of reads

5. Informatics challenges (II)
4. SNP and short INDEL, and structural variation discovery
5. Data visualization
6. Data storage & management

6. Challenge 1. Base accuracy and base calling
machine read-outs are quite diverse
read length, read accuracy, and sequencing error profiles are variable (and change rapidly as machine hardware, chemistry, optics, and noise filtering improves)

7. Roche/454 pyrosequencer insertion and deletions dominate
error rates are nucleotide-dependent
base quality values are underestimated

8. Illumina/Solexa system
Error rate grows with cycle number
Actual base accuracy for a fixed base quality value is a function of base position within the read

9. AB/SOLiD system 2-base, 4-color: 16 probe combinations
dibase encoding: base transitions rather than individual bases are read
conversion between base-space and “color-space”
base quality value assignment is tricky A C G T
2nd Base
1st Base 1 2 3
2-base, 4-color: 16 probe combinations

10. PYROBAYES: 454 base calling program
data likelihoods
priors
posterior base number probability 10

11. Challenge 2. Resequenceability
Reads from repeats cannot be uniquely mapped back to their true region of origin
RepeatMasker does not capture all micro-repeats, i.e. repeats at the scale of the read length
Near-perfect micro-repeats can be also a problem because we want to align reads even with a few sequencing errors and / or SNPs

12. Finding perfect and near-perfect micro-repeats
Hash based methods (fast but only work out to a couple of mismatches)
Exact methods (very slow but find every repeat copy)
Heuristic methods (fast but miss a fraction of the repeats)

13. Challenge 3. Read alignment and assembly
resequencing requires reference sequence-guided read alignment
Step 1. initial short-hash based scan for possible read locations
Step 2. evaluation of candidate read locations with SW method 13

14. Alignment to reference in “color-space”
There is no need to ever leave color space as you can convert your reference sequence to its color space and then align the two sequences. This is in fact desirable given the error properties of color space, our analysis software will use color space. note the misalignment in here are explained in next two slides)
Working in color space:
Reverse-complementation becomes simply reverse
Apply color transition rules to remove measurement errors from partial assemblies
If reference of Sanger reads are combined, translate to color space 14

15. Challenge 4. Polymorphism discovery
shallow and deep read coverage
most candidates will never be “checked”  only very low error rates are acceptable
we updated PolyBayes to deal with new read types
made the new software (PBSHORT) much more efficient

16. SNP calling in short-read coverage
C. elegans reference genome (Bristol, N2 strain)
Pasadena, CB4858
(1 ½ machine runs)
SNP calling error rate very low:
Validation rate = 97.8% (224/229)
Conversion rate = 92.6% (224/242)
Missed SNP rate = 3.75% (26/693)
SNP
INS
INDEL candidates validate and convert at similar rates to SNPs:
Validation rate = 89.3% (193/216)
Conversion rate = 87.3% (193/221) 16

17. Mutational profiling: deep 454/Illumina/SOLiD data
Pichia stipitis reference sequence
Image from JGI web site
collaboration with Doug Smith at Agencourt
Pichia stipitis converts xylose to ethanol (bio-fuel production)
one mutagenized strain had especially high conversion efficiency
determine where the mutations were that caused this phenotype
we resequenced the 15MB genome with 454 Illumina, and SOLiD reads
14 true point mutations in the entire genome
In about 15X nominal coverage each technology can find every point mutation with essentially no false positives

18. Structural variation discovery
copy number variations (deletions & amplifications) can be detected from variations in the depth of read coverage
structural rearrangements (inversions and translocations) require paired-end read data
Ask Chip to provide images for this one slide

19. Challenge 5. Data visualization
aid software development: integration of trace data viewing, fast navigation, zooming/panning
facilitate data validation (e.g. SNP validation): co-viewing of multiple read types, quality value displays
promote hypothesis generation: integration of annotation tracks

20. Challenge 6. Massive data volumes
two connected working groups to define standard binary data formats
Short-read format working group
(Asim Siddiqui, UBC)
Assembly format working group
Boston College

21. Our software is available for testing

22. Credits http://bioinformatics.bc.edu/marthlab
Elaine Mardis (Washington University)
Andy Clark (Cornell University)
Doug Smith (Agencourt)
Research supported by: NHGRI (G.T.M.)
BC Presidential Scholarship (A.R.Q.)
Michael Stromberg
Chip Stewart
Michele Busby
Aaron Quinlan
Damien
Croteau-Chonka
Eric Tsung
Derek Barnett
Weichun Huang