Trying to download the cow genome: where's the beef?

Cows on the UC Davis campus, photo by Keith Bradnam

I want to download the cow genome. A trip to bovinegenome.org seemed like a good place to start. However, the home page of this site is quick to reveal that there is not only assembly Btau_4.6 that was released by the Baylor College of Medicine Human Genome Sequencing Center, there is also an alternative UMD3.1 assembly by the Center for Bioinformatics and Computational Biology (CBCB) at the University of Maryland.

In many areas of life, choice is a good thing. But in genomics, trying to decide which of two independent genome resources to use, well that can be a bit of a headache. The bovinegenome.org site tells me that 'Btau_4.6' was from November 2011. I have to go to the CBCB site to find out that UMD3.1 was made available in December 2009.

Bovinegenome.org helpfully provides a link to the UMD3.1 assembly. Not so helpfully, the link doesn't work. At this point I decide to download the 'Btau_4.6' assembly and associated annotation. The Btau downloads page provides a link to the Assembly 4.0 dataset (in GFF3 format) under a heading of 'Datasets for Bovine Official Gene Set version 2'. So at this point, I have to assume that maybe that the '4.0' might really be referring to the latest 4.6 version of the assembly.

The resulting Assembly 4.0 page has links for each chromosome (available as GFF3 or gzipped GFF3). These links include a 'ChrUn' entry. You will see this in most modern genome assemblies as there is typically a lot of sequence (and resulting annotation) that is 'unplaced' in the genome, i.e. it cannot be mapped to a known chromosome. In the case of cow, ChrUn is the second largest annotation file.

This page also has four more download links for what bovinegenome.org refers to as:

• 'Reference Sequences'
• 'CDS Target'
• 'Metazoa Swissprot Target'
• 'EST Target'

View fullsize

From the Assembly 4.0 downloads page (which may or may not mean version 4.6) 

No README file or other explanation is provided for what these files contain and if you try to download the GFF3 links (rather than the gzipped versions), you will find that these links don't work. The 'CDS Target' link lets you download a file named 'incomplete_CDS_target.gff3'. I do not know whether the 'incomplete' part refers to individual CDSs (missing stop codons?) or the set of CDSs as a whole. I also do not know what the 'target' part refers to.

The chromosome-specific GFF3 files only contain coordinates of gene features, which utilize several different GFF source fields:

• NCBI_transcript
• annotation
• bovine_complete_cds_gmap_perfect
• ensembl
• fgenesh
• fgeneshpp
• geneid
• metazoa_swissprot
• sgp

Again, no README file is provided to make sense of these multiple sources of gene information. In contrast to the chromosome-specific files, the 'reference_sequences.gff3' and 'incomplete_CDS_target.gff3' contain coordinate information and embedded sequence information in FASTA format (this is allowed in the GFF3 spec).

If you look at the coordinate information in the reference sequence file, you see multiple chromosome entries per biological chromosome. E.g.

View fullsize

From the reference_sequences.gff3 file

I do not know what to make of these.

If you want to get the sequences of all the genes, you have to go to yet another download page at bovinegenome.org where you can download FASTA files for CDS and proteins of either the Full Bovine Official Gene Set v2 or the Subset of Bovine Official Gene Set v2 derived from manual annotations. What exactly is the process that generated the 'manual annotations'? I do not know. And bovinegenome.org does not say.

So maybe the issue here is that I should get the information straight from the horse's — or should that be cow's? — mouth. So I head over to the Bovine Genome Project page at BCM HGSC. This reveals that the latest version of the genome assembly is 'Btau_4.6.1' from July 2012! This would seem to supercede the 2011 'Btau_4.6' assembly (which is not even listed on the BCM HGSC page), but I guess I'll have to download it to be sure. The link on the above BCM HGSC page takes me to an FTP folder which is confusingly named 'Btau20101111' (they started the 2012 assembly in 2010?).

So it seems obvious that all I need to do now is to: check whether the bovinegenome.org 'Assembly 4.0' dataset is really listing files from the 'Btau_4.6' release; find out whether those files are different from the BCM HGSC 'Btau_4.6.1' release; understand what the differences are between the nine different sets of gene annotation; decipher what the 'CDS target' (a.k.a. 'incomplete_CDS_target') file contains, understand why there are 14,195 different 'chromosome' features in the genome GFF file that describes 30 chromosomes; and finally compare and contrast all of these sequences and annotations with the alternative UMD3.1 assembly.

Bioinformatics is fun!

Update 1 — 2014-03-14 14.58 pm: The BCM HGSC has an extremely detailed README file that explains the (slight) differences between Btau_4.6 and Btau_4.6.1 (as well as explaining many other things about how they put the genome together). I like README files!

Update 2 — 2014-03-14 15.18 pm: Of course I could always go the Ensembl route and use their version of the cow genome sequence with their annotations. However, the Ensembl 75 version of the cow genome has been built using the older UMD3.1 assembly, whereas an older Ensembl 63 version of the cow genome that is based on the Btau_4.0 assembly is also available (life is never simple!). One advantage to working with Ensembl is that they are very consistent with their annotation and nomenclature.

Update 3 — 2014-03-14 15.35 pm: I now at least realize why there are 14,195 'chromosome' features in the main GFF file provided by bovinegenome.org. They are treating each placed scaffold as its own 'chromosome' feature. This is particularly confusing because a GFF feature field of 'chromosome' is usually taken to refer to an entire chromosome. GFF feature names should be taken from Sequence Ontology terms (e.g. SO:0000340 = chromosome), so in this case the SO term chromosome_part would be more fitting (and less confusing). 

Update 4 — 2014-03-17 11.16 am: I have received some very useful feedback since writing this blog post:

• I was kindly pointed towards the NCBI Genome Remapping Service by @deannachurch (also see her comment below) that would let me project one annotation data from one coordinate system to another. This could be handy if I end up using different cow genomes.
• A colleague of mine pointed me to a great 2009 article by @deannachurch and LaDeanna Hillier that talks about how the simultaneous publication of two cow genomes (from the same underlying data) raises issues about how we deal with publication and curation of such data sets.
• Several people contacted me to give their opinion as to the relative merits of the UMD and the BCM HGSC assemblies.
• Several people also echoed my sentiments about working with the same cow data, but by using Ensembl's architecture instead.
• I was contacted by someone from bovinegenome.org who said that they would fix the issues that I raised. This was after I contacted them to tell them I had blogged about the problems that I had encountered. 

The last point is perhaps the most important one which I would like to dwell on for a moment longer. If you are a scientist and experience problems when trying to get hold of data from a website (or FTP site) don't sit in silence and do nothing!

If nobody tells a webmaster that there is an issue with getting hold of data, then the problem will never be fixed. Sometimes files go missing from FTP sites, or web links break, or simple typos make things unclear....the curators involved may never know unless it is pointed out to them.

Next-generation sequencing must die (part 2) — understanding the generation gap

As a brief follow-up to my previous post, I'd like to clarify that next-generation sequencing may refer to technologies from Illumina, 454, SOLiD, Helicos, Ion Torrent, Complete Genomics, PacBio, or Oxford Nanopore (these links all refer to different papers).

If we want to get more specific, we need to recognize that Complete Genomics is a second generation technology...except when it is a third generation technology. In contrast, we should be clear that Oxford Nanopore is the only example of fourth generation technology...apart from when it is third generation technology. We can at least be sure that Ion Torrent is definitely a second generation platform...unless it's a third generation platform. One paper clarifies this situation by observing that Ion Torrent "sits between" the second and third generation categories.

Further illumination on this subject is provided by the confirmation that PacBio is either a second generation, third generation...or even a "2.5th" generation technology. Likewise, Helicos is also a second generation, third generation, or lies "in between the transition of next-generation sequencing to third generation" sequencing technologies.

So hopefully that's a lot clearer now.

Next-generation sequencing must die!

I hate the phrase next-generation sequencing (NGS) with a passion. Here's why...

The first published use of this phrase (that I can find) is from an article in Drug Discovery Today: Technologies by Thomas Jarvie in 2005. This paper had the succinct title Next generation sequencing technologies, and while this may represent the first time this phrase made it into print, it certainly wouldn't be the last.

Illumina sequencing may be the most obvious technology that springs to mind when people think of NGS, but there is also Pyrosequencing (developed circa 1996), Massively Parallel Signature Sequencing (circa 2000), ABI SOLiD sequencing (circa 2008), and Ion semiconductor sequencing (circa 2010).

Of course we also have single molecule real time sequencing by Pacific Biosciences. They were founded in 2004 but didn't launch their PacBio RS machines until 2010. The current darling of the sequencing world is nanopore technology, something which has been in development since the mid 1990s.

So do we refer to this entire period (from development to finished technology) as the NGS era? If so, then NGS technologies have already been around for almost 20 years. It doesn't strike me as particularly helpful to keep on labeling all of these different technologies with the same name.

Some people have tried to make things clearer by introducing yet more levels of obfuscation. This has led to some of these technologies being referred to as either second generation, third generation, fourth generation, next-next-generation, and even next-next-next generation. And of course these definitions are all subjective and one person's third-generation technology is another person's fourth-generation technology.

Other alternatives to NGS such as high-throughput sequencing or long-read sequencing are equally useless because 'high' and 'long' are both relative terms. The output from a high-throughput sequencing platform of 2008 might seem like 'low-throughput' today. The weakness of length-based descriptions is the reason why the 'Short Read Archive' was (thankfully) reborn as the Sequence Read Archive.

So here is my proposed three-step solution to rid the world of this madness:

1. Don't use any of these terms, ever again.
2. Just refer to a technology by a name that describes the methodology (e.g. sequencing-by-synthesis) or by the name of a company that has developed a specific product (e.g. Oxford Nanopore).
3. You could even just use the term 'current sequencing technologies'...as long as your paper/talk/blog/book has a date associated with it, then I'm confident people will know what you mean.

Update 12th March: I have added a follow-up post to this one.

Some more JABBA awards to highlight bad bioinformatics acronyms

More JABBA awards for bogus bioinformatics acronyms. These all come from a recent issue of Bioinformatics:

• PEAR: a fast and accurate Illumina Paired-End reAd mergeR — if they allowed you to download two copies of a stripped down version of this tool, would that be a pair of pared PEARs?
• SecStAnT: Secondary Structure Analysis Tool for data selection, statistics and models building — What is it with this trend towards using mixed-case for everything?
• ORMAN: Optimal resolution of ambiguous RNA-Seq multimappings in the presence of novel isoforms — The title is not quite revealing the source of the acronym. The tool's name stems from Optimal Resolution of Multimapping Ambiguity of rNa-seq reads. Would ORMAR — or even ORMARR — have been such a terrible name?

Honorable mention:

• orthAgogue: an agile tool for the rapid prediction of orthology relations — This is not an acronym or initialism, but I still have to ask...why is it just the 'A' that is upper case?

Celebrating an unsung hero of genomics: how Albrecht Kossel saved bioinformatics from a world of hurt

The following image is from one of the first publications to ever depict a DNA sequence in a textual manner:

From Wu & Kaiser, Journal of Molecular Biology, 1968. 

This is taken from the 1968 paper by Wu and Kaiser: Structure and base sequence in the cohesive ends of bacteriophage lambda DNA. For almost half a century since this publication, it has become the norm to simply represent the sequence of any DNA molecule as a string of characters, one character per base. But have you ever stopped to consider why these bases have the names that they do?

Most of the work to isolate and describe the purines and pyrimidines that comprise nucleic acids came from the work of the German biochemist Albrecht Kossel. Between 1885 and 1901 he characterized the principle nucleobases that comprise the nucleic acids: adenine, cytosine, guanine, and thymine (though guanine had first been isolated in 1844 by Heinrich Gustav Magnus). The fifth nucleobase (uracil) was discovered by Alberto Ascoli, a student of Kossel.

Kossel's work would later be recognized with the 1910 Nobel Prize for Medicine. It should be noted that Kossel didn't just help isolate and describe these bases, he was also chiefly responsible for most of their names.

Albrecht Kossel, Image from wikimedia

Guanine had already been named based on where it had first been discovered (the excrement of seabirds known as guano). Adenine was so named by Kossel because it was isolated from the pancreas gland ('adenas' in Greek). Thymine was named because it was isolated in nucleic acids from the thymus of a calf. Cytosine — the last of the four DNA bases to be characterized — was also discovered from hydrolysis of the calf thymus. Its name comes from the original name in German ('cytosin') and simply refers to to the Greek prefix for cell ('cyto').

While this last naming choice might seem a little dull, all bioinformaticians owe a huge debt of thanks to Albrecht Kossel. Thankfully, he ensured that all DNA bases have names that start with different letters. This greatly facilitates their representation in silico. Imagine if he had — in a fit of vanity — instead chosen to name these last two bases that he characterized after himself and his daughter Gertrude. If that had been the case then maybe we would today be talking about the bases adenine, albrechtine, guanine, and gertrudine. Not an insurmountable problem to represent with single characters — we already deal with the minor headache of representing purines and pyrimidines differently (using R and Y respectively) — but frankly, it would be a royal pain in the ass.

Thank you Albrecht. The world of bioinformatics is in your debt.

The growth of bioinformatics papers that mention 'big data'

I very much enjoyed Stephen Turner's recent blog post There is no Such Thing as Biomedical "Big Data" and I agree with his central point that a lot of the talk about 'big data' is not really what others would consider 'big'. Out of curiosity, I had a quick dive into Google Scholar to see just how popular this particular cliche is becoming. My search term was "big data" biology|genomics|bioinformatics.

Growth of bioinformatics papers on Google Scholar that mention "big data".

Clearly, this term is on the rise and might become as much of an annoyance as another phrase I loathe: next generation sequencing. A phrase that has been used to describe everything from 25 bp reads from early Solexa technology (circa 2005) to PacBio subreads that can exceed 25,000 bp.

As more people use N50 as a metric, fewer genomes seem to be 'completed'

If you search Google Scholar for the term genome contig|scaffold|sequence +"N50 size|length" and then filter by year, you can see that papers which mention N50 length have increased dramatically in recent years:

Google Scholar results for papers that mention N50 length. 2000–2013.

I'm sure that my search term doesn't capture all mentions of N50, and it probably includes a few false positives as well. It doesn't appear to be mentioned before 2001 at all, and I think that the 2001 Nature human genome paper may have been the first publication to use this metric.

Obviously, part of this growth simply reflects the fact that more people are sequencing genomes (or at least writing about sequenced genomes), and therefore feel the need to include some form of genome assembly metric. A Google Scholar search term for "genome sequence|assembly" shows another pattern of growth, but this time with a notable spurt in 2013:

Google Scholar results for papers that mention genome sequences or assemblies. 2000–2013.

Okay, so more and more people are sequencing genomes. This is good news, but only if those genomes are actually usable. This led me to my next query. How many people refer to their published genome sequence as complete? I.e. I searched Google Scholar for "complete|completed genome sequence|assembly". Again, this is not a perfect search term, and I'm sure it will miss some descriptions of what people consider to be complete genomes. But at the same time it probably filters out all of the 'draft genomes' that have been published. The results are a little depressing:

Google Scholar results for papers that mention genome sequences or assemblies vs those that make mention of 'completed' genome sequences or assemblies. 2000–2013.

So although there were nearly 90,000 publications last year that mentioned a genome sequence (or assembly), approximately just 7,500 papers mentioned the C-word. This is a little easier to visualize if you show the number of 'completed' genome publications as a percentage of the number of publications that mention 'genome sequence' (irrespective of completion status):

Numbers of publications that mention 'completed' genomes as percentage of those that mention genomes. 2000–2013.

Maybe  journal reviewers are more stringent about not allowing people to use the 'completed' word if the genome isn't really complete (which depending on your definition of 'complete' may include most genomes)? Or maybe people are just happier these days to sequence something, throw it through an assembler and then publish it, regardless of how incomplete it is?

2013 ended with a bumper crop of new JABBA awards for bogus bioinformatics acronyms

I have a huge backlog of JABBA awards to hand out. These are all collated from the annual publication of the (voluminous) 2013 Nucleic Acids Research Database Issue. So without further ado, here are the recipients of the latest batch of JABBA awards:

• IDEAL in 2014 illustrates interaction networks composed of intrinsically disordered proteins and their binding partners — If you want to know what IDEAL means, I can tell you that it is an acronym for Intrinsically Disordered proteins with Extensive Annotations and Literature. The 'E' for Extensive is clearly added to make the word IDEAL. They could have also used 'Elaborate', 'Electronic', or maybe 'Eclectic'. To their credit, a Google search for IDEAL database does put them as the top hit. Now that's ideal.
• Updates on the web-based VIOLIN vaccine database and analysis system — The full name of this resource is The integrative Vaccine Investigation and Online Information Network, so they missed an opportunity to name it 'iVIOLIN'.
• HRaP: database of occurrence of HomoRepeats and patterns in proteomes — Good job they were not building a database of 'canonical' repeat patterns.
• RADAR: a rigorously annotated database of A-to-I RNA editing — Strikes me as more 'ridiculous' than 'rigorous'.
• NPInter v2.0: an updated database of ncRNA interactions — You have to read the abstract for version 1 of this tool to discover where the 'P' comes from, the database is fully named The noncoding RNAs and protein related biomacromolecules interaction database.
• The YEASTRACT database: an upgraded information system for the analysis of gene and genomic transcription regulation in Saccharomyces cerevisiae — The acronym here is derived from YEAst Search for Transcriptional Regulators And Consensus Tracking (as if that wasn't obvious).
• 3did: a catalog of domain-based interactions of known three-dimensional structure — In which '3did' refers to 3D Interacting Domains. It is a fairly widespread convention that 3D uses an uppercase 'D' when referring to dimensionality, and yet here they go for all lower case letters, making it read as 'three did'.
• NECTAR: a database of codon-centric missense variant annotations — Another case of highly selective choices of letters to form the acronym. The full name for this database is the Non-synonymous Enriched Coding muTation ARchive.
• LoQAtE—Localization and Quantitation ATlas of the yeast proteomE. A new tool for multiparametric dissection of single-protein behavior in response to biological perturbations in yeast — Presumably pronounced 'locate', but if someone hears about the tool without seeing the spelling, then I doubt they will have much luck finding it online. Given the highly selective use of letters to form the name, they could have just called it 'LOCATE'.
• HoPaCI-DB: host-Pseudomonas and Coxiella interaction database — The 'Pa' comes from Pseudomonas aeruginosa, but the other species (Coxiella burnetii) only contributes a 'C'. More importantly, how do you pronounce this database name? Hoe-packy? Hoe-pacey? Hoppa-cee? Hoppa-sigh?

Honorable mention:

• canSAR: updated cancer research and drug discovery knowledgebase — This looks like it should be an acronym for something, but I can't find any details. It would make more sense to me if it was an acronym, otherwise it just looks like someone can't spell 'cancer'.

The three ages of CEGMA: thoughts on the slow burning popularity of some bioinformatics tools

The past

CEGMA is a bioinformatics tool that was originally developed to annotate a small set of genes in novel genome sequences that lack any annotation. The logic being that if you can at least annotate a small number of genes and have some confidence about their gene structure, you can then use them as a training set for an ab initio gene finder to go on and annotate the rest of the genome.

This tool was developed in 2005 and it took rejections from two different journals before the paper was finally published in 2007. We soon realized that the set of highly conserved eukaryotic genes that CEGMA used could also be adapted to assess the completeness of genome assemblies. Strictly speaking, we can use CEGMA to assess the completeness of the 'gene space' of genomes. Another publication followed in 2009, but CEGMA still didn't gain much attention.

CEGMA was then used as one of the assessment tools in the 2011 Assemblathon competition, and then again for the Assemblathon 2 contest. It's possible that these publications led to an explosion in the popularity of CEGMA. Alternatively, it may have become more popular just because more and more people have started to sequence genomes, and there is a growing need for tools to assess whether genome assemblies are any good.

The following graph shows the increase in citations to our two CEGMA papers since they were first published. I think it is unusual to see this sort of delayed growth in citations to a paper. The current citations from 2014 suggest that this year will see the citation count double compared to 2013.

The present

All of this is both good and bad. It is always good to see a bioinformatics tool actively being used, and it is always nice to have your papers cited. However, it's bad because the principal developer left our group many years ago and I have been left to support CEGMA without sufficient time or resources to do so. I will be honest and say that it can be a real pain to even get CEGMA installed (especially on some flavors of Linux). You need to separately install NCBI BLAST+, HMMER, geneid, and genewise, and you can't just use any version of these tools either. These installation problems have meant that I recently tried making it easier for people to submit jobs to us, which I run locally on their behalf.

These submission requests made me realize that many people are using CEGMA to assess the quality of transcriptomes as well as genomes. This is not something we ever advocated, but it seems to work. These submissions have also let me take a look at whether CEGMA is performing as expected with respect to the N50 lengths of the genomes/transcriptomes being assessed (I can't use NG50 which I would prefer, because I don't know the expected size for these genomes).

Generally speaking, if your genome contains longer sequences, then you have more chance of those sequences containing some of the 248 core genes that CEGMA searches for. This is not exactly rocket science, but I still find it surprising — not to mention worrying — that there are a lot of extremely low quality genome assemblies out there, which might not be useful for anything.

The future

We are currently preparing a grant that would hopefully allow us to give CEGMA a much needed overhaul. Currently we have insufficient resources to really support the current version of CEGMA, but we have many good ideas for how we could improve it. Most notably, we would want to make new sets of core genes based on modern resources such as the eggNOG database. The core genes that CEGMA uses were determined from an analysis of the KOGs database which is now over a decade old! A lot has changed in the world of genomics since then.

The problem that arises when Google Scholar indexes papers published to pre-print servers

The Assemblathon 2 paper, on which I was lead author, was ultimately published with the online journal Gigascience. However, like an increasing number of papers, it was first released to the arXiv.org pre-print server.

If you are a user of the very useful Google Scholar service and you have also published a paper such that it appears in two places, then you may have run into the same problems that I have. Namely, Google Scholar appears to only track citations to the first place where the paper was published.

It should be said that it is great that Google tracks citations to these pre-print articles at all, though see another post of mine that illustrates just how powerful (and somewhat creepy), Google Scholar's indexing power is. However, most people would expect that when a paper is formally published, that Google Scholar should track citations to that as well. Preferably separately from the pre-print version of the article.

For a long time with the Assemblathon 2 paper, Google Scholar only seemed to show citations to the pre-print version of the paper, even when I knew that others were citing the Gigascience version. So I contacted Google about this, and after a bit of a wait, I heard back from them:

Hi Keith,

It still get indexed though the information is not yet shown:

http://scholar.google.com/scholar?q=http%3A%2F%2Fwww.gigasciencejournal.com%2Fcontent%2F2%2F1%2F10+&btnG=

If one version (the arXiv one in this case) was discovered before the last major index update, the information for the other versions found after the major update would not appear before the next major update.

Their answer still raises some issues, and I'm waiting to hear back from my follow up question...how often does the index get updated? Checking Google Scholar today, it initially appears as if they are still only tracking the pre-print version of our paper:

However, after checking I see that 9 out of 10 of the most recent citations are all citing the Gigascience version of the paper. So in conclusion:

1. Google Scholar will start to track formal versions of a publication even after the paper was first published on a pre-print server.
2. Somewhat annoyingly, they do not separate out the citations and so one Google Scholar entry ends up tracking two versions of a paper.
3. The Google Scholar entry that is tracking the combined citations only lists the pre-print server in the 'Journal' name field; you have to check individual citations to see if they are citing the formal version of the publication.
4. Google Scholar has a 'major' indexing cycle and you may have to wait for the latest version of the index to be updated before you see any changes.