Welcome to Tech Bio Talks, where we explore the bold ideas, breakthrough technologies, and visionary people reshaping the future of AI and medicine. My guest today is both a pioneer in the field of image-based profiling and a fierce advocate for open science, a combination that has led to hundreds of labs across the industry adopting her team's tools and data sets to accelerate drug discovery. She's also one of our earliest supporters and scientific advisers at Recursion. In fact, her work is so closely tied to the founding of our company, and we'll get into more of that later, that we named one of our labs in Salt Lake City in honor of her, the Anne E. Carpenter Building. Currently, she's an institute scientist at the Broad Institute of MIT and Harvard. I am so excited to have this conversation. So, let's get into it. Welcome, Ann Carpenter.
Thanks. So excited to have a chance to catch up with you, Chris.
Likewise. So, I remember back in October of 2013, my co-founder Blake and I were at some sort of drug discovery meeting in Cambridge and we saw you and you were like the most exciting scientist that we had ever wanted to meet in the field. You had pioneered this new kind of omics called phenomics. And we, as I recall, tried to get you to go to dinner with us to talk about science and you said no. Let's start there maybe as a good place to start.
Oh great, let's start with the most embarrassing story. So I was a very young professor at the time trying to establish my lab and that was the first meeting that I had brought a postdoc with me. And my goal, my one goal for this entire conference was like make sure my postdoc meets very important people and you know has the opportunity which
was not Blake and me, not
it turns out a couple of grad students who hadn't even finished their PhDs was not high on my agenda. So I was like, "Oh, sorry." You know, we have plans because we did have plans to like find somebody better to go to dinner with, right? So it's embarrassing in retrospect, but turns out everybody who goes to Boston to a conference like knows somebody and I was a young professor. I was not a VIP to the others. So long story short, it was like 5:00. We hadn't found anybody. I was like, "Uh maybe I can find those grad students because they seemed cool. Like let's go see if we can find them." And like thank goodness we did because you guys told us that was before Recursion existed. Maybe even before you had a name and so well yeah in retrospect that was an excellent choice and really glad it happened.
Yeah we actually came up with the name for the company when Blake and I went out for drinks after we had dinner with you. But it was an amazing dinner. We learned so much about the way you were thinking about the field. It felt a little bit like a dissertation defense. You asked us a lot of questions which was super fun. Um, you know, as you think back to the time when you were pioneering this field, like what initially led you to believe in the incredible potential of images of human cells as this breakthrough path to understanding biology and chemistry?
Yeah, when we met, I was excited about the potential, but I would not have, you know, it was not certain. It was definitely an uncertain time and so I mean for me I had grown up in a generation that had just discovered mRNA profiling in the late 90s and so profiling was really a new concept and I just was absolutely in love with the idea. So then when I got into images like hey why can't we use the cell's morphology which as a cell biologist I had already fallen in love with I knew it was incredibly informative to my eye but like what if that could be extracted by computational methods and what if we could use that for profiling but it was by no means a done deal at the time. Um it was clear there were some studies um the most important one was Perlman 2004 um from Mitchison and Ulrich labs and um it was clear that it could work for really strong phenotypes but it was not at all clear that it was going to generally be useful. So I spent some of my startup money, I spent some of my postdoc's time working on it. Nobody's main project because it was too risky at the time. But over time it became clear like ah this could actually be just as powerful as mRNA profiling and yet we're using a different image type and a different data type and therefore getting really different complementary information. Um, so it was I think it wasn't really until 2016 um that my postdoc showed me the mo like the moment the moment that I knew this phenomics image based profiling was like really going to work for real stuff in the field was when my postdoc Muhammad Rohban showed me this circular dendrogram of around 100 genes that had been overexpressed and they fell into such delightful clusters that as a biologist I knew matched to real biology. Um it was like literally like a teary-eyed moment that hey this is this is really going to work. But you guys were ahead of the curve because that was two years after you guys had launched Recursion.
And as you think back to like the beginning of your work in this space there were many people not many there were a subset of people in our field who were interested in images. One of the things that you did that I think is really fascinating is that you actually built tools and open sourced them for other people. I remember in my dissertation and even the early days of Recursion, we were using CellProfiler and CellProfiler Analyst. Talk a little bit about what led you to make the tools themselves. And do you think that ended up fundamentally shifting the way that this technology was adopted?
Yeah. Yeah, for sure. I mean, I think everybody's career makes more sense maybe in retrospect than it does as you're going along. You just make decisions as you go. But if I look back, it actually is a really nice narrative that we spent the first maybe 10 years of my lab making software that would allow people to get information out of their images and then the next five years or so making data sets that we could so we could actually tell are we doing a good job at matching phenotypes and matching genes to compounds and genes to genes and so on. Um and now the past five years we've really shifted to what are all the different applications we can actually do. So, I feel like we spent a lot of time building the foundations of this field to even be able to do what we're doing now. But this is really the fun part for me is accomplishing real tasks in the drug discovery space.
I think you've done a lot of great work to show the field that sometimes you won't find what you want in a specific cell type or a cell type that's not associated with a disease, but very often you can find the relevant biology. I mean that's a huge surprise to traditionalists I would say in the
Yeah. And we find for example we find the vast majority of FDA approved drugs produce a Cell Painting phenotype in just a regular old cancer cell line growing in a dish. Um that is not something that I would have guessed or expected 10 years ago. Um but that's the data. That's the reality. And so like goodness why not pick all the low-hanging fruit first and then we can move to the diseases that actually do require those much more complex model systems with those trade-offs that come with it.
So Ann, you've talked publicly about your purpose which is centered around creating quality data and sharing strategies for drug discovery applications versus building the models themselves. What would you say academia's role is in pushing the industry forward? What would you say your role is in pushing the industry forward?
Yeah, it's easier for me to answer, you know, for my own lab. I think for the whole field of academia, I think the world works best when everybody finds a cool niche and does their thing. You know, the thing they're passionate about, the thing that really matters and hopefully getting input from the people who can advise them on what really matters. Um so, of course, in our space, it's really crucial for machine learning labs to be pretty connected to biology labs and drug discovery labs. But speaking for my own lab, we personally have not prioritized lately like dorking around with model architectures and trying to figure out how to improve the training of these kinds of models. It tends to give you like percentage point increases over time. But I think for me, given my unique niche in academia, the thing that we can do best and most helpful to the community is to work with drug discovery researchers to figure out what tasks are most important to solve and then what data do we need to create a benchmark that then the machine learning people can go to town on. Um but really it's like framing the problem I think is a real bottleneck and that's what we've really decided to focus on. So I would say deciding what data to make is by nature determining the problems that you can actually solve well and that's really where we like to live. And I don't think anybody disagrees that biology is data limited. I mean we have certain large data sets but data sets that can answer particular questions like will this drug be toxic or what's this gene's function? Those kinds of things. Um we still have plenty of data that could be made.
And what kinds of data do you think are most important for the field to make next? Like what is on your list for the next big data set?
Yeah. So I mean we have lots of ideas of things we would like to work on. But I think one that is very timely because there's been a lot of discussion about reducing the use of animal models. Um partly because people care about not using animals, partly because they're very expensive, and partly because they actually just don't work that well. You know, we have so many drugs getting to human clinical trials and then failing for tox. Not as many as in past decades to be fair. Um, but it's still a sort of a terrible problem. So, that is a big goal of this data generating OASIS consortium that Recursion is actually part of and that I'm leading together with a colleague at Bayer on the agrochemical side. So actually this is a problem that both pharma and agrochemical really wrestles with is how can we make sure we produce chemicals that are least harmful to human and other animal species. So unless we're trying to kill them as pesticides of course. So it's a cool consortium because it is also multimodal. We're gathering Cell Painting image data, transcriptomics and proteomics which is usually not we're not able to do at very much scale. Um, so pretty excited to do that in some different liver cell models. And the whole goal of the consortium is which of these data types and which of these cell models can best predict actual liver outcomes in actual living organisms. Um, so I think that'd be a huge win towards reducing the cost of drug discovery, which of course is a major goal for our industry to be able to keep making medicines.
