2015 genome-center
- 1. A data intensive future: How can biology take full advantage of the coming data deluge? C.Titus Brown School ofVeterinary Medicine; Genome Center & Data Science Initiative 11/13/15
- 2. Outline 0. Background 1. Research: what do we do with infinite data? 2. Development: software and infrastructure. 3. Open science & reproducibility. 4. Training
- 3. 0. Background In which I present the perspective that we face increasingly large data sets, from diverse samples, generated in real time, with many different data types.
- 4. DNA sequencing rates continues to grow. Stephens et al., 2015 - 10.1371/journal.pbio.1002195
- 5. Oxford Nanopore sequencing Slide viaTorsten Seeman
- 6. Nanopore technology Slide viaTorsten Seeman
- 11. “Fighting EbolaWith a Palm- Sized DNA Sequencer” See: http://www.theatlantic.com/science/archive/2015/09/ebola- sequencer-dna-minion/405466/
- 12. “DeepDOM” cruise: examination of dissolved organic matter & microbial metabolism vs physical parameters – potential collab. Via Elizabeth Kujawinski
- 13. 1. Research In which I discuss advances made towards analyzing infinite amounts of genomic data, and the perspectives engendered thereby: to whit, streaming and sketches.
- 14. 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 De Bruijn graphs (sequencing graphs) scale with data size, not information size.
- 15. Why do sequence graphs scale badly? Memory usage ~ “real” variation + number of errors Number of errors ~ size of data set
- 16. Practical memory measurements Velvet measurements (Adina Howe)
- 17. Our solution: lossy compression 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. Shotgun sequencing and coverage “Coverage” is simply the average number of reads that overlap each true base in genome. Here, the coverage is ~10 – just draw a line straight down from the top through all of the reads.
- 19. Random sampling => deep sampling needed Typically 10-100x needed for robust recovery (30-300 Gbp for human)
- 26. Graph sizes now scales with information content. Most samples can be reconstructed via de novo assembly on commodity computers.
- 27. Diginorm ~ “lossy compression” Nearly perfect from an information theoretic perspective: – Discards 95% more of data for genomes. – Loses < 00.02% of information.
- 28. This changes the way analyses scale. 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
- 29. Streaming lossy compression: for read in dataset: if estimated_coverage(read) < CUTOFF: yield read This is literally a three line algorithm. Not kidding. It took four years to figure out which three lines, though…
- 30. Diginorm can detect information saturation in a stream. Zhang et al., submitted.
- 31. This generically permits semi-streaming analytical approaches. Zhang et al., submitted.
- 32. e.g. E. coli analysis => ~1.2 pass, sublinear memory Zhang et al., submitted.
- 33. Another simple algorithm. Zhang et al., submitted.
- 34. Single pass, reference free, tunable, streaming online variant calling. Error detection variant calling
- 35. Real time / streaming data analysis. Raw data (real time, from sequencer?) Error trimming Variant calling De novo assembly
- 36. My real point - • We need well founded, and flexible, and algorithmically efficient, and high performance components for sequence data manipulation in biology. • We are building some of these on a streaming and low memory paradigm. • We are building out a scripting library for composing these operations.
- 37. 2. Software and infrastructure Alas, practical data analysis depends on software and computers, which leads to depressingly practical considerations for gentleperson scientists.
- 38. Software It’s all well and good to develop new data analysis approaches, but their utility is greater when they are implemented in usable software. Writing, maintaining, and progressing research software is hard.
- 39. The khmer software package • Demo implementation of research data structures & algorithms; • 10.5k lines of C++ code, 13.7k lines of Python code; • khmer v2.0 has 87% statement coverage under test; • ~3-4 developers, 50+ contributors, ~1000s of users (?) The khmer software package, Crusoe et al., 2015. http://f1000research.com/articles/4-900/v1
- 40. khmer is developed as a true open source package • github.com/dib-lab/khmer; • BSD license; • Code review, two-person sign off on changes; • Continuous integration (tests are run on each change request);
- 41. Challenges: Research vs stability! Stable software for users, & platform for future research; vs research “culture” (funding and careers)
- 42. How is continued software dev feasible?! Representative half-arsed lab software development Version that worked once, for some publication. Grad student 1 research Grad student 2 research Incompatible and broken code
- 43. A not-insane way to do software development Stable version Grad student 1 research Grad student 2 research Stable, tested code Run tests Run tests Run tests Run tests Run tests Run tests Run tests
- 44. Infrastructure issues Suppose that we have a nice ecosystem of bioinformatics & data analysis tools. Where and how do we run them? Consider: 1. Biologists hate funding computational infrastructure. 2. Researchers are generally incompetent at building and maintaining usable infrastructure. 3. Centralized infrastructure fails in the face of infinite data.
- 45. Decentralized infrastructure for bioinformatics? Compute server (Galaxy? Arvados?) Web interface + API Data/ Info Raw data sets Public servers "Walled garden" server Private server Graph query layer Upload/submit (NCBI, KBase) Import (MG-RAST, SRA, EBI) ivory.idyll.org/blog/2014-moore-ddd-award.html
- 46. 3. Open science and reproducibility In which I start from the point that most researchers* cannot replicate their own computational analyses, much less reproduce those published by anyone else. *This doesn’t apply to anyone in this audience; you’re all outliers!
- 47. My lab & the diginorm paper. • All our code was on github; • Much of our data analysis was in the cloud (on Amazon EC2); • Our figures were made in IPython Notebook. • Our paper was in LaTeX. Brown et al., 2012 (arXiv)
- 48. IPython Notebook: data + code => IPython)Notebook)
- 49. To reproduce our paper: git clone <khmer> && python setup.py install git clone <pipeline> cd pipeline wget <data> && tar xzf <data> make && cd ../notebook && make cd ../ && make
- 50. This is standard process in lab -- Our papers now have: • Source hosted on github; • Data hosted there or onAWS; • Long running data analysis => ‘make’ • Graphing and data digestion => IPython Notebook (also in github) Zhang et al. doi: 10.1371/journal.pone.0101271
- 51. Research process Generate new results; encode in Makefile Summarize in IPython Notebook Push to githubDiscuss, explore
- 52. Literate graphing & interactive exploration Camille Scott
- 53. Why bother?? “There is no scientific knowledge of the individual.” (Aristotle) More pragmatically, we are tired of struggling to reproduce other people’s results. And, in the end, it’s not all that much extra work.
- 54. What does this have to do with open science? This is a longer & larger conversation, but: All of our processes enable easy and efficient pre-publication sharing. Source code, analyses, preprints… When we share early, our ideas have a significant competitive advantage in the research marketplace of ideas.
- 55. 4.Training In which I note that methods and tools do little without a trained hand wielding them, and a trained eye examining the results.
- 56. Perspectives on training • Prediction: The single biggest challenge facing biology over the next 20 years is the lack of data analysis training (see: NIH DIWG report) • Data analysis is not turning the crank; it is an intellectual exercise on par with experimental design or paper writing. • Training is systematically undervalued in academia (!?)
- 57. UC Davis and training My goal here is to support the coalescence and growth of a local community of practice around “data intensive biology”.
- 58. Summer NGS workshop (2010-2017)
- 59. General parameters: • Regular intensive workshops, half-day or longer. • Aimed at research practitioners (grad students & more senior); open to all (including outside community). • Novice (“zero entry”) on up. • Low cost for students. • Leverage global training initiatives.
- 60. Thus far & near future ~12 workshops on bioinformatics in 2015. Trying out soon: • Half-day intro workshops; • Week-long advanced workshops; • Co-working hours. dib-training.readthedocs.org/
- 61. The End. • If you think 5-10 years out, we face significant practical issues for data analysis in biology. • We need new algorithms/data structures, AND good implementations, AND better computational practice, AND training. • This can be either viewed with despair… or seen as an opportunity to seize the competitive advantage! (How I view it varies from day to day.)
- 62. Thanks for listening! Please contact me at ctbrown@ucdavis.edu! Note: I work here now!
Editor's Notes
- 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.
- Goal is to do first stage data reduction/analysis in less time than it takes to generate the data. Compression => OLC assembly.
- Taking advantage of structure within read
- Analyze data in cloud; import and export important; connect to other databases.
- Lure them in with bioinformatics and then show them that Michigan, in the summertime, is qite nice!