So I’ve been doing a lot of correlation analyses lately. Right now they’re exploratory. I get some insight into what may be going on, it guides work that we’re doing, that kind of thing.
What if want to know if these correlations are actually meaningful? I’m not talking cause-and-effect meaningful, just “Are they consistent enough that we could base some other work off of them?”
Fortunately, there’s a post for that.
One of the interesting studies linked to from that post by Hui Xiang Chua is thisÂ study from Schonbrodt and Perugini that finds that in most cases your n should be 250+ to have stable correlations (that is, correlations that you wouldn’t expect to change if you sampled a different 250 cases from the same population).
This article popped up in some feed of mine in the past week, and it’s a quick, fascinating take on the octopus as our best on-earth example of an alien intelligence. It’s a fast read, and makes a compelling case for why octopi, from whom we diverged a long time ago, display signs of consciousness and (thus) are good alien analogs.
Pretty neat both from the “Hey, octopi are cool” perspective and if you’re an SF author interested in writing plausible alien intelligence.
The article by Olivia Goldhill is here.
Perhaps unsurprisingly, one of my interests is functional, accurate protein annotations. The default way to annotate new sequences, especially in a high-throughput manner, is to use sequence identity with some form of BLAST and use the best hit to annotate your sequence of interest.
There are some limitations here. It’s been shown that enzyme function is not necessarily conserved even with fairly similar sequences. We’ve demonstrated that each orphan enzyme we find a sequence for can lead to the re-annotation of hundreds of genomes.
In their paper DomSign: a top-down annotation pipeline to enlarge enzyme space in the protein universe, Wang et al from Tokyo Institute of Technology and Tsinghua University apply a protein-domain based approach to try and expand our ability to predict enzyme activities for proteins.
Things have changed quite a bit in the last decade and especially the last five years in scientific publishing. Starting with open access efforts (notably the PLOS journals, which launched when I was in grad school and have driven a revolution in publishing) and continuing in venues like Retraction Watch and PubPeer, the move toward making science open, clear, honest, and above all accurate is powerful and completely unlike when I started doing research.
Retraction Watch put up a guest post today by Drummond Rennie and C.K. Gunsalus titled If you think it’s rude to ask to look at your co-author’s data, you’re not doing science. In it they talk about how a couple recent and slightly less recent high-profile scientific frauds have broken down, and the failure of senior authors to actually do their part to validate the data and really know what’s going on. They also provide a truly helpful breakdown of approaches to make sure everyone knows what work is being done, that the work is actually being done (and funded, and so forth), and that everyone is credited properly.
If you’re doing, well, science, or any kind of collaboration at all, I recommend this genuinely helpful piece.
In our work on orphan enzymes, we’ve consistently seen a “rich get richer” effect. Research tends to accumulate on those proteins that already have assigned sequences. This is a systematic issue, since annotation based on sequence similarity probably means that we’re often assuming that a newly identified gene does the same function as a known protein…when in reality, it is more like a highly similar orphan enzyme for which we lack sequence data.
We saw this occasionally in the generally awesome BRENDA enzyme database. A curator had assigned a sequence to an orphan enzyme when that sequence was actually for a highly similar enzyme that did not catalyze the orphan enzyme activity. This kind of over-assignment likely prevents further research on the orphan enzyme and tends to focus more research on the enzyme for which we had sequence data in the first place.
Cracking the brain’s “ignorome”
In their recent paper Functionally Enigmatic Genes: A Case Study of the Brain Ignorome, Pandey and colleagues tackle this problem from the other side of the mirror – uncharacterized genes.
They surveyed those genes that show “intense and highly selective” expression in the brain (ISE genes, for short) and asked “How well-characterized are they?” After all, one of the promises of modern high-throughput methods is that we can look at features such as tissue-specific expression and use that as a guide for which genes to devote our research attention to.
What they found is that despite our knowing that these genes are all intensely and selectively expressed in the brain, research about them has been tremendously lopsided.
Iâ€™ll quote them on just how off-kilter the research distribution is:
The number of publications for these 650 ISE genes is highly skewed (Figure 1). The top 5% account for ~68% of the relevant literature whereas the bottom 50% of genes account for only 1% of the literature.
Hereâ€™s Figure 1:
So that shows us that despite somehow being specific and important to the brain, many of these genes remain understudied.
What makes the ignorome different?
The short answer is â€œage.â€
Much like the “rich get richer” phenomenon I talked about for orphan enzymes, there is (unsurprisingly) a correlation between when a gene was first characterized and how much research there has been on it. Nothing else really differs between the genes that are understudied and those that have been the focus of significant study.
That brings up the natural corollary question of â€œOkay, so are we figuring out what the other genes do?â€
The answer here seems to be that we were for a while, but now the rate of advancing discovery is flattening out. Iâ€™ll quote the authors here as well:
While the average rate of decrease was rapid between 1991 and 2000 (âˆ’25 genes/year), the rate has been lethargic over the past five years (âˆ’6.4 genes/yr, Figure 5). This trend is surprising given the sharp increase in the rate of addition to the neuroscience literature. As a result, the number of neuroscience articles associated with the elimination of a single ignorome gene has gone up by a factor of three between 1991 and 2012 (Figure 5). The rate at which the ignorome is shrinking is approaching an asymptote, and without focused effort to functionally annotate the ignorome, it will likely make up 40â€“50 functionally important genes for more than a decade.
So what do we do about it?
One of the core reasons for â€œrich get richerâ€ effects is that known genes (or proteins) simply have more â€œhandlesâ€ you can work with. If your expression analysis tells you that 20 genes are significantly enriched in your test condition and you can find some functional characterization for 10 of them, itâ€™s only natural to focus on those 10 first.
â€¦and given how time and work tend to play out, â€œfirstâ€ can quickly become â€œonly.â€ Given how daunting a completely uncharacterized gene can be, who would fault researchers for spending the majority of their effort on those genes that have some functional characterizations (or predictions) available for them? That certainly fits the whole 80/20 rule idea of focusing most of your effort where youâ€™ll have the most gain.
Pandey and colleagues attempt to address this by making more handles. They show how we can leverage high-throughput and large-scale phenotype databases to generate additional functional characterization for at least some of the ignorome genes without significant additional effort. Now, instead of flying relatively blind, a researcher can have both sequence-similarity-based predictions of function and some best guesses at phenotype associations for these genes.
I really like this kind of leveraging of existing data to make avenues of research more accessible and thus more likely. This kind of thing is going to be very important in tackling those dark areas of unknown function that exist all over biology.
Who did this research
Ashutosh K. Pandey, Lu Lu, Xusheng Wang, Ramin Homayouni, and Robert W. Williams.
(â€¦and hey, Robert Williams is another UC alum!)
The full citation:
Pandey AK, Lu L, Wang X, Homayouni R, Williams RW (2014) Functionally Enigmatic Genes: A Case Study of the Brain Ignorome. PLoS ONE 9(2): e88889. doi:10.1371/journal.pone.0088889
Figure and quotes were used under the Creative Commons Attribution License.
We celebrated the end of 2013 with the release of our new paper, Rapid identification of sequences for orphan enzymes to power accurate protein annotation in PLOS ONE.
So what’s the big deal? What are orphan enzymes and why do we need to identify them?
Sequences are card catalog numbers for everything
In modern biology, protein and nucleotide sequence data are the glue that hold everything together. When we sequence a new genome, for example, we make a “best guess” for what each gene does by comparing its sequence to a vast collection of sequences we already have. Essentially, that lets us go from this amino acid sequence:
1 MSLPLKTIVH LVKPFACTAR FSARYPIHVI VVAVLLSAAA YLSVTQSYLN
51 EWKLDSNQYS TYLSIKPDEL FEKCTHYYRS PVSDTWKLLS SKEAADIYTP
101 FHYYLSTISF QSKDNSTTLP SLDDVIYSVD HTRYLLSEEP KIPTELVSEN
151 GTKWRLRNNS NFILDLHNIY RNMVKQFSNK TSEFDQFDLF IILAAYLTLF
201 YTLCCLFNDM RKIGSKFWLS FSALSNSACA LYLSLYTTHS LLKKPASLLS
251 LVIGLPFIVV IIGFKHKVRL AAFSLQKFHR ISIDKKITVS NIIYEAMFQE
301 GAYLIRDYLF YISSFIGCAI YARHLPGLVN FCILSTFMLV FDLLLSATFY
351 SAILSMKLEI NIIHRSTVIR QTLEEDGVVP TTADIIYKDE TASEPHFLRS
401 NVAIILGKAS VIGLLLLINL YVFTDKLNAT ILNTVYFDST IYSLPNFINY
451 KDIGNLSNQV IISVLPKQYY TPLKKYHQIE DSVLLIIDSV SNAIRDQFIS
501 KLLFFAFAVS ISINVYLLNA AKIHTGYMNF QPQSNKIDDL VVQQKSATIE
551 FSETRSMPAS SGLETPVTAK DIIISEEIQN NECVYALSSQ DEPIRPLSNL
601 VELMEKEQLK NMNNTEVSNL VVNGKLPLYS LEKKLEDTTR AVLVRRKALS
651 TLAESPILVS EKLPFRNYDY DRVFGACCEN VIGYMPIPVG VIGPLIIDGT
701 SYHIPMATTE GCLVASAMRG CKAINAGGGA TTVLTKDGMT RGPVVRFPTL
751 IRSGACKIWL DSEEGQNSIK KAFNSTSRFA RLQHIQTCLA GDLLFMRFRT
801 TTGDAMGMNM ISKGVEYSLK QMVEEYGWED MEVVSVSGNY CTDKKPAAIN
851 WIEGRGKSVV AEATIPGDVV KSVLKSDVSA LVELNISKNL VGSAMAGSVG
901 GFNAHAANLV TALFLALGQD PAQNVESSNC ITLMKEVDGD LRISVSMPSI
951 EVGTIGGGTV LEPQGAMLDL LGVRGPHPTE PGANARQLAR IIACAVLAGE
1001 LSLCSALAAG HLVQSHMTHN RKTNKANELP QPSNKGPPCK TSALL*
…to predicting that this protein is probably an “HMG-CoA Reductase,” an enzyme that carries out a key step in cholesterol synthesis.
We can also get more specific, tying part of this sequence information to the specific activity of the protein. In the case of my example enzyme, the “business end” is the second half of the protein.
This kind of sequence data powers so much of what we do in modern biology, from guessing what individual proteins do all the way to generating entire metabolic models and then predicting literally every food source a microbe can grow on.
We’re missing a lot of sequences
Hundreds upon hundreds, in fact. For a lot of critical enzymes.
As part of our Orphan Enzymes Project, we’ve tried to figure out how we can find sequences for these hundreds of enzymes.
After all, each enzyme represents hundreds of thousands of dollars in lost research…and each enzyme sequence we don’t have undercuts the value of all of our fantastic sequence-based tools.
We can rapidly identify a lot of orphan enzymes
Our new paper describes a few case studies on how we can identify orphan enzymes in the lab and just how big an impact identifying sequence for each orphan enzyme has.
We found several cases where we were actually able to buy samples of enzymes that had never been sequenced. We were also able to collaborate with Charles Waechter and Jeffrey Rush of the University of Kentucky to find sequence data for an enzyme they’d been working hard to characterize.
The key point of this part of our work is that many enzymes that are “tricky” for one set of researchers to sequence may be entirely doable for another group that specializes in sequencing. The more we collaborate, the more value we get out of all of our work.
Identifying orphan enzymes has a big impact
The second part of our work asks the simple question, “Does it matter?”
For each enzyme for which we found sequence data, we asked “How many enzymes should we now re-annotate?”
In other words, for all those guesses that have been made about what proteins do, for how many is our enzyme the best guess based on closeness of its sequence to the one we found.
It turns out that each enzyme sequence we identified led to anywhere from 130 to 430 proteins getting new, better guesses about their functions.
That’s hundreds of potential incorrect predictions or misled researchers averted by just “finishing the job” of sequencing a handful of enzymes.
Given the tremendous amount of work that has gone into characterizing each of these enzymes, it’s essential that we take every opportunity to apply modern sequencing expertise to existing samples.
Comments on the paper are welcome, whether here or on the paper itself at PLOS ONE.
Randy Schekman, James Rothman, and Thomas SÃ¼dhof were just awarded the 2013 Nobel Prize in Physiology or Medicine for their discovery of the how the cell transports materials around. Theirs were fundamental discoveries that play into every aspect of cell biology, including fun things like “How neurotransmitters get into and out of nerves.”
Randy taught part of my undergraduate molecular and cell bio course way back in the late 90s (one of the other instructors was Nicholas Cozzarelli – I was blessed with some excellent teachers).
My grad school lab (the Hampton lab) had a close working relationship with Schekman’s group, as his ongoing discoveries about how molecules and materials move within the cell directly tied into our own research on protein degradation (gotta move all those protein parts somehow…).
Given the nature of Randy’s discoveries and the quality of his work, we all thought it was only a matter of time before we was awarded the Nobel. It’s awesome to see it finally happen.
You can read the Nobel press release here.
You can read Randy’s charming phone interview here. It’s wonderful that he received the Nobel while his dad is still alive.
A lawn of E. coli stained with Coomassie Brilliant Blue dye.
Last year I found myself needing to visualize growth of a relatively thin lawn of E. coli on imperfectly translucent minimal medium plates. It was part of testing growth based on our predictions in the recently published Computing minimal nutrient sets from metabolic networks via linear constraint solving (I’ll have a separate post about that soon). Trying to get the lawns to stand out via backlighting or dark backgrounds didn’t do the trick, but staining the cells finally gave me the lovely picture you see above.
The protocol I used came from this page at the Center for Polymer Studies at Boston University.
Protocol: How to coomassie stain a bacterial plate
First, make staining solution.
1 liter staining solution:
1) To 400 mL distilled water, add 500 mL methanol
2) Add 100 mL acetic acid
3) Add 1 gram 0.1% Coomassie Brilliant R stain powder
4) Mix until the solution is a uniform blue
1 liter rinse solution:
1) To 400 mL distilled water, add 500 mL methanol
2) Add 100 mL acetic acid
(In other words, the staining solution minus the stain.)
To stain your bugs:
1) Pour the staining solution onto the plate so that it just covers the surface of the agar.
2) Let stand for 45 seconds.
3) Pour off the solution.
4) Pour on the rinse solution.
5) Swirl it for 10 seconds, then let stands for another 50 seconds.
6) Pour off the rinse solution.
It’s easy, and the resulting plates are quite pretty.
One of the most unnerving aspects of biological research is the possibility that your samples aren’t what you think they are. Most of my lab work has involved yeast (cerevisiae) and a smattering of types of bacteria (largely coli and some cyanobacteria). In general, we didn’t maintain the cells by passaging them, and there were some obvious antibiotic and auxotrophic (nutrition-based) markers we could use to tell that the cells were basically what we thought they were.
But as I’ve learned since entering the exciting world of genome analyses, there is just a ton of variation between the “same” organism and strain in different labs…or in the same lab at different times. The “default” E. coli strain, K-12 MG1655, has a neat little mutation in amino acid biosynthesis that easily reverts to wild type, which plays all sorts of havoc with computational models that assume that it’s nonfunctional.
I’m quite interested in how we can account for these kinds of differences and make modeling and predictive tools that are resilient to them.
In their recent paper Hiding in plain view: Genetic profiling reveals decades old cross-contamination of bladder cancer cell line KU7 with HeLa, Jager et al applied a very basic kind of DNA profiling to many samples of a popular and widely used bladder cancer cell line. These are cells that were supposed to have been derived from a fairly mild bladder cancer sampled from a patient in 1980. They’ve been widely used since then to study and model bladder cancer.
Except it turns out that they’re not bladder cancer cells. As Jager and his colleagues discovered, basically all the KU7 cell lines in the world are actually a completely different kind of cell (the most common cancer cell line in the world, HeLa). This apparently started with cross-contamination back at the source.
So what does this mean for studies based on those cells? Presumably we’d want to have a way to mass-tag those publications and all the databases or other informatics resources derived from them with the true identity of the cells used. Is this reasonably achievable, and is there a good way to track areas where the ideas or conclusions drawn from experiments using these misidentified cells ended up?
I’m not especially familiar with the bioinformatics and quantitative bio of cancer biology, so I don’t know how much impact this specific discovery has on large-scale data resources those fields rely on. Presumably this kind of thing is going to keep happening – we’ve certainly seen it in the misidentification and renaming of microbial samples from which enzyme and other metabolic data were derived. It would be handy to have consistent mechanisms in place to add additional metadata to publications so that this kind of “switch” can be tracked and propagated into downstream resources.
There’s more discussion of this discovery and its consequences for publications that used the misidentified cells over at Retraction Watch.