Talk for ASLO Ocean Science Meeting, Honolulu, HI
March 3, 2017
Lisa J. Cohen, Harriet Alexander, C. Titus Brown
The Marine Microbial Eukaryotic Transcriptome Sequencing Project (MMETSP) facilitated the generation of 678 Illumina RNA sequence datasets from a wide diversity of organisms spanning more than 40 phyla of cultured microbial eukaryotes collected from a variety of marine environments. This is the largest publicly available set of RNA sequencing data from a diversity of eukaryotic taxa with a standardized library preparation. We developed an automated and modularized de novo transcriptome assembly pipeline for the MMETSP data set that is extensible to accommodate both future software updates and additional samples. With this large set of assemblies from a diversity of species, we were able to quantitatively evaluate the qualities of individual transcriptomes. Moreover, a meta-analysis across the dataset revealed lineage-specific transcriptome characteristics, such as predicted open reading frames, contig features, unique k-mers and evaluation scores. Ultimately, a better understanding of these assemblies and annotations will enhance our ability to accurately identify and characterize genes of ecological and biogeochemical significance.
REASSEMBLING 600+ MARINE TRANSCRIPTOMES: AUTOMATED PIPELINE DEVELOPMENT AND EVALUATION
1. REASSEMBLING 600+ MARINE TRANSCRIPTOMES:
AUTOMATED PIPELINE DEVELOPMENT AND
EVALUATION
Lisa Cohen, Harriet Alexander, C. Titus Brown
Lab for Data Intensive Biology (DIB), UC Davis
ASLO Aquatic Sciences meeting
Session 016: Advances in Aquatic Meta-Omics
March 3, 2017
@monsterbashseq
ljcohen@ucdavis.edu
2. Marine Microbial Eukaryotic Transcriptome
Sequencing Project (MMETSP)
- 678 Illumina RNA sequence datasets = 1 TB raw data
- Wide diversity spanning more than 40 phyla
- Original assemblies by the National Center for Genome Resources (NCGR)
Keeling et al. 2014
PMID: 24959919
Caron et al. 2016
PMID: 27867198
3. Need for a modularized, extensible RNA-seq pipeline:
o Software and best practices for RNA-seq analysis changing rapidly
(Conesa et al. 2016, PMID: 26813401)
o Accumulating more and more data!
o MMETSP: awesome data set to test software and pipelines!
o What to do if:
• New samples to add?
• New software tool is developed?
6. Questions:
1. Did we generate more biologically-
meaningful content with re-assemblies?
1. Are there phylogenetic patterns in the
assemblies?
7. Smith-Unna et al. 2016
PMID: 27252236
Transrate score = overall quality
of the final assembly (scale 0-1.0)
Qualities of our re-assemblies are higher:
1. Did we generate more biologically-meaningful content with re-assemblies?
Transratescore
0.31
0.22
NCGR DIB
8. Re-assemblies generally contain most of the information in the
NCGR assemblies, plus ~30% more content:
Comparison:
DIB vs. NCGR
DIB
NCGR
Proportionofcontigs(CRB-BLAST)
Comparison:
NCGR vs. DIB
1. Did we generate more biologically-meaningful content with re-assemblies?
NCGR DIB
9. Similar Open Reading Frame (ORF) and
Benchmarks of Universal Single Copy Orthologs (BUSCO)
1. Did we generate more biologically-meaningful content with re-assemblies?
MeanORFpercentage
CompleteBUSCOpercentage
NCGR DIBNCGR DIB
10. Scott, C. in prep. 2016.
www.camillescott.org/dammit
‘dammit’ annotation pipeline: Pfam, Rfam, OrthoDB
annotated absent transcripts
transcripts absent from NCGR
#Transcripts
MMETSP sample (sorted)
1. Did we generate more biologically-meaningful content with re-assemblies?
After annotation, ~30% extra content appears real
DIB
NCGR
Extra content
11. Some DIB assemblies have more unique content.
Unique k-mers (k=25), unique word combinations
1. Did we generate more biologically-meaningful content with re-assemblies?
Probably.
Unique k-mers
(DIB)
Unique k-mers
(NCGR)
12. Assemblies from Dinophyta have more unique k-mers and lower qualities.
Dinoflagellates: steady-state gene expression, translational gene regulation
Aranda et al. 2016 PMID: 28004835
Lin 2011 PMID: 21514379
Hou and Lin 2009. PMID: 27426948
N=
173
111
73
61
60
60
25
22
2. Can we detect phylogenetic differences in the assemblies?
Unique k-mers = unique word combinations (k=25)
13. Ciliophora have lower ORF percentagesN=
173
111
73
61
60
60
25
22
Ciliates: alternative triplet codon dictionary, STOP codon different purpose
Alkalaeva and Mikhailova 2016, PMID: 28009453
Heaphy et al. 2016, PMID: 27501944
Swart et al. 2016, PMID: 27426948
2. Are there phylogenetic differences in the assemblies?
Trends.
Mean % ORF
# contigs
14. Future work:
• In-depth annotation analysis
• Orthologous groupings of contigs
• Co-expression network analysis
• Better reference transcriptomes for MMETSP available:
https://doi.org/10.6084/m9.figshare.3840153.v6
• Strain-specific trends in assemblies support previously-reported
transcriptomic features
• De novo transcriptome assembly pipeline available:
https://github.com/dib-lab/dib-MMETSP
Conclusions
@monsterbashseq
ljcohen@ucdavis.edu
Contact:
15. Acknowledgements
• Data Intensive Biology Lab
– Camille Scott, Luiz Irber
• MSU iCER
• NSF’s XSEDE, Jetstream cloud
• Substituting for my NPB101D
sections today:
– Natalia Caporale, Sheryar
Siddiqui, Pearl Chen, Arik
Davidyan, Karl Larson Photo by James Word
Data Intensive Biology Summer Institute, applications due March 17th!
http://ivory.idyll.org/dibsi/
16. Files available for download!
Cohen, Lisa; Alexander, Harriet; Brown, C. Titus (2017): Marine Microbial
Eukaryotic Transcriptome Sequencing Project, re-assemblies. figshare.
https://doi.org/10.6084/m9.figshare.3840153.v6
https://github.com/dib-lab/dib-MMETSP
@monsterbashseq
ljcohen@ucdavis.edu
Data Intensive Biology Summer Institute, applications due March 17th!
http://ivory.idyll.org/dibsi/
Editor's Notes
Hi, my name is Lisa Cohen, I’m a PhD student at UC Davis. Thank you for this opportunity to speak today. I would like to first acknowledge my co-authors, Harriet Alexander, who is sitting in the audience today and my advisor, Titus Brown.
The Marine Microbial Eukaryotic Sequencing Project is a unique set of mRNA sequence data generated by a consortium of PIs who all got together and submitted their favorite marine microbial eukaryotes to one sequencing facility. These species represent 40 pelagic and endosymbiotic phyla, such dinoflagellates, ciliates, diatoms. They are both phylogenetically diverse and geographically diverse, collected from all over the world.
This is a really exciting set of data for a few reasons, one is because it is one of the largest publicly available sets of RNA data with a standardized library preparation from different organisms with a total of about 1 TB of raw sequence data.
Second, it’s purposefully built, not a metatranscriptome. We technically know who is supposed to be in this data set, so we are generating reference transcriptomes for all of these species, some of which have never had any reference transcriptomes or genomes before.
Right after data were sequenced, the NCGR assembled the transcriptomes as references with their own pipeline, using the genome assembler ABySS with some modifications and post-processing for transcriptomes.
====================
Bottom panel, left to right:
Elphidium margaritaceum
http://zoology.bio.spbu.ru/Eng/Sci/Korsun/Foram2_E-margaritaceum.jpg
2. Acanthamoeba
https://upload.wikimedia.org/wikipedia/commons/thumb/1/1b/Parasite140120-fig3_Acanthamoeba_keratitis_Figure_3B.png/220px-Parasite140120-fig3_Acanthamoeba_keratitis_Figure_3B.png
3. Gonyaulax spinifera
http://www.sms.si.edu/IRLSpec/images/Gonyaulax_Lg.jpg
4. Asterionellopsis glacialis
http://www.smhi.se/oceanografi/oce_info_data/plankton_checklist/diatoms/asterionellopsis_glacialis.gif
5. Tetraselmis
http://cfb.unh.edu/phycokey/Choices/Chlorophyceae/unicells/flagellated/TETRASELMIS/Tetraselmis_06_500x345.jpg
6. Oxyrrhis marina
http://cfb.unh.edu/phycokey/Choices/Dinophyceae/NonPS-dinos/OXYRRHIS/Oxyrrhis_04_300x246_marina.jpg
7. Alexandrium
http://www.whoi.edu/cms/images/dfino/2006/6/Alexandrium_en_11187_26907.jpg
8. Pseudonitzschia
https://upload.wikimedia.org/wikipedia/commons/5/5e/Pseudonitzschia2.jpg
9. Chlamydomonas
https://web.mst.edu/~microbio/BIO221_2009/images_2009/chlamydomonas-3.jpg
10. Emiliania_huxleyi
https://upload.wikimedia.org/wikipedia/commons/d/d9/Emiliania_huxleyi_coccolithophore_(PLoS).png
11. Symbiodinium
http://www.personal.psu.edu/tcl3/index.html
12. Phaeocystis antarctica
http://www.esf.edu/antarctica/images/Phaeo_montage2.jpg
13. Micromonas
http://roscoff-culture-collection.org/sites/default/files/field/image/micromonas-colored-350_0.jpg
14. Karenia brevis
http://www.sms.si.edu/irlspec/images/Kareni_brevis_2.jpg
15. Thalassiosira pseudonana
http://genome.jgi.doe.gov/Thaps3/Tpseudonana.jpg
16. Ditylum_brightwellii
https://cimt.pmc.ucsc.edu/images/HAB%20ID/diatom/Ditylum_brightwellii.jpg
So, when I was starting my PhD about a year and a half ago, it was becoming apparent that software and best practices for RNA sequencing analysis and de novo transcriptome assembly are not standard and changing rapidly.
Pipelines developed for model animal species do not necessarily hold true for all species.
We’re also collecting more and more data! RNA sequence data in particular.
The MMETSP data are great to use to test software and analysis pipelines! Because of its size and because the organisms are diverse, we can better understand how these tools are performing with data from difference species.
Some of the problems that I and others in Titus Brown’s lab think about is what happens if a PI wants to submit just one more sample? What happens if there are shiny new tools developed?
Our modularized pipeline, which I wrote in Python, attempts to address these issues. It takes metadata from any data set in NCBI as input and decides which samples to run.
Raw sequence reads are downloaded from NCBI, quality trimmed, checked with fastqc, run through digital normalization, then assembled using the Trinity transcriptome assembler.
I’m glossing over a lot of details here because there is not enough time, but if you are interested please see me after to talk. There is a tutorial also available, called the “Eel pond protocol”, which is open access and has a small subset of data to run through the steps of a de novo assembly with Trinity.
A benefit of this pipeline to highlight is that you can pick up from where you left off if something crashes. As anyone who has used an institutional high performance computing cluster knows, stuff breaks, stops running. With this pipeline, if something stops, you can start it again.
I also want to mention that this data set pushes the limits of high performance computing clusters with 1 TB raw data, in terms of storage and compute resources. This took more than 8,000 computing hours, We have found that the resources required for these >600 assemblies are not trivial, and should be a consideration when planning for a project of this size in the future.
In evaluating our assemblies, it appears that our re-assemblies have more contigs. A contig is a linear prediction of a full transcript by the assembly software. In subsequent slides, I’ll be showing similar figures like this, so want to orient you first. On the y-axis is what we’re measuring – here it’s the number of contigs. This is a split violin plot showing the frequency distribution around the mean of each pipeline. In the blue on the right shows our re-assemblies, which I’ve labeled “DIB” because we’re the data intensive biology lab. In the gray on the left are assemblies from NCGR. The number on top in blue shows the numbers of assemblies where DIB has a higher value than NCGR or in gray where NCGR has a higher number.
In this case, we see that there were more DIB assemblies with higher numbers of contigs in comparison to the NCGR.
The mean of DIB is around 48,000 contigs, with some samples producing up to 190,000 contigs up here towards the tail of the distribution. While the mean of NCGR is around 25,000 contigs and fewer assemblies have high numbers of contigs, the highest is about 100,000.
So, these differences were interesting for us – and we came up with some questions (click)
One, we’re interested to see whether we’ve assembled more things. It could be that this is just fragmented junk. But it could also be relevant, being able to resolve allelic variants or alternative splicing. Or just pieces of the same transcript. Theoretically, each contig is supposed to represent one transcript, but we can’t really say that yet.
The second question has to do with the biological differences in our samples. They are from different taxonomic groupings. So, we’re wondering if the software is performing differently based on what species the data come from. The relationship between raw data content and assembly quality is not well understood. So, with this data set, we’re wondering whether the nucleic acid sequence information is being handled differently by the software tools.
The qualities of our assemblies appear to be higher. Transrate is a software tool that was developed to help you understand your transcriptome based on a variety of metrics. One of those metrics is an overall synthetic quality score for a transcriptome, which is called the “transrate score”.
Our mean transrate score, while the NCGR transrate score is 0.22
In addition to have higher quality scores, there appears to be more content. The proportion of contigs from a comparison called a reciprocal best blast of NCGR vs. our DIB assemblies indicates that most of the content found in NCGR is also found in the DIB re-assemblies. But also that there is extra information in the DIB assemblies not found in the NCGR assemblies. This information was obtained by aligning the two assemblies against each other both ways. First with NCGR as the reference, then the reverse with DIB as the reference.
Engage with audience: As you can see here…our peak is about 0.8, or 80%. This means that we’re capturing 80% of the content in the NCGR assemblies. On the other hand, NCGR assemblies capture about 50% of the content of our assemblies. The difference is about 30%.
The ~30% difference between these 2 blast comparisons leads us to still question whether we have just assembled junk or if we actually have higher resolution assemblies.
Orient audience to graphs: left ORF on Y axis
Even though we have more contigs, the open reading frame protein coding regions detected is similar if not more tightly distributed towards the upper range. Most of the assemblies have slightly higher ORF content.
And on the right are BUSCO percentages, which is a set of benchmarking universal single copy orthologs expected to be found in all eukaryotic transcriptomes, like housekeeping genes.
While there are problems with using BUSCO scores as an absolute measurement of assembly quality, they can serve as a comparative metric relative to another pipeline. Our assemblies have a similar if not slightly higher BUSCO content relative to NCGR. So, at least these haven’t gone down. The extra content we found is probably not all junk.
In digging deeper into the extra content, this is a plot of ONLY this extra content in the blue part. Samples are across the x axis, sorted by the number of extra contigs on the y axis. (pause, let this sink in, take a drink or something)
Highlighted in green is the number of these extra contigs that are actually annotated to a known gene.
I annotated the re-assemblies using this really great tool out of our lab by Camille Scott called ‘dammit’. No, it’s not an acronym, it was named out of frustration: “Just annotate it, dammit!” The dammit pipeline uses the highly-curated Pfam and Rfam known protein domain databases as well as ORthoDB with conserved orthology domains. About 1/3 of the extra content has annotations.
This is a pared-down example of what the annotations look like from one of the Dinoflagellate samples, to illustrate some of our frustrations with contigs and annotations. The assembler will recognize a contig as a transcript, then the dammit pipeline will find matches with the databases. There are usually multiple proteins that match, so I’ve chosen the top e-value match so that there is only one protein annotated per transcript. Here you can see that there are multiple contigs annotated as the same protein, glycoprotein glucosyltransferase.
So, this has been a challenge to sift through all of this. But – again - it is great to have these annotations.
Porocentrum minimum
https://www.eoas.ubc.ca/research/phytoplankton/dinoflagellates/prorocentrum/p_minimum.html
Here we are comparing the raw sequence content, regardless of annotation, in terms of the number of kmers or unique word combinations with a k length of 25. We see that our assemblies fall above the 1:1 expectation, meaning that our assemblies have more unique words compared to the NCGR assemblies. This is kind of like taking two versions of the same book and digesting them down into individual 25 letter words found in the book. We found that our assemblies have more unique words than NCGR.
Therefore, we are able to answer that our assemblies probably have a bit more biologically-meaningful content
To address our second question about whether we can detect phylogenetic differences in the assemblies, we took a look at some of the assembly metrics grouped by taxa.
Explain figures: unique k-mers on the y, input reads on the x, colors indicate different taxa, plotting mean and stdev
The Dinoflagellates appear to have more unique kmer content. This seems to make sense, knowing that Dinoflagellates have this steady-state gene expression thing going on, where they just keep expressing genes on and one, then regulate more at the translational level.
As far as the software, it might be useful to incorporate strain-specific information like this into assembly software.
Here again, colors are different taxonomic groupings, mean percentage of open reading frame predictions on the y, number of transcripts on the x
We see here that Cilliate assemblies appear to have a lower open reading frame percentage. This is interesting since it has recently been found Ciliates have an alternative triplet codon dictionary, with codons normally encoding STOP serving a different purpose.
Dinoflagellates here have this high open reading frame content, and lots of contigs.
In this case, it is useful to know that our assembly evaluation tools might perform outside the range of what is normal for the organisms in question. The assemblies are not necessarily lower quality, but may be perceived as lower in quality because of cool and unique features like this.
Strain-specific trends may lead to understanding how raw data content affects the overall assembly quality