It’s that time of year again, folks: Bioinformatics Christmas. Or is it rather Thanksgiving, where the whole family gets together whether we like it or not?
Either way, I’ll be there next week, bells on and business cards in hand. I’d love to catch up with you, if you’ll be there. Send an email or find me on the contact page if you want to get a coffee and chat.
I’m very much looking forward to Carl Zimmer’s talk, as well as some talks by colleagues and friends of mine: John Keilty and Karina Chmielewski from Third Rock, Mike Dinsmore from Editas, and Iain McFaydden from Moderna.
Who do I really want to meet this year? Tanya Cashorali, one of the plenary keynote panelists and another woman who has started her own data science company. There aren’t many of those. We have to stick together.
The world of quantitative biology is large, diffuse and sometimes overwhelming. It’s hard sometimes to even figure out what someone means when they say “bioinformatics”. This can make it hard to figure out what part of the field someone works in.
One way to break it down is to describe bioinformatics as the building of tools and methods for the processing and management of biological data, and computational biology as the pursuit of biological sciences using computational methods. Therefore, bioinformatics is more of an engineering discipline and computational biology more a scientific discipline.
It’s helpful to think about these distinctions, subtle as they seem. It takes a certain mindset and skillset to build a robust sequencing analysis pipeline that will serve the needs of a large group of scientists for years. That mindset and skillset may be very different from the one required to do a deep investigation of the variants that impact risk of heart disease.
We can argue about the naming conventions all we want, but the label we apply to these two types of specialist doesn’t really matter. What matters is what they do; the person I would call a computational biologist writes code, yes, but does it in pursuit of a particular biological problem, and they would love to write less code and more manuscripts. The bioinformatician, on the other hand, wants to spend their time writing robust, high-quality code that does interesting and powerful computations. Papers are more of a nice side-effect.
The truth of the matter is that most programming biologists are a mix of the two disciplines.
When hiring for a small department or a startup, the distinction between these two caricatures becomes very important. Some people will be in the field for the biology specifically, and will choke when pressed to develop a tool for use by a team. Others will jump at the chance to write such a thing. Every group needs both of these. Consider the current needs; will this person be building a pipeline that will be re-used again and again? Or will they investigate particular variants, or particular compound response profiles? Fitting the right person to the job will ensure a happy employee and high productivity.
Figuring out what kind of background and preferences someone has can be as simple as asking them. Their resume or LinkedIn profile can also give clues. A software-focused person will tend to have one or more large, open-source bioinformatics software tools prominently listed. Their reference list may include a few papers describing this project and others (potentially many others) that use that tool. A manuscript-focused person will not be as likely to have a major tool-building segment of their resume. Instead, they will list a series of biology or dataset-focused projects, with manuscripts describing each.
Data Science
But where does data science fit into all this? That, at least, is simple; bioinformatics/computational biology is data science with a biology application, just as computational chemistry is data science for chemistry. Physicists have figured out that they’re all data scientists already, so there is no need for a name for them beyond “physicist”. I hope in the future we’ll do the same and just call ourselves “biologists”.
Most of us in the biotech space are fully onboard with the concept of reproducible research. Who thinks it’s a bad idea to be able to trace back from our results to the data and methods that produced them?
The trick, of course, is in how to do it. In the lab, there’s a strong culture of recording our experiments, tracking samples in LIMS systems and lab notebooks. However, in computational research the processes are often more personalized, or simply aren’t done.
And yet we’ve all experienced the confusion of looking back at a collection of analyses we’ve done, and wondering which was the “real” result, or why the results of one analysis differed from the other. What about that slide deck we showed to management – which of the many versions of the RNA-Seq pipeline did it come from? Why does the code result in a different result than it did the last time I ran it?
I had this problem with my dissertation. My graduate work was done entirely in R, and included nearly 100 figures and tables. Some of these figures were very similar – the same analysis was done on several different datasets. Keeping track of the origin of each figure was extremely difficult; I couldn’t convince myself I had it right.
Two things helped with this: version control and self-documenting code. I’ve written about version control before. It solved the problem of figuring out why the results might change from one day to the next. Today I’ll tell you about why self-documenting code is a critical component of a data analyst’s workflow.
My entire thesis was self-documented. It was written in Sweave, which is a documentation integration package for R and LaTeX. I embedded all of the code from my work into the text document that described it. When I made a change to the text, I would rebuild the whole thing and see a pdf emerge with all of my prose, figures, and tables, all beautifully formatted. Since the code that produced any given figure was located right next to the text that described it, I was fully confident I knew exactly which analysis produced which plot. Since the whole document was checked into version control (I used Subversion at the time), I knew I could track back any changes to the plot to the changes in the analysis that produced it. It may sound like doing this sort of tracking and documentation is a lot of work, and I won’t pretend that there isn’t overhead. But this extra work paid off profoundly one day.
A good four months into the thesis-writing process one of the public datasets I had analyzed was retracted; there were irregularities in the authors’ data curation process. I was mortified – what did this mean for my work in integrating these datasets, and how was I going to extract all of the plots and tabular results that included the disgraced dataset?
Since my thesis was self-documenting, I didn’t need to worry about how the extraction of one dataset would mangle the organization of the document. It was structured like a programmer had written it – with for loops and lists of datasets. I deleted the offending dataset and rebuilt the thesis. My figures were recreated, my tables recalculated. I did need to edit a few things by hand; for example, any mention of that disgraced dataset in the text needed to be changed. But overall, reproducible research had saved me. I could have spent a month rewriting and instead I spent a few days.
Not bad.
I can’t recommend self-documenting code highly enough. It’s helped me on many occasions since, in smaller analyses and in less-spectacular ways, but I no longer even think about doing work for myself or for a client without self-documentation.
Coda
When I wrote my thesis Sweave was pretty state-of-the art and Subversion was not a complete dinosaur. Now, there are better options available, like Markdown and Knitr, or the open lab notebooks put together by the Jupyter folks. Any one of them will make your analysis infinitely more reliable. For version control, Git is the most common choice, but any of them will give you the confidence you need that your code is what you think it is.
I’ve been asked before to explain how repeated measurements can impact statistical models. We often bake repeated measurements – more than one measurement take on the same person, the same experimental animal, etc – into our experiments. They can give us more confidence in noisy data. The downside is that they do need to be accounted for properly when doing the analysis afterwards.
For example, an experiment with repeated measurements probably shouldn’t be analyzed with a simple t-test.
Let’s use an example to give us an intuition for why.
Imagine you wanted to know whether carb-loading resulted in faster race times for mid-distance runners. Say you hire someone to collect some data for you: they recruit runners and assign them to either carb-load or eat a balanced meal before running a timed 5k.
You see a dataset that looks something like this:
Looks pretty good, right? You do a t-test (the data is about normal) and find a significant p-value: .001. Great! This looks significant! You get ready to publish a paper (or a blog post).
But then imagine that your data collector neglected to mention to you that the measurements were all taken on the same person. The data collector is an avid runner themselves, and managed to run 100 5ks. Do you still believe your t-test?
Of course you don’t. Those measurements are all correlated; they’re quite a good measurement of that one person, but who can say how well that one person generalizes to all runners?
And this is the intuition you are looking for: the measurements collected on your sketchy 5k-running data collector are all correlated, and are not independent. A t-test assumes that measurements are independent, and so if you used that test, you would see a falsely-inflated p-value.
A slightly better way to do this experiment would be to identify ten different runners, and have them run ten races each: five with each type of preparatory meal. In this case we would still be looking at correlated data, because the measurements taken on one person will be similar to each other. Even better would be fifty independent runners, each running two races.
What if the study budget only allows for ten runners? Or what if you expect very noisy measurements and so you want to collect more than one measurement from each subject? These experimental designs are still fine, and we can test our carb-loading hypothesis with something called a mixed-effects model. And that’s a topic for another post.
What on earth do people mean when they tell you they’re going to “fit a model”?
Let’s start with what a model is. A model is a description of a system, usually expressed as an equation of some kind. Let’s say we have some data – measurements of variables x and y. We think that in the future, we’ll have measurements of more x-like values, and we’d like to be able to predict those ys.
Some “data”.
It looks like you could draw a nice straight line through this cloud of points. That means that a linear model might be a good choice for this data. We’ve just done the first step in the model-fitting process: we’ve decided to use a line – a simple linear model.
The process of picking the correct line for this model is called “fitting”. There are different ways to do this – least squares is possibly the most familiar one. You could also use the “wiggle a ruler around on paper” method, or the “draw lines in Powerpoint” method. We’ll skip the details of that step because the internet describes least-squares fairly well.
Two bad fits and one good one.
I’m going to use an equation here, but it’s simple and it’s the only one I’ll use in this post!
That fitted line can be described with the equation y=mx+b. When we fit the model what we’re really doing is choosing the values for m and b – the slope and the intercept. The point of fitting the model is to find this equation – to find the values of m and b such that y=mx+b describes a line that fits our observed data well. In the case of the best fit model above, m is close to 1, and b is just a bit larger than 0.
And why do we care about this? Well, that value of m can be really informative. If m is very large and positive, then a small change in the value of x predicts a tremendous positive change in the value of y. If m is small, then changes in x predict small changes in y. A large m implies that x may have a large effect on y, hence m is also sometimes called the effect size. It’s also sometimes called a coefficient.
The principals of this model-fitting can be applied to linear models created on data with many more parameters – many dimensions. But they are fit in similar ways. We may not be able to draw multidimensional datasets in neat graphs but we can still apply least-squares to them!
Testing for significance
Now that we’ve fit a model and found values for m and b, we’d like to know something: does m really matter?
Take these two sets of data:
Two datasets: both could be fit with a model that is a straight line with slope of 1. Even though that line might be the “best” fit for both sets of points, it would be a far better fit for the first dataset.
You got me; the first example is just a copy of the one above. But the second is another dataset that we could imagine fits the *same* linear model – that is, the best-fit linear model could have the exact same values for m and b. But really, does the value of x predict the value of y well here? We can tell by looking that it doesn’t.
That’s why many model-fitting tools return not only a slope for each parameter, but a p-value. This p-value is an indicator of whether that predictor (x) is actually useful in informing you about the state of the response variable (y).
To assess whether a parameter is predictive, we remove the variable (x) and it’s coefficient (m) from the model. And then we see how good we are at predicting y with a model that doesn’t include them.
In this case a model with no x means that we guess: we create a model where the prediction for y is always the same: the mean value of all of the observed values of y.
Predicting y with our linear model including x, and predicting y without x. Sum the lengths of the yellow lines to get a good idea for how good a prediction we’re making.
We compare the predictions between these two models. If our model that includes x is much better at prediction, we assign a low p-value to that coefficient.
We do this kind of testing for significance in many statistical settings, including one of my favorites: testing for differential expression of genes in RNA-Seq experiments. If a linear model that includes the expression level of gene A is better at predicting which group a sample comes from than a model without A, we decide that gene A is significantly differentially expressed.
Last week was the annual BioIT World Convention and Expo. It’s three days of talks, workshops, and row upon row of technology solutions vendors putting their best foot forward. Also, biotech folks from all over the country converge upon Boston, so I get to catch up with many colleagues I see but once a year.
As usual, the conference is a firehose of information about new technology, new methods, and new software. Also as usual, participating in the Best in Show judging was the highlight of the event. It’s a real pleasure to see the best new technologies our community has to offer. Congratulations to the winners!
The terms probability and likelihood are often used interchangeably in day-to-day conversation. They have specific meanings in the world of statistics, however, and understanding the difference is helpful in understanding statistical methods.
We’ll use examples to start. Take a coin flip: if you flip a coin, and you know it’s fair; a lifetime of experience gives you a model describing the behavior of the coin: half the time a flip will result in heads. You use this probability of 0.5 to decide whether you want to take a bet on the outcome of that coin.
If, on the other hand, you wanted to test whether that coin was fair, you might flip it many times. Say you flip it 1000 times, and you observe 505 heads and 495 tails. Now you want to know: is this coin fair? Is my model of the coin’s behavior correct? Now you are talking about a likelihood; what is the likelihood that this is a fair coin?
In short, a probability quantifies how often you observe a certain outcome of a test, given a certain understanding of the underlying data. A likelihood quantifies how good one’s model is, given a set of data that’s been observed. Probabilities describe test outcomes, while likelihoods describe models.
