Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
SlideShare a Scribd company logo

2015 vancouver-vanbug

•Download as PPTX, PDF•
•
C
c.titus.brown

This document discusses new directions for the khmer bioinformatics platform, including developing semi-streaming algorithms for sequence analysis using k-mers. Digital normalization is presented as an initial approach that compresses sequencing data, though it discards information. Later work introduced a two-pass semi-streaming framework using saturation detection to enable error correction and variant calling using minimal memory. Current work includes developing a pair-HMM-based graph aligner and applying it to tasks like variant calling. The khmer platform provides implementations of these streaming algorithms to enable analysis of large genomic and metagenomic datasets.

Related slideshows

Usability and Bioinformatics: experience and research challenges
Usability and Bioinformatics: experience and research challengesUsability and Bioinformatics: experience and research challenges
Usability and Bioinformatics: experience and research challenges
 
Integrative_omics_lecture_feb112016_UAB
Integrative_omics_lecture_feb112016_UABIntegrative_omics_lecture_feb112016_UAB
Integrative_omics_lecture_feb112016_UAB
 
BPIPE: a bioinformatics pipeline framework
BPIPE: a bioinformatics pipeline frameworkBPIPE: a bioinformatics pipeline framework
BPIPE: a bioinformatics pipeline framework
 
1 of 51
Building a platform for bioinformatics: some exciting new directions for khmer. C. Titus Brown ctbrown@ucdavis.edu March 12, 2015
Hello! Associate Professor (#tenure!); School of Veterinary Medicine University of California, Davis. More information at: • ged.msu.edu/ ( URL needs to be updated :) • github.com/ged-lab/ • ivory.idyll.org/blog/ • @ctitusbrown
Warnings This talk contains information that may constitute “forward-looking statements.” Generally, the words “believe,” “expect,” “intend,” “estimate,” “anticipate,” “project,” “will” and similar expressions identify forward-looking statements, which generally are not historical in nature. I have been advised to put this disclaimer in as well: Dr. Brown is not currently under treatment for any disorders related to megalomania.
Introducing k-mers CCGATTGCACTGGACCGA (<- read) CCGATTGCAC CGATTGCACT GATTGCACTG ATTGCACTGG TTGCACTGGA TGCACTGGAC GCACTGGACC ACTGGACCGA
De Bruijn graphs – assemble on overlaps J.R. Miller et al. / Genomics (2010)
K-mers give you an implicit alignment CCGATTGCACTGGACCGATGCACGGTACCGTATAGCC CATGGACCGATTGCACTGGACCGATGCACGGTACCG
K-mers give you an implicit alignment CCGATTGCACTGGACCGATGCACGGTACCGTATAGCC CATGGACCGATTGCACTGGACCGATGCACGGTACCG CATGGACCGATTGCACTGGACCGATGCACGGACCG (with no accounting for mismatches or indels)
The problem with k-mers CCGATTGCACTGGACCGATGCACGGTACCGTATAGCC CATGGACCGATTGCACTCGACCGATGCACGGTACCG Each sequencing error results in k novel k-mers!
The opportunity: CCGATTGCACTGGACCGATGCACGGTACCGTATAGCC CATGGACCGATTGCACTCGACCGATGCACGGTACCG The graph contains information about errors (can be used for error trimming in reads). The graph also contains information about variants (can be used for variant calling).
Conway T C , Bromage A J Bioinformatics 2011;27:479-486 © The Author 2011. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com One big challenge: scalability! De Bruijn graph size scales with # errors.
One big challenge: scalability! De Bruijn graph size scales with # errors. Memory usage ~ “real” variation + number of errors Number of errors ~ size of data set
Conway T C , Bromage A J Bioinformatics 2011;27:479-486 © The Author 2011. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com One big challenge: scalability! De Bruijn graph size scales with # errors.
Goals • Initial goal: can we assemble large data sets?? • Longer-term goal: can we find efficient (De Bruijn?) graph-based approaches to sequence analysis?
First attempt: compressible De Bruijn graphs 1% 5% 15%10% Pell et al., 2012 Can use Bloom filters to store De Bruijn graph structures. => Overall structure remains as you squish graphs down.
Technical challenges met (and defeated) • Exhaustive in-memory traversal of graphs containing 5-15 billion nodes. • Sequencing technology introduces false connections in graph. • Implementation lets us scale ~20x over other approaches. Pell et al., 2012
Technical challenges met (and defeated) • Exhaustive in-memory traversal of graphs containing 5-15 billion nodes. • Sequencing technology introduces false connections in graph. • Implementation lets us scale ~20x over other approaches, but this is not enough. • Although, see Minia assembler (Chikhi et al.) Pell et al., 2012
Second attempt: diginorm Conway T C , Bromage A J Bioinformatics 2011;27:479-486 © The Author 2011. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com
Random sampling => deep sampling needed Typically 10-100x needed for robust recovery (30-300 Gbp for human)
Actual coverage varies widely from the average. Low coverage introduces unavoidable breaks.
But! Shotgun sequencing is very redundant! Lots of the high coverage simply isn’t needed. (unnecessary data)
Digital normalization
Digital normalization
Digital normalization
Digital normalization
Digital normalization
Digital normalization
Contig assembly now scales with underlying genome size • Transcriptomes, microbial genomes incl MDA, and most metagenomes can be assembled in under 50 GB of RAM, with identical or improved results. • Memory efficient is improved by use of CountMin Sketch. Brown et al., 2012, arXiv.
Diginorm is simple:
Diginorm is only a good start: • Diginorm alters the coverage of the data set. • Diginorm also discards lots of data! • Various other infelicities… o Repeats go away! o Coverage estimation approach ~poor.
Diginorm is a good start: • Diginorm works on genomes, metagenomes, and transcriptomes; • Diginorm is streaming and uses sublinear space.
Third attempt: a semi-streaming framework for sequence analysis https://github.com/ged-lab/2014-streaming/
Diginorm can detect graph saturation Zhang et al., submitted.
This generically permits semi- streaming approaches. Zhang et al., submitted.
e.g. E. coli analysis => ~1.2 pass, sublinear memory Zhang et al., submitted.
=> Efficient k-mer error trimming. Zhang et al., submitted. (This all works on metagenomes & transcriptomes, too.)
Moving some sequence analysis to streaming. ~1.2 pass, sublinear memory Zhang et al., submitted. First pass: digital normalization - reduced set of k-mers. Second pass: spectral analysis of data with reduced k-mer set. First pass: collection of low-abundance reads + analysis of saturated reads. Second pass: analysis of collected low-abundance reads. First pass: collection of low-abundance reads + analysis of saturated reads. (a) (b) (c) two-pass; reduced memory few-pass; reduced memory online; streaming.
Sublinear time/space read error analysis -- Zhang et al., submitted. Read error profile from mouse mRNAseq (c.f. Grabherr et al., 2011).
Another simple algorithm. Zhang et al., submitted.
So, that’s pretty cool, right? • We provide simple time- and memory-efficient approaches for k-mer spectral analysis of large data sets. • These semi-streaming approaches provide a general framework for applying k-mer spectral approaches to all (deep) sequencing data, including genomes, metagenomes, and RNAseq. • The khmer software provides a functional and reasonably efficient reference implementation, freely available under the BSD license and actively developed at github.com/ged-lab/.
Stream all the things! (1/2)
Stream all the things! (2/2)
But that’s not all! Buy now, and you can also get sequence-to-graph alignment for the low, low price of free!* graph = khmer.new_counting_hash(…) aligner = khmer.ReadAligner(graph, trusted=5) score, graph_align, read_align, is_truncated = aligner.align(seq) * Terms and conditions may apply. Not all source code fully works :)
Pair-HMM-based graph alignment Jordan Fish and Michael Crusoe
(Full model) Jordan Fish and Michael Crusoe
This is a general API… Many potential uses: • Error correction; • Variant calling; • Counting (to replace mapping) & allelic counts; • Align to multiple references; • Tackle strain variation and polyploidy; • Building consensus graphs from shallow population sequencing; • Consensus graph building from multiple read types; • Protein-guided graph search (BlastGraph & Xander)
2015 vancouver-vanbug
Whole-genome variant calling
Graphalign is still alpha. • We don’t understand parameters well. • Unoptimized. • Not yet competitive with existing approaches. • Broadly applicable! • Hope to engage w/broader community, soon.
Concluding thoughts #1 • None of our theory is particularly limited to De Bruijn graphs, although our implementation is deeply tied to them at the moment. • We view these ideas (streaming; graphs) as a potentially substantial improvement over current mainstream approaches. • We are not alone – there is a larger community exploring these approaches! (GA4GH, esp.)
Concluding thoughts #2 • Our implementations are usable but not yet terribly optimized. • We are moving khmer towards a platform for providing reference implementations of these approaches, as well as for research and development. • We are interested in providing components with decent performance & statistical guarantees, for fun and profit. • Python and C++ FTW!
Thanks! Please contact me at ctbrown@ucdavis.edu!

More Related Content

2015 vancouver-vanbug

  • 1. Building a platform for bioinformatics: some exciting new directions for khmer. C. Titus Brown ctbrown@ucdavis.edu March 12, 2015
  • 2. Hello! Associate Professor (#tenure!); School of Veterinary Medicine University of California, Davis. More information at: • ged.msu.edu/ ( URL needs to be updated :) • github.com/ged-lab/ • ivory.idyll.org/blog/ • @ctitusbrown
  • 3. Warnings This talk contains information that may constitute “forward-looking statements.” Generally, the words “believe,” “expect,” “intend,” “estimate,” “anticipate,” “project,” “will” and similar expressions identify forward-looking statements, which generally are not historical in nature. I have been advised to put this disclaimer in as well: Dr. Brown is not currently under treatment for any disorders related to megalomania.
  • 4. Introducing k-mers CCGATTGCACTGGACCGA (<- read) CCGATTGCAC CGATTGCACT GATTGCACTG ATTGCACTGG TTGCACTGGA TGCACTGGAC GCACTGGACC ACTGGACCGA
  • 5. De Bruijn graphs – assemble on overlaps J.R. Miller et al. / Genomics (2010)
  • 6. K-mers give you an implicit alignment CCGATTGCACTGGACCGATGCACGGTACCGTATAGCC CATGGACCGATTGCACTGGACCGATGCACGGTACCG
  • 7. K-mers give you an implicit alignment CCGATTGCACTGGACCGATGCACGGTACCGTATAGCC CATGGACCGATTGCACTGGACCGATGCACGGTACCG CATGGACCGATTGCACTGGACCGATGCACGGACCG (with no accounting for mismatches or indels)
  • 8. The problem with k-mers CCGATTGCACTGGACCGATGCACGGTACCGTATAGCC CATGGACCGATTGCACTCGACCGATGCACGGTACCG Each sequencing error results in k novel k-mers!
  • 9. The opportunity: CCGATTGCACTGGACCGATGCACGGTACCGTATAGCC CATGGACCGATTGCACTCGACCGATGCACGGTACCG The graph contains information about errors (can be used for error trimming in reads). The graph also contains information about variants (can be used for variant calling).
  • 10. Conway T C , Bromage A J Bioinformatics 2011;27:479-486 © The Author 2011. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com One big challenge: scalability! De Bruijn graph size scales with # errors.
  • 11. One big challenge: scalability! De Bruijn graph size scales with # errors. Memory usage ~ “real” variation + number of errors Number of errors ~ size of data set
  • 12. Conway T C , Bromage A J Bioinformatics 2011;27:479-486 © The Author 2011. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com One big challenge: scalability! De Bruijn graph size scales with # errors.
  • 13. Goals • Initial goal: can we assemble large data sets?? • Longer-term goal: can we find efficient (De Bruijn?) graph-based approaches to sequence analysis?
  • 14. First attempt: compressible De Bruijn graphs 1% 5% 15%10% Pell et al., 2012 Can use Bloom filters to store De Bruijn graph structures. => Overall structure remains as you squish graphs down.
  • 15. Technical challenges met (and defeated) • Exhaustive in-memory traversal of graphs containing 5-15 billion nodes. • Sequencing technology introduces false connections in graph. • Implementation lets us scale ~20x over other approaches. Pell et al., 2012
  • 16. Technical challenges met (and defeated) • Exhaustive in-memory traversal of graphs containing 5-15 billion nodes. • Sequencing technology introduces false connections in graph. • Implementation lets us scale ~20x over other approaches, but this is not enough. • Although, see Minia assembler (Chikhi et al.) Pell et al., 2012
  • 17. Second attempt: diginorm Conway T C , Bromage A J Bioinformatics 2011;27:479-486 © The Author 2011. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com
  • 18. Random sampling => deep sampling needed Typically 10-100x needed for robust recovery (30-300 Gbp for human)
  • 19. Actual coverage varies widely from the average. Low coverage introduces unavoidable breaks.
  • 20. But! Shotgun sequencing is very redundant! Lots of the high coverage simply isn’t needed. (unnecessary data)
  • 27. Contig assembly now scales with underlying genome size • Transcriptomes, microbial genomes incl MDA, and most metagenomes can be assembled in under 50 GB of RAM, with identical or improved results. • Memory efficient is improved by use of CountMin Sketch. Brown et al., 2012, arXiv.
  • 29. Diginorm is only a good start: • Diginorm alters the coverage of the data set. • Diginorm also discards lots of data! • Various other infelicities… o Repeats go away! o Coverage estimation approach ~poor.
  • 30. Diginorm is a good start: • Diginorm works on genomes, metagenomes, and transcriptomes; • Diginorm is streaming and uses sublinear space.
  • 31. Third attempt: a semi-streaming framework for sequence analysis https://github.com/ged-lab/2014-streaming/
  • 32. Diginorm can detect graph saturation Zhang et al., submitted.
  • 33. This generically permits semi- streaming approaches. Zhang et al., submitted.
  • 34. e.g. E. coli analysis => ~1.2 pass, sublinear memory Zhang et al., submitted.
  • 35. => Efficient k-mer error trimming. Zhang et al., submitted. (This all works on metagenomes & transcriptomes, too.)
  • 36. Moving some sequence analysis to streaming. ~1.2 pass, sublinear memory Zhang et al., submitted. First pass: digital normalization - reduced set of k-mers. Second pass: spectral analysis of data with reduced k-mer set. First pass: collection of low-abundance reads + analysis of saturated reads. Second pass: analysis of collected low-abundance reads. First pass: collection of low-abundance reads + analysis of saturated reads. (a) (b) (c) two-pass; reduced memory few-pass; reduced memory online; streaming.
  • 37. Sublinear time/space read error analysis -- Zhang et al., submitted. Read error profile from mouse mRNAseq (c.f. Grabherr et al., 2011).
  • 38. Another simple algorithm. Zhang et al., submitted.
  • 39. So, that’s pretty cool, right? • We provide simple time- and memory-efficient approaches for k-mer spectral analysis of large data sets. • These semi-streaming approaches provide a general framework for applying k-mer spectral approaches to all (deep) sequencing data, including genomes, metagenomes, and RNAseq. • The khmer software provides a functional and reasonably efficient reference implementation, freely available under the BSD license and actively developed at github.com/ged-lab/.
  • 40. Stream all the things! (1/2)
  • 41. Stream all the things! (2/2)
  • 42. But that’s not all! Buy now, and you can also get sequence-to-graph alignment for the low, low price of free!* graph = khmer.new_counting_hash(…) aligner = khmer.ReadAligner(graph, trusted=5) score, graph_align, read_align, is_truncated = aligner.align(seq) * Terms and conditions may apply. Not all source code fully works :)
  • 44. (Full model) Jordan Fish and Michael Crusoe
  • 45. This is a general API… Many potential uses: • Error correction; • Variant calling; • Counting (to replace mapping) & allelic counts; • Align to multiple references; • Tackle strain variation and polyploidy; • Building consensus graphs from shallow population sequencing; • Consensus graph building from multiple read types; • Protein-guided graph search (BlastGraph & Xander)
  • 48. Graphalign is still alpha. • We don’t understand parameters well. • Unoptimized. • Not yet competitive with existing approaches. • Broadly applicable! • Hope to engage w/broader community, soon.
  • 49. Concluding thoughts #1 • None of our theory is particularly limited to De Bruijn graphs, although our implementation is deeply tied to them at the moment. • We view these ideas (streaming; graphs) as a potentially substantial improvement over current mainstream approaches. • We are not alone – there is a larger community exploring these approaches! (GA4GH, esp.)
  • 50. Concluding thoughts #2 • Our implementations are usable but not yet terribly optimized. • We are moving khmer towards a platform for providing reference implementations of these approaches, as well as for research and development. • We are interested in providing components with decent performance & statistical guarantees, for fun and profit. • Python and C++ FTW!
  • 51. Thanks! Please contact me at ctbrown@ucdavis.edu!

Editor's Notes

  1. A sketch showing the relationship between the number of sequence reads and the number of edges in the graph. Because the underlying genome is fixed in size, as the number of sequence reads increases the number of edges in the graph due to the underlying genome that will plateau when every part of the genome is covered. Conversely, since errors tend to be random and more or less unique, their number scales linearly with the number of sequence reads. Once enough sequence reads are present to have enough coverage to clearly distinguish true edges (which come from the underlying genome), they will usually be outnumbered by spurious edges (which arise from errors) by a substantial factor.
  2. A sketch showing the relationship between the number of sequence reads and the number of edges in the graph. Because the underlying genome is fixed in size, as the number of sequence reads increases the number of edges in the graph due to the underlying genome that will plateau when every part of the genome is covered. Conversely, since errors tend to be random and more or less unique, their number scales linearly with the number of sequence reads. Once enough sequence reads are present to have enough coverage to clearly distinguish true edges (which come from the underlying genome), they will usually be outnumbered by spurious edges (which arise from errors) by a substantial factor.
  3. High coverage is essential.
  4. High coverage is essential.
  5. Goal is to do first stage data reduction/analysis in less time than it takes to generate the data. Compression => OLC assembly.