1. billion-piece genome puzzle
a new standard for
assembling a
billion-piece genome puzzle
ALLPATHS-LG

2. Sante Gnerre et al. (20 Authors)
CS 681
High-quality draft assemblies of mammalian genomes from massively parallel sequence data
ALLPATHS-LG by
Sante Gnerre et al. (20 Authors)
Jan 25th, 2011
presented by
Ömer Köksal

3. Agenda Introduction Results Discussion Model for Input Data
Sequencing Data
ALLPATHS-LG Assembly Method
Uncertainty in Assemblies
Human and Mouse Assemblies
Human Genome
Mouse Genome
Segmental Duplications
Understanding Gaps
Discussion

4. Introduction High-quality assembly of a genome sequence is critical
Particularly challenging for large, repeat rich genomes such as those of mammals
Using traditional capillary-based sequencing (>700 bases) such assemblies produced for multiple mammals at a cost of tens of millions dollars each.
New massively parallel technologies are expected to lower cost dramatically but they could not, because of
short sequencing (~100 bases in length)
less accuracy
difficult to assemble

5. Introduction (cont’d)
ALLPATHS-LG
de novo assembly of large (and small) genomes
it should be possible to generate high quality draft assemlies of Large Genomes
~1000 fold lower cost than a decade ago
Previous versions:
ALLPATHS 1.0 (2008)
ALLPATHS 2.0 (2009)

6. Results RESULTS Model for Input Data Sequencing Data
ALLPATHS-LG Assembly Method
Uncertainty in Assemblies
Human and Mouse Assemblies
Human Genome
Mouse Genome
Segmental Duplication
Understanding Gaps

7. Results - Model for Input Data
De novo genome assembly depends on
computational methods
nature and quantity of sequence data used
Fairly standard model of Capillary-based sequence was modified
Sets a target of 100 fold sequence coverage to compensate shorter reads & nonuniform coverage
Despite using higher coverage, proposed model is dramatically cheaper since the per-base cost ~10000 fold lower than capillary sequencing
illumina sequencing was used (Table-1)

8. Results - Model for Input Data (cont’d)
Table 1 – Provisional sequencing model for de novo assembly

9. Results – Sequencing Data
Using the model above generated sequences are:
Human Genome
Mouse Genome
Human Genome:
GM12878 (Coriell Institute) of 1000 Genomes Pilot Project
NCBI Short Read: Human_NA_12878_Genome_on_illumina
Mouse Genome:
C57BL/6J female DNA
NCBI Short Read: Mouse_B6_Genome_on_illumina

10. Results - ALLPATHS-LG Assembly Method
previous versions were improved extensively
can assembly small genomes
freely available at:

11. Results - ALLPATHS-LG Assembly Method (cont’d)
Some key innovations in ALLPATH-LG
Handling repetitive sequences
more resilient to repeats
Error Correction
for every 24-mer the algorithm examines the stack of all reads containing 24-mer
incidence of incorrect error correction was reduced
Use of jumping data
it coult work even with such data: it trim bases beyond junction points and treats each read pair as belonging to one of two possible distributions
Efficient memory usage
can asseble human genomes on commercial servers (48 processors & 512 GB Ram) in a few week
3 week for mouse & 3.5 weeks for human)

12. Results – Uncertainty in Assemblies
The goal of assembly is to reconstruct the genome as accurately as possible
However in some locations the data may be compatible with more than one solutions
Rather than making an arbitrary choice (& introducing errors) algortihm retains alternatives
ALLPATHS-LG algorithm generates an assembly graph whose edges are sequences and braches represent alternate choices
ATC{A,T}GGTTTTTTT{T,TT}ACAC
Variant Call Format (.VCF file)

13. Results – Uncertainty in Assemblies (cont’d)
NOTE:
Current version of ALLPATHS-LG only captures single-base and simple sequence indel uncertainties
Better way to capture alternatives are needed (many of which are still lost in the current version and giving rise to errors)
It would be desirable to assign probabilities to each alternative

14. Results – Human & Mouse Assemblies
Resulting genome assemblies provide good coverage of the human and mouse genomes
ALLPATHS-LG assemblies were compared with previously published assemblies
Capillary-based sequencing
SOAP (massively sequencing parallel sequencing)

15. Results – Human & Mouse Assemblies (cont’d)

16. Results – Human Genome N50 contig length of 24 kb
Scaffold length of 11.5 Mb
Contiguity is > 4 fold longer than SOAP algorithm
Connectivity is > 25 fold longer than SOAP algorithm
Assembled sequence contains 91.1% of the reference genome (SOAP: 74.3%)
Assembled sequence contains 95.1% of the reference genome (SOAP: 81.2%)
Results are similar to capillary based assemblies

17. Results – Human Genome (cont’d)
Local assembly error: 3.5 %
Capillary: 4.1 %
SOAP: 6.2 %
Long range accuracy: 99.1%
Capillary: 99.7 %
SOAP: %

18. Results – Mouse Genome
Results are broadly similar for the mouse genome
N50 contig length of 16 kb
Scaffold length of 7.2 Mb
Connectivity is > 20 fold larger than SOAP algorithm
Approach Capillary results (contig: 25 kb, scaffold: 16.9 Mb)
Assembled sequence contains 88.7% of the reference genome (Capillary: 94.2%)
Assembled sequence contains 96.7% of the reference genome (Capillary: 97.3%)
Results are considerably better than SOAP

19. Results – Mouse Genome (cont’d)
Local assembly error: 3.0 %
Capillary: 2.7 %
SOAP: 14.2 %
Long range accuracy: 99.0 %
Capillary: 99.1 %
SOAP: %

20. Results – Segmental Duplications
Segmental duplications shows a challange
ALLPATHS-LG assemlies (human and mouse) cover only ~40% segemental duplications
Capillary: 60%
SOAP: 12%
NOTE:
Clearly additional work is needed here

21. Results – Understanding Gaps
Rougly three quarters of the gaps captured
Remaining gaps are not spanned
Majority of the sequence within the gaps consists of repetitive elements, 61.9% of gaps:
For mouse: LINE elements are major contributors to GAPS
For human: LINE & SINE elements

22. Discussion
High quality vertebrate genomes provided an essential foundation for comperative analysis of human genome
Costing tens of millions $ each to generate with capillary based sequencing
In this work, ALLPATHS-LG was presented lowering the cost ~1000 fold.

23. Discussion (cont’d) ALLPATHS-LG Good long range connectity,
Good accuracy,
Good coverage
wrt capillary based sequencing and
better than SOAP
Quality of the assembliesis considerably better:
scaffolds are > 25 times longer

24. Discussion (cont’d)
ALLPATHS-LG is anticipated to yield even better results in the improved version.
ALLPATHS-LG introduced a preliminary syntax for expressing alternatives: TTTT{T, TT}
Computational hardware requirements:
SOAP is faster (takes 3 days) but accuracy is low
ALLPATHS-LG is slower but produces high quality assemblies
ALLPATHS-LG is anticipated to be speeded up with algorithmic improvements (considering in mind the trade-off between speed and the accuracy)

25. Thank you.
Questions ?