As part of our ongoing AI In your Community series, I sat down with Dr Anne Carpenter who leads the Carpenter Lab at the Broad Institute of MIT and Harvard.
Anne Carpenter
Tara Chklovski: What problem are you working on?
Anne Carpenter: My group is developing software to accelerate drug discovery. Normally a researcher works on a single disease, devoting her entire life to studying it. We are working on a general software tool for image analysis. You may wonder why we need image analysis to cure different diseases. This is because one of the favorite tools of a biologist is a microscope, and even though we have robotic microscopes that capture millions of images a year, no biologist wants to look at all those images by eye.
So we write software that can automatically identify what the biologist thinks is important or interesting in that image. The software could measure how metastatic the cells look, how how much they are infected by tuberculosis etc.
Another very common experiment we run is to help Pharma companies identify effective drugs. A pharma company may have a million compounds that they’ve created that all have different structures and different impact on humans or on cells. So we test each individual compound to see if it has a desirable effect on the disease. We grow metastatic cells in a dish; we add a compound and determine whether they look less or more metastatic. If we have tuberculosis growing in a dish we see if the drug seems to kill the bacteria or not.
TC: Interesting! So what are you finding?
AC: What we’re finding is that this is a efficient way to discover new drugs. It’s highly automated — it involves the physical robotics of the automated sample preparation that pipettes tiny amounts of drugs onto cells; the microscopy is completely automated, and now the image analysis is automated. It’s a just a very efficient way to test compounds to figure out which might be effective drugs. Of course, there are a lot of steps and clinical trials after that to go through, but this is a really efficient approach.
TC: Are humans involved in annotating the data or saying, “This image means this or that”?
AC: Humans are involved usually at the beginning, setting up the pipeline to make sure that it measures important, valuable things, and checking that the software doesn’t get fooled by common, confusing artifacts, and when we do use deep learning, we need to provide some annotation, labeling what a metastatic cell looks like, or what an infected cell looks like.
TC: Have you found any kind of patterns or insights that humans haven’t been able to find?
AC: Yes! It happens more often than we are able to follow up on. One thing about biologists is that they don’t necessarily trust the computer model when it says, “Oh look, this is interesting.” They’ll say, “Oh, neat! But I’m busy so I’m going to carry on with my work.” I don’t think that’s an unreasonable thing to do because if you have a plan to discover a drug, and you are making progress on your plan, you don’t want to be distracted by something that the computer found. You have plenty of useful things to do! So we’re always encouraging people to let the computer take a little more control, and in some cases that’s really paid off. For example, you may want to know when a cell is about to divide. You may have seen pictures of cells when they change their shape and squish into two pieces. Normally if you look at a cell under a microscope and you see a round shape you would not be able to say when that cell was about to divide. We ran an experiment using deep learning and it can do an amazing job of identifying seven stages of the cell cycle — only two of which are detectable by the human eye! It can detect the different stages with eighty percent accuracy.
There’s a pharmaceutical company, Recursion Pharma, that is based on this idea. They use images of healthy normal cells and images of cells that have a mutation associated with a rare disease. They ask the computer if it can spot any differences between those two sets of cell images. And when they do find a difference, they screen drugs to try to find ones that can make the unhealthy cells look more like healthy cells. The company has seen some really phenomenal results using this approach.
TC: I am curious, how did you get into this field of computational drug discovery?
AC: I only got into software development when I was around twenty six. I finished my PhD in cell biology and knew absolutely nothing about programming. I was a nerd, but not that kind of a nerd! I wasn’t interested in computers at all. But I had a problem that I wanted to solve. I was a biologist and I had microscopy images and I didn’t want to look at these all by eye and the software that I could find wasn’t doing a good job. So I tried to figure this out myself, and now I’m leading a group at the Broad Institute of M.I.T. and Harvard that is doing computer science on images. It’s actually not that hard to pick up and learn and also, it’s never too late! That’s my message to both parents and children, that it’s never too late to pick up on these kinds of skills.
TC: What in your opinion makes a good problem?
AC: I think a good problem is an important one. It’s nice if it’s feasible or easy, but you should take something that really matters even if you feel you can’t solve it!
TC: What in your opinion makes a good product?
AC: My experience is in building software for biologists. It was helpful that I was a biologist myself. I think really close interaction with the end users, and ideally even being a user of the software yourself, can really help. But don’t assume that everyone is exactly like you; think how might this be used by somebody who’s older than you, younger than you, a different gender, or in some kind of different life circumstance. Ideally, try to talk to those people and have them test your software as well. So many apps these days are written by young professional Caucasian men, leaving a lot of opportunity out there if you see a need in the world that they don’t.
TC: What inspires you?
AC: Definitely making an impact on the world. It’s my primary motivation. I really want to make a difference and I find it most inspiring when it’s clear that what we’re doing is is making a difference.
TC: What do you find difficult?
AC: The vast majority of what we do, as researchers, fails. So we have to keep picking ourselves up and trying something new. We as scientists are highly skilled and knowledgeable, but the whole point of doing research is to do something new, explore the unknown and so you have to figure it out as you go. A lot of success has to do with being able to persevere and to believe in your mission enough to push through the depressing parts of it!
TC: What field of A.I. excites you the most?
AC: Well of course the one I’m in. I find that the most fascinating, but if I had to pick something other than biomedical research, I would pick text analysis (Facebook posts and and other written material) as early indicators of whether somebody is a suicidal or having an early onset of Dementia or Alzheimer’s. I think it’s really neat to be able to take the data that’s already out there and use it for public good.
