Документ взят из кэша поисковой машины. Адрес оригинального документа : http://mccmb.belozersky.msu.ru/2015/proceedings/abstracts/7.pdf Дата изменения: Mon Jun 15 15:39:58 2015 Дата индексирования: Sat Apr 9 23:19:49 2016 Кодировка:

Improving genome assemblies using multi-platform sequence data Ё Pinar Kavak 1,2, , Bekir Erguner 1 , Bayram Yuksel 3 , Duran Ustek 4 , Mahmut Samil Ё Ё ё 1 2 5, Sa glu , Tunga GungЁ , and Can Alkan giro Ё or
1

Ё Advanced Genomics and Bioinformatics Research Group (IGBAM), BILGEM, TUBITAK, 41470 Gebze, Kocaeli, Turkey, pinar.kavak@tubitak.gov.tr; 2 Department of Computer Engineering, Bo ci University, 34342 Beb ek, Istanbul, Turkey gaziё 3 Advanced Genomics and Bioinformatics Research Group (IGBAM), MAM, TUBITAK, 41470 Gebze, Ko caeli, Turkey Ё 4 Department of Medical Genetics, Istanbul Medip ol University, 34810 Beykoz, Istanbul, Turkey
5

Department of Computer Engineering, Bilkent University, 06800 Bilkent, Ankara, Turkey, calkan@cs.bilkent.edu.tr

Abstract De novo assembly using short reads generated by next generation sequencing technologies is still an open problem. Although there are several assembly algorithms developed for data generated with different sequencing technologies, and some that can make use of hybrid data, the assemblies are still far from being perfect. There is still a need for computational approaches to improve draft assemblies. Here we propose a new method to correct assembly mistakes when there are multiple types of data obtained using different sequencing technologies that have different strengths and biases. We apply our method to Illumina, 454, and Ion Torrent data, and also compare our results with existing hybrid assemblers, Celera and Masurca.

1

Intro duction

Since the introduction of high throughput next generation sequencing (NGS) technologies, traditional Sanger sequencing is being abandoned especially for large-scale sequencing pro jects. Although cost effective for data production, NGS also imposes increased cost for data processing and computational burden. In addition, the data quality is in fact lower, with greater error rates, and short read lengths for most platforms. One of the main algorithmic problems to analyze NGS data is the de novo assembly: i.e. "stitching" billions of short DNA strings into a collection of larger sequencees, ideally the size of chromosomes. However, "perfect" assemblies with no gaps and no errors are still lacking due to many factors, including the short read and fragment (paired-end) lengths, sequencing errors in basepair level, and the complex and repetitive nature of most genomes. Some of these problems in d e novo assembly can be ameliorated through using data generated using different sequencing platforms, where each technology has "strengths" that may be used to fix biases introduced by others.


to whom correspondence should be addressed. pinar.kavak@tubitak.gov.tr, calkan@cs.bilkent.edu.tr

1


In this work, we propose to improve draft assemblies (i.e. produced using a single data source, and/or single algorithm) by incorporating data generated using different NGS technologies, and applying novel correction methods. To achieve better improvements, we exploit the advantages of both short but low-error and long but erroneous reads. We show that correcting the contigs built by assembling long reads through mapping short (and high quality) read contigs produce the best results, compared to the assemblies generated by algorithms that use hybrid data. 2 Metho ds

We first cloned a bacterial artificial chromosome (BAC) from human chromosome 13. We then sequenced this BAC separately using Illumina, Roche/454, and Ion-Torrent platforms. Illumina data is paired-end, where the others are single-end. The read lengths are 101bp for Illumina, 10bp-1Kbp for Rochhe/454, and 5bp-201bp for Ion Torrent. We also obtained a "gold standard" reference assembly using template-based assembly with Mira [7] with Roche/454, which is then corrected with the Illumina reads. Since Roche/454 and Ion Torrent platforms have similar sequencing biases (i.e. problematic homopolymers), we worked on two separate groups: Illumina & 454 and Illumina & Ion-Torrent, which gives us an opportunity to compare Roche/454 and Ion-Torrent. Pre-pro cessing: We first discarded the reads that has low average quality value (phred score 17, i.e. 2% error rate). Next, we removed the reads with high N-density (with >10% of the read consisting of Ns). We then trimmed groups of bases that seem to be nonuniform according to sequence base content. We also inevitably applied each assembler's pre-processing operations. Assembly: We used several assembly tools: Velvet[3], a de Bruijn graph based assembler to assemble the short reads; and two different overlap-layout-consensus (OLC) assemblers: Celera [1], and SGA [2] to assemble the long read data sets (Roche/454 and Ion Torrent) separately. Finally, we also used a de Bruijn based assembler, SPAdes[4] on the long read data. We then mapped all draft assemblies to the E. coli reference sequence to identify and discard E. coli contamination due to the cloning process. At the end, we obtained one short read, and three long read assemblies. Correction: We mapped the contigs obtained with the short reads onto the contigs generated by assembling long reads using BLAST[8]. Since BLAST may report multiple 2


mapping locations due to repeats, we accepted only the "best" map locations. Reasoning from the fact that the short reads show less sequencing errors, we opted for the sequence reported by the short read based contigs over the long read contigs assemblies when there are disagreements between the pair, and patched the "less fragmented" long read assemblies. We repeated this process for each of the three long read assembly data sets. Evaluation: We mapped each of the final corrected assemblies onto the reference genome we constructed, calculated various statistics based on the comparisons, and estimated assembly qualities (Table 1). We also used two hybrid assemblers, Celera-CABOG [5] and Masurca [6] on the same data to compare our correction methodology with those of hybrid assembly algorithms. 3 Results and Conclusion

We present a summary of the results in Table 1. Briefly, the Velvet assembly using only the Illumina reads showed better coverage (99%) and high average identity (97.5%) rates compared to Celera assembly using Celera. Correcting the Celera assembly with our method improves both coverage and average identity rates, which are then further improved by reiteratively applying our method. We also observe that the hybrid assemblers used in this study did not produce better results with this data set. Here we presented a new method to improve draft assemblies by correcting high contiguity assemblies using high quality short read contigs. However, the need to develop new methods that exploit different data properties of different NGS technologies remains. Funding The pro ject is supported by the Republic of Turkey Ministry of Development Infrastructure
Ё Grant (no: 2011K120020), BILGEM - TUBITAK (The Scientific and Technological Research Council of Ё Turkey) grant (no: T439000), and a TUBITAK grant to C.A.(112E135).

References
[1] E.W.Myers et al (2000) A Whole-Genome Assembly of Drosophila, Science, 287:2196-2204. [2] J.Simpson et al (2012) Efficient de novo assembly of large genomes using compressed data structures, Genome Research, 22:549-556. [3] D.Zerbino, E.Birney (2000) Velvet: Algorithms for de novo short read assembly using de Bruijn graphs, Genome Research, 18(5):821-829. [4] A.Bankevich et al (2012) SPAdes: A New Genome Assembly Algorithm and Its Applications to Single-Cell Sequencing, Journal of Computational Biology, 19(5):455-477. [5] J.R.Miller et al (2008) Aggressive assembly of pyrosequencing reads with mates, Bioinformatics, 24(24):2818-2824. [6] A.Zimin et al (2013) The MaSuRCA genome Assembler, Bioinformatics, 29(21):2669-2677. [7] B.Chevreux et al (1999) Genome sequence assembly using trace signals and additional sequence information, Computer Science and Biology:Proceedings of the German Conference on Bioinformatics (GCB), 99:45-56. [8] S.Altschul et al (1990) Basic lo cal alignment search to ol, Journal of Molecular Biology, 215(3):403-410.

3


Table 1: Results of assembly correction method on BAC data.
Name Length # of Contigs # of Mapp ed Contigs # of Covered bases Coverage Avg. Identity # of Gaps Size of Gaps

Reference Velvet Ill. Velvet Celera 454 Celera Ion Celera Corrected Celera Ill-454 Celera Ill-454 Celera2 * Ill-454 Celera3 Ill-Ion Celera Ill-Ion Celera2 Ill-Ion Celera2 SGA 454 SGA Ion SGA Corrected SGA Ill-454 SGA Ill-454 SGA2 Ill-Ion SGA Ill-Ion SGA2 Ill-Ion SGA2 SPADES 454 SPADES Ion SPADES Corrected SPADES Ill-454 SPADES Ill-454 SPADES2 Ill-Ion SPADES Ill-Ion SPADES2 Masurca Ill-454 Masurca Ill-Ion Masurca Celera-CABOG Ill-454 Celera Ill-Ion Celera

176.843

197,040

455

437

175,172

0.99055

0.97523

39

1,671

908,008 39,347

735 27

735 27

172,563 47,638

0.97580 0.26938

0.92599 0.96932

18 47

4,280 129,205

4,945,785 5,078,059 5,086,627 93,909 145,262 216,167

895 890 890 30 30 30

270 265 265 28 28 28

176,368 176,640 176,640 81,819 91,962 99,645

0.99731 0.998852 0.998852 0.46267 0.52002 0.56347

0.94370 0.944527 0.944560 0.96327 0.97412 0.98066

5 4 4 36 33 34

475 203 203 95,024 84,881 77,198

62,909,254 842,997

108,095 6,417

101,514 6,122

176,546 153,092

0.99832 0.86569

0.97439 0.99124

1 197

297 23.751

295,009 279,034 197,509 203,064 204,524

335 305 291 291 291

335 305 291 291 291

176,757 176,757 175,052 175,676 175,677

0.99951 0.99951 0.98987 0.99340 0.99341

0.96823 0.96769 0.97501 0.97413 0.97405

5 5 45 34 34

86 86 1,791 1,167 1,166

12,307,761 176,561

49,824 110

49,691 107

176,843 167,890

1.0 0.94937

0.98053 0.92909

0 9

0 8,953

290,702 290,917 198,665 200,307

298 297 52 52

298 297 52 52

176,454 176,454 171,977 172,101

0.99780 0.99780 0.97248 0.97319

0.96538 0.96530 0.94215 0.94230

5 5 4 2

389 389 4,866 4,742

380 2,640

1 8

0 8

0 1,952

0 0.01104

0 0.98223

0 9

0 174,891

1,101,716 0

891 0

891 0

174,330 0

0.98579 0.0

0.92452 0.0

12 0

2,513 0.0

Name: the name of the data group that constitute the assembly; # of contigs: the numb er of contigs that b elong to the resulting assembly; # of Mapp ed Contigs: the numb er of contigs that successfully mapp ed onto the reference sequence; # of Covered bases: the numb er of bases on the reference sequence that are covered by the assembly; Coverage: p ercentage of covered reference; Avg. identity: p ercentage of the correctly predicted reference bases; # of Gaps: The numb er of gaps that cannot b e covered on the reference genome; Size of Gaps: total numb er of bases on the gaps.

*

"2" represents the results of the second cycle of correction, "3" represents the third cycle.

4