The Six Degrees of Functional Genomics

I enjoyed a whirlwind functional genomics talk by Steve Kay today about the circadian rhythm in a number of model organisms. Perhaps most interesting was his outline of the functional genomic approach. They were, paraphrased, as follows.

1) Identify the elements of the circadian rhythm machinery
2) Model and generate quantitative hypotheses about this machinery
3) Synthetically reproduce the machinery

He went on to describe how we are working on wrapping up step 1 and just beginning to dip our toes into step 2.

I should start by saying that I think the research described was truly impressive and brought to bear a large number of high throughput techniques to answer a question in a way that went beyond the usual model of figuring out what "your favorite gene" has to do with process X. He's turned things on their head and asked what does gene Y have to do with "my favorite process", and answered the question en masse.

However after he delved into the previous work on circadian rhythm I was left a little worried. In a process which, like the cell cycle machinery, transcription and translational control is so key, how can you avoid the fact that a vast swath of the cell's general control mechanisms for these fundamental processes will in some way also affect the circadian rhythm? At some point, any element of the cell that isn't completely inert is going to affect any process you can choose, albeit in perhaps a small way.

Which brings me to the Six Degrees of Functional Genomics. No part of the cell exists in isolation. No process in the cell can, really, be excised from the context of the larger cell. What we're talking about, after all, a tiny little bag of water packed chock full of different proteins. True, there are compartments, but these compartments have a nasty way of communicating with each other. In the end, every protein in the cell is functionally related to every other protein of the cell, given enough degrees of separation. Quantitative models are likely to look more like weather models than a nice damped oscillator. In the end, it's all very dependent on initial conditions and most predictions will be probabilistic in nature.

To try to develop quantitative models of cellular behavior given current knowledge may be like trying to model all of human social interaction by the Facebook "friends" network. It's heavily biased for certain kinds of interactions, like those between college dorm mates. It's true, these relationships are very important, and they tell you a lot about how the cell behaves day to day. But there's almost certainly a set of interactions that we don't know we don't know, like the interactions with parents and family (not everyone wants their parents seeing all those facebook photos).

To this extent it probably is important to try to develop the networks we currently have. We do still want to find key elements of our processes of interest. Hopefully we don't get so carried away looking for the next gene that we forget the complexity of cellular function try to build ever bigger piles of genes in our category of interest.

I'm aware that I am setting up a bit of a straw man, here, and I don't pretend that Dr. Kay is unaware of these kinds of concerns. But I think when the broader scientific community looks to functional genomics and computational biology for answers, they need to be aware of the fundamental limitations. Frankly I am not sure if the broader scientific community looks to functional genomics and computational genomics for much of anything, but that has its own problems.

Tool Time

To start my note off on a tangent, I want to recommend the "getting to work early" paradigm. Arriving to work before 7am finds very few souls clogging the arteries of the building, and very few distractions to sidetrack this grad student's taxed neural network. It is a time for settling in and thinking about the big picture. It is also a time for using all those adverbs that your PI has stricken from your science writing. Even the word "abrogate" gets boring if repeated endlessly like a sitcom laugh track.

So let me expound, veritably explicate, upon the following question: What are the bioinformatic tools that I wish I had for my research? The answer comes like a dam burst. There is simply too much material to stay above water.

• Multiple Reference Short Read Mapping: By this I mean a tool that corrals the multiplicity of reference human genomes and variant annotations and links them together for read mapping. This might sound silly, since most reference mappers can handle a SNP here or there. But there are a number of "everything that can go wrong will" sorts of scenarios, where a common SNP variant or two can lead to horrifically erroneous mapping. This propagates into very confusing results down the line. Such results have to be carefully untangled by hand before they reveal their fundamentally invalid core. With the tools we now have available in the human genome, multiple reference mapping is becoming a must have app. So if you're out there WashU, Broad, Sanger, BGI, hear my prayer.
• Base Quality Retouching: Like the photographs that grace the covers of the latest supermarket magazines, the output that spools off of an Illumina GA needs a little retouching. Sometimes it needs a lot of retouching. There are the issues of PCR duplicates, nucleotide chemistry and failed cycles. These are technical problems that may or may not go away any time soon. In the meantime, we need better base quality numbers. Is that set of 8 reads calling an A instead of G a SNP? Hard to say if you can't believe your base call qualities. MarkDuplicates (picard) and the GATK coming out of the Broad might have this problem mostly solved, but they remain to be packaged into a neat little bundle and handed out like candy to the rest of us.
• The Mapping Quality Problem: To anyone who has played with the high throughput sequencing technology should know about this problem. What does it mean that a read maps to a given location? Suppose it maps to one location perfectly, but 25 with one mismatch. Suppose instead that it had mapped to one location with one mismatch and only two with two mismatches. Which gets the better mapping quality? How are these situations even comparable? I have my own thoughts on a Bayesian way of handling this situation. Maybe just saying the word Bayesian is enough to conjure my solution, and maybe it's too naive to be useful in implementation. Regardless, we need an answer sooner rather than later, lest interesting loci perish for want of a good sequencing read to feed them.
• The SNP caller to end all SNP callers: This really does not deserve an explanation. (1) We want SNPs. (2) We want confidence scores for those SNPs that is remotely close to correct. The first part is done, we can call SNPs, but I'll be damned if I believe the kinds of confidence scores we assign to them. Getting this problem solved really requires getting the three problems above solved first.
• Structural Variants for the Rest of Us: The gsMapper has a nice little tool for calling structural variants. Of course, 454 reads are quite amenable to this kind of work due to their length. Paired end Illumina reads should be perfectly functional too, though. I have yet to see an easy to use structural variant caller whose results I can sink my teeth into. I've seen a number of ad hoc tools, and some very high level tools which are nearly impossible to use. To get this problem solved rightly we probably need the mapping quality problem solved first. 
• De PseudoNovo RefSembly: No I am not just trying to smash words together to sound smart. I bring this tool up because, in my ideal world, the tools are bountiful, the data overfloweth, and every grad student is above average. In this imagined world we also have this little gem of a tool for particular problems. Sometimes you have a reference. Sometimes you have multiple references. Sometimes you have some reads that map to the reference, and some reads that you think might represent some new genetic material. You'd like to map to the genome, but you'd also like to put together those delicious additional morsels into something that approximates a meal. For this you want Ref-Sembly, a tool that uses the a priori information from the genome you are working from but also openly allows and embraces the possibility of additional sequence. Such a tool should make a best guess at what such underlying sequence is and provide information about how that sequence might connect to the reference you've dutifully provided. Currently, I think people map reads to the genome and just cram the unmapped refuse into a de novo assembler. I'm not going to say that this is wrong, but, in my heart of hearts, I don't feel that it is fully right. Assembly off of a reference needs to be more nuanced than a garbage compactor. 
• A Visualization Suite that Doesn't Crash My Computer When I Try to Look at Tens of Thousands of Reads: Does my request defy the bounds of computer science? Is my measly 8 gigs of RAM insufficient for your hungry java app? All I know is this: there is currently no replacement for putting eyes on data. I can see an indel coming from a mile away if I can visualize my reads. IGV is my current tool of choice, but it craps out (for me) when the coverage gets deep. Unfortunately this is where I need the tool the most. Maybe the answer is that I should get some more sticks of RAM, but I have to imagine that the coverage is only going to get deeper, and the problem will continue to mount. 
• A Visualization Suite that Produces Poster-Ready Images: UCSC genome browser comes close. Very very close. But the difficulty in customizing the visualization and the the granularity of the images (with their horrific font) makes this a step down from my ideal. If only there was a "whimsical" button to enhance the graphic appeal of the data it already displays, then I think we would be there. If I am looking across 100 kb of sequence I need my exons to have a little more flair then a vertical line one pixel thick. My guess is that my hypothetical reader is now laughing that I didn't notice the "Visualize, with Feeling" button, tagged with the infamous 'blink' html, that sits dead center on the home page at genome.ucsc.edu. Maybe that person will email me.

That's it for now. I think I have exorcised the adverb demon that haunts my scientific writing. I return to the keyboard and the pipette knowing that my salvation is temporary, and that the thirst for flowery exposition shall rise again.

Presenting to Peers

So this is just a random comment, but I realized recently how infrequently a grad student gets the opportunity to present to his/her peers. Well actually I didn't realize it, a friend of mine in the program realized in and brought it to all of our attention. He offered to start off a set of presentations that we, as students, would give to each other.

This may sound a little ridiculous at first. You're thinking: "What on earth is he talking about? Poster sessions, journal clubs, lab meetings, maybe a departmental presentation or two aren't enough?" Well, yea, I guess those are quite considerable. Our peers form most of the audience for those presentations. But in those circumstances there is almost always an authority figure present. There's your PI or even other PIs and members of your thesis committee. There's a program director or maybe even a judge who is looking for the presentation that wins a prize.

And that is great. Don't get me wrong; we need that. But where is the opportunity for us to grow into independent scholars? Where is the opportunity for us to shape a presentation style that is designed to speak to an audience that understands the material at our level, and on our own terms? I've noticed that the talks that are most captivating at conferences are those that are delivered with a sense of familiarity and comfort with both the subject mater and the audience. They're often delivered with humor and the occasional hint of wry, self-deprecating humility. If we want to train not just scientists, but communicators, we should give them the chance to train their art in an uninhibited setting. I'm concerned that always reporting our results in front of people we need to impress may further entrench systems of jargon and insular academic perspective.

I'm not trying to badmouth lab meeting or committee meetings. Those are opportunities to expose our line of research to outside challenge, and sometimes even outside attack. We need that. We need to learn to think like scientists. That means constantly revisiting our own assumptions and our own familiarity with our discipline. But maybe every once in a while the big guys could step out of the room and us baby scientists could talk about what we do to each other. It might save a little adrenalin for another day, and it might foster lines of cooperation between students that could last through our careers.  And who knows, maybe we wouldn't have to sit through so many presentations that sound like someone reading through their alphabet soup.

Too many articles, too little time

How on earth do you drudge through the literature? I guess a good number of people read the abstracts and skim for interesting figures. A few others hone their area of interest into such a tiny corner that they can actually be completely up to date (in, say, VHL and HIF1alpha interactions in hypoxic conditions in HeLa cells under 10% FBS concentration, or whatever).

But where is the fun in that? As someone who got interested in science because it was just, so damn cool, I don't want to give up that connection to the broader picture. I'm not particularly interested in participating in the construction of my very own, brand new pigeon hole. I can see the wooden outlines now: using sequencing technology for subtyping of malignant melanoma in a clinical setting using gene list X. I will know everything about how to use sequencing technology Y to look at melanoma genes X in patient group Z. I mean I will be able to take the kids to school with my most scholarly, mind numbingly in depth knowledge of XYZ. But I will have forgotten the wider goal, the bigger picture. I will have specialized in the war of 1812 only to find that World War-Eleventy-Two is raging on without me.

Solutions? Google Reader sure seems like a good shot at trying to stay mildly up to date. Even there I get ruggedly behind. What I really need is a second brain that can sift through all the sludge to find those nuggets of thrilling wisdom. Anyone know where I can get one of those?

Paradigm 4

If there's one thing I hate, it's the monotony of generating data. I'm working on running sequencing on 48 samples right now, and let me tell you that is not a bag of laughs. I am all too aware that what I am doing could probably be done faster and better by a halfway decently designed machine, and consequently I just sit there daydreaming of a world in which I have that machine. This leads to pipetting errors which leads to more frustration which, as you can imagine, leads to more daydreaming. It's a cycle of violence on the microliter scale.

Which is all just an intro to explaining why I loved this article in the New York Times on the data deluge. Sequencing machines (among other things) are generating so much data that it's actually the analysis that becomes the limiting step. Eventually there will be so much data output that there will be little need for pipetters and a great need for analyzers. That's right, the limiting reagent is actually human brain hours.

Mind you not just any human brain hours will do. To understand this data we need people well versed in computer science. Computers, after all, are the only things capable of reading off the billions of bases involved in anything approaching reasonable time. A human reading of just a single human genome at one base per second (no breaks, no sleeping) would take over 95 years to complete.

Our limiting reagent brain needs to be versed in statistics, to allow for the fact that any comparisons made on the genomic scale and possibly between large populations. Signal is well hidden by noise, and the noise isn't even necessarily as random as we would like it to be. After all, this is a living, breathing genome we're talking about, not a string of A's, C's, G's and T's as we often imagine it.

Which brings us to the third necessity. The analytical brain that we need also, ideally, should have a strong understanding of molecular biology and the biology of any disease in question. Cells are complicated. Ridiculously complicated. But we do know a pretty enormous amount about how they operate. A thorough understanding of this prior knowledge helps us ask more pointed questions of the data in hand.

Have I overdetermined the system yet? Probably. In the end, deciphering this data is going to take a lot of collaboration. I've seen a lot of attempts at all-in-one prepackaged analysis engines for sequencing data. None of them, so far, looks very impressive. Moreover understanding the output of such packages is its own special challenge, since their inner workings is often closed source or poorly documented. Thus it's often hard to trust or interpret the results that you don't generate yourself.

So will this data flood answer the big questions of our age? Are we going to find the cure for cancer? Perhaps at least the cure for a cancer? If we do it will be not only because of our ability to design and execute good experiments, but also our creativity in sifting the results.

The Panopticon, Part II: Control

Having just opened a machine of revolutionary scientific power, you convene your graduate students to discuss the possibilities.  

Yes I know this is ridiculous, what PI respects and trusts his grad students enough to share this kind of information? He'd just go straight for his fellow PIs, right? Just roll with it, people.

Sitting down with your students you first carefully explain to them the circumstances of finding the machine and the mysterious booklet with its unbelievable claims (see Part I). Your students sit around and listen attentively and with increasing eagerness lean forward in their chairs. They've seen the strange device stowed away in the corner of the lab, and they know this is not one of your endless hypothetical. You end your explanation quickly and ask:

"So, what should we do with the machine? We have 25 uses and we'd better use them well!"

Abe: "Well clearly we should start analyzing people with Our Favorite Disease! We can look at 25 of them, which should give us a good sense of what's going on in OFD."


Ben: "What's going on in OFD? Do we even know what we're looking for?"

Abe: "Well first of all we can see if parasite X is present in OFD. I mean that theory has been kicking around for a long time now."

Charles: "But what if there are no parasites, or only a few parasites? Do you want to waste all of these experiments just looking for parasites?"

Abe: "Well that's the great thing about it! I mean if you really can look at everything, then we can answer a lot of questions at once. Like what about the theory that OFD patients have greater p123 signaling? We'll be able to count the p123 and answer that one right off the bat, too."

Ben: "Hold on bucko, what are you saying? If we just look at 25 OFD patients we won't have any idea whether p123 is high or low. We'll just know what it is on average in those patients, we have no basis for comparison."

Charles: "Yea Abe, slow down. We need to think through this. What are the proper controls?"

Abe: "Proper controls? Look we only have 25 runs of this thing, we need to be focusing on interesting samples, not normal everyday people. What if we don't look at enough sick patients and miss something important?"

Charles: "We can't do this without controls. There's just no way. You have to be able to compare the patients to some estimate of what's normal, and there's no other way to know what's normal without using some runs of the Panopticon. I mean, we could use previous estimates of p123 or parasite prevalence in the general population, but there's no way that we can really believe that those are accurate."

Ben: "Yea, I remember hearing that p123 may have three or more isoforms that have been undetected in our blotting assay. The Panopticon is powerful enough to see those."

Delta: "It will see those... (Mysteriously) but what else will it see?"

Ben: "What do you mean, Delta?"

Delta: "What else will it see, under the surface? It's true we may see p123 isoforms we expect, but what about those we don't expect?"

Ben: "(Condescending) Well we'll look at those, too. Now Abe do you see why we need to run some normal patients through the machine, too?"

Abe: "Yea I guess, but I think we should do as few as possible."

Ben: "Well yea, I mean you only need to run a few controls."

Charles: "Do you guys just completely not get it? We're not running the kind of control where we know what to expect. This isn't like running a PCR with water instead of DNA template. We don't just need to know what's normal, we need to know the variability of normal. Sure, we might put two people in and they might not have parasites, but what if the third one would have? If we see parasites in half of our patients we still won't know if that's normal for the general population?"

Abe: (Sighs) "Well what do you recommend? We can't waste all of these runs."

Charles: "Split it down the middle. Half the runs, or I guess 13 if it makes you happier, could be patients, 12 could be normal people."

Abe: "I'd say it was a shame, but I guess we can still answer so many different questions."

Delta: "The more questions you ask, the more will slip through your fingers."

Ben: "Well I know that doesn't make any sense. The more questions the merrier."

Delta: "Seeing everything is like seeing nothing. It is a true Panopticon. The original Panopticon was a tower in the middle of a prison. Each cell faced the center, and the guards could see all cells from their vantage point. So will you be within the Panopticon. You see all, but in this sight you become imprisoned. Just as you can view each cell, you will find you can see none of them."

Abe: (Laughing) "So speaks the oracle! Did that make any sense?"

Ben: "Not that I can tell"

Charles: "No"

Delta: "Go ahead then."

Abe: "I don't know what Delta is talking about, but let's just run 20 people, 10 healthy, 10 sick. We'll have 5 left over just in case something goes wrong."

You decide to allow your graduate students to proceed. They agree to try the machine on ten patients and ten healthy people, analyze the data. You are pleased that they've come to the right conclusion, and put together a case-control design. You worry, though, about Delta's ominous prophecy for this experiment. Perhaps you will find out what she means when the data comes in...

Continued later.

On purposeful learning

I was cleaning out my Onenote files the other day and came across a poem by WB Yeats I liked and saved:

"What Then?"

His chosen comrades thought at school
He must grow a famous man;
He thought the same and lived by rule,
All his twenties crammed with toil;
`What then?' sang Plato's ghost. `What then?'

Everything he wrote was read,
After certain years he won
Sufficient money for his need,
Friends that have been friends indeed;
`What then?' sang Plato's ghost. `What then?'

All his happier dreams came true -
A small old house, wife, daughter, son,
Grounds where plum and cabbage grew,
Poets and Wits about him drew;
`What then?' sang Plato's ghost. `What then?'

`The work is done,' grown old he thought,
`According to my boyish plan;
Let the fools rage, I swerved in naught,
Something to perfection brought';
But louder sang that ghost, `What then?'

I mean to post a science-based post soon but this has been a continuation of conversations we've had in the past semester. What is the end goal of accruing knowledge for all of you? How will we use it? Or will we end up spinning our wheels with minutiae discoveries only for the sake of accruing grants we require for personal survival?

A Risky Statement on Gender

Before I get myself into too much trouble, I want to mention that N=2.

So I am sitting in my bioinformatics class during presentations and I am noticing an interesting phenomenon that I can't help but post about. Each student is supposed to present, but students can form teams.

Here's the result: The two girls in the group are part of a two person team. There are no two person teams with two males (of five). In the teams with a male and a female, the female begins the presentation and the male jumps in about 2 minutes into the 15 min presentation and (loudly) completes the remainder of the presentation. The female sits patiently for the balance of the presentation, but the male never hands the torch back.


My tentative conclusion, the barriers to women in science are often subtle, and are not only institutional.