Monday, June 29, 2015

What I Did This Year

My master’s thesis, which I spent most of this year working on with my supervisor, Chris Yau, contains 83 pages, 82 equations, and 13,000 lines of code, but because you are brilliant I am going to explain it here in 800 words (plus one 300-word note on what I learned from the project) and you are going to understand all of it.


The thesis develops better ways of performing two important statistical tasks on a widely used type of biological data. First I’ll explain the data, then I’ll explain the tasks, then I’ll explain what we did.


The data is called single-cell gene expression data. In a previous post I explained why we like gene expression data: it tells us how much RNA cells produce from their genes, which tells us which genes are important to the cell’s functions.


Before, we could only measure gene expression data by combining hundreds of cells: otherwise there wasn’t enough RNA to measure. But single-cell gene expression data lets us measure RNA levels within single cells. This is exciting because rather than pretending that all the cells in a tissue, or a tumor, are identical, we can look for groups of cells which are different in important ways. It’s like the difference between asking “What are Americans like on average?” and interviewing individual Americans. (We’re not all bad!)


These differences between cells matter because they have implications for how we treat disease. For example, single-cell analysis reveals that brain tumors which are classified as a single type in fact contain mixtures of multiple types of cells. And this diversity makes the tumors harder to treat: the more diverse the tumor is, the worse the patient’s prognosis, possibly because treatments that work on one group of cells don’t work well on the rest of the cells.


So people are writing papers saying single-cell data will “revolutionize” science. (I would write papers saying things like that, but no one pays attention to me.) Unfortunately, single-cell data is also hard for standard statistical models to deal with, because it has a lot of “zeros”: cells where we cannot detect any RNA produced from a gene. If you plot how much RNA is measured from a gene, it looks like this:
The spike at zero is bad because many statistical models assume data is Gaussian -- that is, it makes a nice blob, pictured below.


Obviously, the spike at zero in single-cell data makes the data very not Gaussian, and this makes it harder to perform many statistical tasks. Specifically, there are two things we often want to do with data -- cluster it, where we divide cells into groups so that similar cells are put into the same group, and reduce its dimension, where we plot cells in two (or more) dimensions such that similar cells appear close together [1]. Standard methods for performing these tasks assume the data is Gaussian.


In my thesis, we first develop a model for the patterns of zeros in single-cell data, and, based on this model, show that zeros really do hurt the performance of standard methods. We show, in several different datasets, that genes that are expressed at a higher level are less likely to be zero, and develop a model that describes this relationship. (I describe the mathematical details here [2].) Based on our model, we generate simulated data and show that the performance of standard statistical methods gets worse as we add zeros. Here’s a demo: move the slider to add zeros, and watch how the clusters (which are supposed to be separate) get mushed together. This is bad. We have an important new type of data; we have important statistical tasks that we often want to perform; standard methods for performing the tasks do not work well on the new data.


Diagnosing the problem is easy; how do we fix it? We develop two new models: a clustering model (ZIMM, for zero-inflated mixture model) and a dimensionality-reduction model (ZIFA, for zero-inflated factor analysis -- new statistical models are required to have a cool acronym). I give more details on how we developed the models in this footnote [3]. We show that our models do a better job than standard models on both simulated data and real biological data and put the code online so other people can use them. We are giving talks on the work at various venues, and a paper is under review. If you’re interested in using or extending these models, feel free to send me or Chris an email.


While we developed these models for single-cell data, they are potentially useful for lots of other datasets which have many zeros or missing datapoints. In the thesis we discuss several examples: recommendation datasets (how much will a person will like a product? -- see, for example, Netflix’s million dollar competition to predict movie ratings), usage datasets (how often a person will use a product?) and network datasets (how strongly are two nodes, eg airports, connected in a network?) Feel free to read the last chapter of the thesis and let me know if you have ideas for further applications (translation: do my homework for me).


Working on this project taught me about the role of luck in research. I often feel there’s a great deal of luck when you investigate some cool statistical hypothesis (“older women with cats are more likely to be single”). The hypothesis may or may not turn out to be true, but its truth doesn’t depend on how good a scientist you are. This luck evens out, though, because better scientists a) look at more things b) look at more interesting things c) look at them more rigorously.


You might think, though, that when you’re just doing math rather than observing the world, there’s no luck involved, only skill; equations are within your control. Still, my experience this year makes me feel there’s still luck involved, and I’ll draw an analogy to chess. Chess, in theory, is a game with no luck; there’s no dice, no drawn cards. But you will still hear chess players say “I got lucky” (or, more often, “I got unlucky”) because they calculated ten moves ahead, and on the eleventh move, their queen happened to be in the right place, and so they won, but they didn’t foresee it.


Similarly, my thesis involves about ten pages of algebra, at the end of which the math works out nicely. But I didn’t know this would happen when I began; this is what I mean by getting lucky.


Again, though, I feel the luck evens out here. I may occasionally beat a chessplayer who sees farther ahead than me because my pieces end up in the right place, but most of the time I’m going to lose. Similarly, I have met mathematicians who can see further down a proof tree than I can (and also have an eerie instinct for which branches will yield fruit) and they can prove trickier things than I can. In the case of my thesis, while we got “lucky” in the end, we also had to fix my screwups revise our approach several times; the final product feels like the result of a lot of blood, sweat, and tea [4]. Or maybe I’m just saying that so Oxford will let me graduate.


Notes:
[1] For each cell, we have data for tens of thousands of genes. It’s very hard to visualize or analyze data with ten thousand dimensions. So we want to reduce ten thousand to something more manageable, like two, in a way that preserves basic patterns in the data.
[2] Call the probability that a gene is zero p0, and the average expression of a gene when it’s not zero μ. Then, p0  = exp(-lambda*μ2), where exp is the exponential function and lambda is a parameter we fit to the data. Here are four different real biological datasets: the orange line is the actual data, and the red is our fitted model. You can see the orange line fits the red line pretty closely.


[3] Here’s a two-step procedure for developing statistical models:


  1. Write down (using math) the process you think generates the data. This is usually the straightforward part. In the case of ZIFA and ZIMM, our process is “generate data using a standard statistical model. Then, add zeros to that data”.
  2. Your model will have “parameters” -- numbers that affect the predictions it makes. For example, if your model is
    number_of_colleges_you_get_into = A*your_SAT_score + B*your_parents’_income
    then A and B are parameters of the model. Obviously, in order to calculate useful things with a model (in this case, how many colleges you are expected to get into) you need a way of figuring out the values of the parameters (for example, A = .01, B = .001). This is the hard part; it took me most of the winter.
[4] The British do not cry, or if they do, they cry Earl Grey.




Tuesday, June 9, 2015

So Maybe I Am (Partly) An Affirmative Action Admit


The day I got into MIT, I listened to the boys in my high school physics class whisper that girls had gotten in just because they were girls. Half the female undergrads at MIT have heard such comments, in part because female applicants to MIT are accepted at a much higher rate than male applicants. The explanation given by the MIT admissions office is that girls don’t apply unless they’re especially well-qualified, because MIT is a school that excels in stereotypically male areas.

Though I ended up turning down MIT for Stanford, still I wondered if my gender had helped me get in. My high school class elected me “nerdiest girl”; my application highlighted my interests in male-dominated areas like math, physics, and artificial intelligence. As a nerdy girl, I was an outlier; as a nerdy boy, I would’ve been a stereotype. Two of my acceptance letters commented on the fact that I played chess, a male-dominated game.

So I was excited when I learned that Stanford was required to provide the comments admissions officers had made on applications. When I tried to obtain these comments, I discovered that Stanford does pretty much everything it can to prevent you from getting them [1]. You get an email asking you to “consider the administrative and logistical burden” and whether your life will “be better for having reviewed them”. They will not send you the records via email; when I showed up in person, an employee told me I would not be allowed to use a phone or computer to take notes.

“Can I just have a copy to take home?” I asked.

“No,” he said. “You have twenty minutes.” He started a timer, and I began copying out the comments word for word.

My application was reviewed by two people. It is clear my gender mattered to the first reviewer, who referred to me using terms that emphasized that I was female and discussed my essay about playing chess on Chicago’s South Side with a bunch of older black men. “I can just visualize this young white woman talkin’ trash and kickin’ butt in the South Side chess joint,” the reviewer wrote. Would the image have been as charming had I been male?

Then I emailed every other college I applied to to request my admissions records, because that is a totally normal thing to do. They all told me, politely, to go away; probably the “Stanford” in my email address didn’t help.

Based on the comments on my admissions letters and my application, I conclude that my gender may have helped me get in. A non-rigorous survey of my Facebook friends revealed that I am not the only female scientist who wonders about this. About half of the female scientists who answered said they thought it was possible they had earned an opportunity in part because of their gender. (Some fiercely rejected the idea -- “I topped my university because I am smart and worked hard and not because I am a girl!” -- while some took a more moderate position -- “I think it doesn't hurt to be female when applying to places... In general, I try to work hard enough so that my ability/merit is at the level where it's so clear they should accept/hire me that gender isn't even a question.”)

How much does the possibility that my gender helped bother me? Not much, for three reasons.

  1. Most of me believes that I would’ve gotten in had I been male, that I am not just good at math “for a girl”. I did better on most college math tests than most of the the guys in the class. I also took one math class under the psuedonym “Andrew Chang” and the gender switch didn’t make much difference. (If anything, incidentally, graders in math classes are biased against girls).
  2. Even if I wouldn’t have gotten into the same schools, I’m okay with schools giving preference to girls in male-dominated fields. Colleges should admit people based on their potential, not just their current position. And a high school girl in the male-dominated fields faces so many handicaps a boy does not -- social pressure to leave, stereotype threat, bias -- that if she nonetheless achieves the same score on a math test, that seems like evidence that she has more potential. (Similarly, we’re more impressed by the kid who manages to teach himself integrals out of a textbook than the kid who goes to a school where everyone takes BC Calculus.) I also believe that a gender-balanced college environment is a better one, both because of the sexism which can flourish in male-dominated environments and because men and women have reliably different views on certain issues, making gender balance in academic discussions important. (More on this here [2].)
  3. But let’s say you don’t buy any of this and think I was unfairly admitted to college. Honestly, I still don’t really care because in terms of things I’ve been unfairly given, I’ve got way bigger problems. My graduate education is funded by a) a guy who’s literally getting feces thrown on his effigy b) the US military and c) an organization with a strong history of supporting the nuclear arms race. I am constantly forgetting my keys and being let into buildings anyway because I am white and am carrying a nice laptop, and let’s not even talk about where the laptop came from. In contrast, the people harmed by an affirmative action preference towards me are men who are better than I am at computer science. I’ve met a few of those guys. They’re doing okay.

You might say this is a non sequitur because the presence of a larger injustice does not mean we should ignore a smaller one. I agree that it is important to consider how much affirmative action we really want, and to study, for example, whether or not female scientists have an advantage in applying for jobs. But in terms of injustices that I personally contribute to, I have bigger things to worry about. And if you are one of the dozens of guys from my high school who is annually rejected from MIT: before making nasty comments about how unjust the system is, you might consider how many injustices have already gone your way. 

Notes:
[1] This is not the first time I have found college officials reluctant to provide information on admissions processes. While touring computer science graduate programs, I spoke with one dean who told me his department had changed their admissions criteria to increase the fraction of women admitted to the program. Without thinking, I asked, “So am I statistically dumber than my male peers?” I have no idea why I said that, but he hastily backtracked and said, no, it wasn’t me he was talking about, it was the undergraduates. (In any case, “dumber” was an inflammatory and incorrect word to use; he had been talking about correcting for women’s lack of experience coding.)

[2] I am troubled by the sexism that often seems to take root in male-dominated environments like fraternities, tech companies, venture capital firms, and gaming communities; women who work in male-dominated environments also often report feeling isolated, which makes them more likely to leave. Gender balance in academic discussions is also important because, while gender gaps for most traits are small, men and women have reliably different views in some areas -- political orientation, sex discrimination, sexual assault. Even within the supposedly objective academic research world, it is clear that a researcher’s gender influences what they choose to focus on. It is no coincidence that many of the high-impact papers studying gender -- on bias in orchestral auditions, gender gaps in desire to compete, and gender inequality in deliberative participation -- have female authors, or that gender studies departments tend to be mostly female. Would medical experiments have been performed only on male lab animals (controlling for the menstrual cycle is a pain) had more of the experimenters been female?

pearson pearson pearson