Understanding CEGMA output: complete vs partial

On Friday I posted a reply to a thread on SEQanswers about CEGMA. I thought I'd include a modified version of that response here as it is an issue that gets raised fairly frequently. It concerns the 'complete' and 'partial' results that CEGMA includes in the final output file that it generates (typically called 'output.completeness_report'). Here were the two questions that were posted:

1) If a partial score is higher than a complete score then does this indicate that the assembly is fragmented?

2) Also, should the partial score be lower than the complete score in an ideal situation?

Remember, these are not scores per se. Both of these figures describe a number of core eukaryotic genes (CEGs) that the CEGMA pipeline predicts to be present in the input assembly file. The 'complete' set  refers to those gene predictions which CEGMA classes as 'full-length'. Note that even if CEGMA says something is 'complete' there is still the possibility that parts of the protein is missing.

This is because CEGMA is taking each CEG that it has predicted and aligns the protein sequence of that CEG to the HMM profile generated from the corresponding core gene family (made up of six proteins from Schizosacchromyces pombe, Saccharomyces cerevisiae, Caenorhabditis elegans, Drosophila melanogaster, Arabidopsis thaliana, and Homo sapiens). As I recall from memory, if the alignment spans more than 70% of the protein profile the CEG is considered to be 'complete'. This 70% threshold is an arbitrary cut-off, but seems to work well in finding genuine orthologs of CEGs.

Somewhat confusingly, although we consider 'partial' matches to be those below 70% (but above some unspecified minimum score), the output in output.completeness_report uses 'partial' to include both 'complete' and 'partial' matches. So the number of partial matches will always be at least as high as the number of complete matches.

You should look at both results. If you don't have 248 core genes 'completely' present, the next thing is look at how many additional partial matches there are. If you have a result like 200/240 (i.e. 200 complete CEGs and 40 additional partial matches) then this at least suggests that most of the core gene set is present in your assembly, but some may be split across contigs or missing from the assembly. Remember, CEGMA only looks for genes that are located inside individual contigs or scaffolds. Theoretically, you could have an assembly that splits every gene across contigs which might lead to a 'complete' result of zero, and a partial result of '248'.

From looking at results of many different runs of CEGMA, it is common to see something like 90–95% of core gene present in the 'complete' category, and another 1–5% present as partial genes (for good assemblies at least). I have also seen one case where the results were 157/223. This is more unusual, suggesting that a relatively large number (27%) of the core genes were present as fragments. This might simply reflect lots of short contigs/scaffolds in the assembly. In contrast to this, one of the best results that I have seen is 245/248. It is rare to see all core genes present, even when you allow for partial matches.

Below is a chart that shows the results from 50 runs of CEGMA against different assemblies. The x-axis shows the percentage of 248 CEGs that were completely present, and the y-axis shows the percentage of CEGs that were only partially present.

Is yours bigger than mine? Big data revisited

Google Scholar lists 2,090 publications that contain the phrase 'big data' in their title. And that's just from the first 9 months of 2014! The titles of these articles reflect the interest/concern/fear in this increasingly popular topic:

One paper, Managing Big Data for Scientific Visualization, starts out by identifying a common challenge of working with 'big data':

Many areas of endeavor have problems with big data…while engineering and scientific visualization have also faced the problem for some time, solutions are less well developed, and common techniques are less well understood

They then go on to discuss some of the problems of storing 'big data', one of which is listed as:

Data too big for local disk — clearly, not only do some of these data objects not fit in main memory, but they do not even fit on local disk on most workstations. In fact, the largest CFD study of which we are aware is 650 gigabytes, which would not fit on centralized storage at most installations!

Wait, what!?! 650 GB is too large for storage? Oh yes, that's right. I forgot to mention that this paper is from 1997. My point is that 'big data' has been a problem for some time now and will no doubt continue to be a problem.

I understand that having a simple, user-friendly, label like 'big data' helps with the discussion, but it remains such an ambiguous, and highly relative term. It's relative because whether you deem something to be 'big data' or not might depend heavily on the size of your storage media and/or the speed of your networking infrastructure. It's also relative in terms of your field of study; a typical set of 'big data' in astrophysics might be much bigger than a typical set of 'big data' in genomics.

Maybe it would help to use big dataTM when talking about any data that you like to think of as big, and then use BIG data for those situations where your future data acquisition plans cause your sys admin to have sleepless nights.

The problem with posters at academic conferences

I recently attended the Genome Science: Biology, Technology, and Bioinformatics meeting in the UK, where I presented a poster. As I was walking around, looking at other people's posters, I was reminded of the common problem that occurs with many academic posters. Here are some pseudo-anonomous examples to show what I mean (click images to enlarge):

The problem here is not with the total amount of text — though that can sometimes be an issue — but with the width of the text. These posters are 84 cm (33 inches) wide, and it is not ideal to create text blocks that span the entire width of the poster. The reasons behind this are the same reasons why you never see newspapers display text like this…we are not very good at reading information in this manner.

To quote from Lynch & Horton's Web Style Guide; specifically the section on Page Width and Line Length:

The ideal line length for text layout is based on the physiology of the human eye. The area of the retina used for tasks requiring high visual acuity is called the macula. The macula is small, typically less than 15 percent of the area of the retina. At normal reading distances the arc of the visual field covered by the macula is only a few inches wide—about the width of a well-designed column of text, or about twelve words per line. Research shows that reading slows as line lengths begin to exceed the ideal width, because the reader then needs to use the muscles of the eye or neck to track from the end of one line to the beginning of the next line. If the eye must traverse great distances on a page, the reader must hunt for the beginning of the next line.

In contrast to the above examples, there were a couple of posters at the #UKGS2014 meeting that I thought were beautifully displayed. Bright, colorful, clearly laid out, not too much text, and good use of big fonts. Congratulations to Warry Owen et al. and Karim Gharbi et al. for your poster presentation prowess!

When is a citation not a citation?

Today I received a notification from Google Scholar that one of my papers had been cited. I often have a quick look at such papers to see how our work is being referenced. The article in question was from the Proceedings of the 3rd Annual Symposium on Biological Data Visualization: Data Analysis and Redesign Contests:

FixingTIM: interactive exploration of sequence and structural data to identify functional mutations in protein families

The paper describes a tool that helps "identify protein mutations across a family of structural models and to help discover the effect of these mutations on protein function". I was a bit surprised by this because this isn't a topic that I've published on. So I looked to see what paper of mine was being cited and how it was being cited. Here is the relevant sentence from the background section of the paper:

To improve the exploration process, many efforts have been made, from folding the sequences through classification [1,2], to tools for 3D view exploration [3] and to web-based applications which present large amounts of information to the users [4].

Citation number 2 is the paper on which I am a co-author:

  • Chen N, Harris TW, Antoshechkin I, Bastiani C, Bieri T, Blasiar D, Bradnam K, Canaran P, Chan J, Chen C, Chen WJ, Cunningham F, Davis P, Kenny E, Kishore R, Lawson D, Lee R, Muller H, Nakamura C, Pai S, Ozersky P, Petcherski A, Rogers A, Sabo A, Schwarz EM, Van Auken K, Wang Q, Durbin R, Spieth J, Sternberg PW, Stein LD: Wormbase: A comprehensive data resource for Caenorhabditis biology and genomics. Nucleic Acids Res 2005, 33(1):383-389.

The cited paper simply describes the WormBase database and includes only a passing reference to the fact that WormBase contains some links to protein structures (when known), but that's about it. The WormBase paper doesn't mention 'folding' or 'classification' anywhere, which makes it seem a really odd choice of paper to be cited. It makes me wonder how many other papers end up gaining seemingly spurious citations like this one.