Carli: Hey everyone and welcome to Science in Real Time. This is the show where we sit down with the people pushing science forward and talk about what's really happening behind the breakthroughs. Before we dive in, make sure to subscribe so you never miss an episode. We've got some incredible conversations coming your way. All right, let's get into it. I was wondering if to start this episode you can share a little bit about your background and what drew you from neuroscience in particular and behavior into machine learning.
Reese: So my background is in behavioral neuroscience. I did my PhD in olfactory search behaviors in unrestrained mice and I also did some post-doctoral work on that and now I work as an AI data scientist at ViQi. What drew me into machine learning, it was kind of a natural progression. Behavioral neuroscience is very, can be very naturalistic, very messy. That's what I was attracted to. That's what I wanted to study is complex environments, complex behavior and so how do you structure complex continuous signals. Machine learning is a really obvious way to do it especially unsupervised machine learning and so that's what drew me into it. It started with as a collaboration with one of the theoretical neuroscience labs and my interest in skill set just kind of grew from there.
Carli: It sounds like you've spent more than a few years studying animals and kind of like navigating their complex environments.So I was wondering how has that informed how you think about data and pattern recognition specifically in biology?
Reese: Yeah, I think I've always been most interested in solving the really difficult messy problems and that's what attracted me to behavior. It's what attracted me to animal behavior and neuroscience and I think it develops an interesting perspective because most biologists will tell you that we're the field where patterns are meant to be broken. There's almost always an exception to a rule that you set and I think that I'm very well set up to understand that because I was at the kind of extreme edge of that in my research and I got to experience computational people and physicists joining the field of neuroscience, studying behavior and watching their expectations be not met in the messiness of well right now the olfactory nerve is going to be right right here when you do the surgery and then on the next mouse it's going to be for no reason two millimeters back. So I think that I take a very conservative approach to pattern recognition because of knowing about those exceptions.
Carli: That makes a lot of sense and I one thing that is coming up for me in your academic work and considering this very messy side of biology that we're speaking about and how much of a wild west has AI been becoming as it is raising in popularity and everybody wants to use it. I'm wondering you've used AI to uncover hidden behavioral states in mice and I'm wondering then how does that experience translate from going from behavioral states in mice to detecting hidden cellular phenotypes for example? How do we bridge both of those things?
Reese: Well it's extremely different and extremely similar at the same time because the process is very similar although I did use different types of AIs in doing my behavioral state work. The idea of trying to quickly and accurately identify pattern outputs or clustering or target phenotypes without having to see them obviously on an image, on a video, on a slide. The process remains largely the same. I think that experimental design is at the forefront of using AI responsibly and effectively in biology.
Carli: That makes complete sense and I'm wondering when you joined ViQi then how did your approach designing machine learning models for complex self-imaging databases change if it changed at all or how was that experience for you integrating yourself in that environment?
Reese: I think when I was studying behavior one issue that I had was when I initially started was we wanted to study search with airborne airborne odorants with wind and try as we might it was pretty much impossible to control the olfactory stimulus to the extent that we wanted with the wind blowing across the arena and then the odor coming through. So instead we had to design our experiment around the fact that we could not control that stimulus and we couldn't know the exact stimulus at any point in the box. So that's kind of where the experimental design comes in. With ViQi it's a little bit of a alteration of that but it's the same idea. We work with very well-defined controls, known controls, and I think that identifying the controls in your experiment, I mean everybody knows that that's important, but specifically with AI when you're sending it into that black box you want to have an understanding of what bounds your output and what structures your output.
Carli: That makes complete sense and I think that that actually ties pretty nicely to what I was thinking in terms of if we have a very messy data set then how do we go about interpreting it and how do we go about emphasizing this interpretability? So I was wondering how do you ensure then that AI outputs are made interpretable and useful for biologists beyond these controls? Is there something that folks should consider beyond these controls that they use for AI experimentation?
Reese: Well it is the controls. Like having well-defined controls and having target phenotypes that you're interested in, in our case, that really helps interpret the AI output. Now if you don't have that we do run those experiments too. We do this phenotypic profiling where we don't have any known phenotypes and we look at how the phenotypes cluster across AI confusion and therefore in phenotypic space. And in that case what I think is most important is we can do all these statistics to identify how close we think those clusters are, how tight we think those clusters are, but it comes back to eventually validation and evaluation of the model. And I think that's at the forefront of the field of AI right now is how do we evaluate our models and make sure that they are effective to the extent that we want them to to be. And that unfortunately I don't think is much of a shortcut. Like we have to use traditional methods to validate.
Carli: That makes a lot of sense and one thing that I'm thinking now is how do we communicate these very complex ideas and how do we interpret these ideas? And I'm wondering if you have learned, what have you learned about communicating complex computational ideas to teams grounded in experimental biology? Considering that beyond there's many things that are different amongst the different groups and how they approach the question. So I'm wondering if you have learned anything in communicating to scientists that work more with the experimental biology side of things?
Reese: Yeah, I think generally experimental biologists have a pretty good understanding of basic traditional statistics, usually advanced statistics. And so being able to communicate in that language and then apply these AI computational ideas in that language, again using controls but also using tests that they recognize, making sure you're using the correct statistical tests that they recognize. That's how you communicate across the field, is being able to speak the language to a certain extent that there is a common shared vocabulary.
Carli: How can data scientists build trust and shared understanding with collaborators, especially those who are skeptical about the black box of AI as you were saying earlier?
Reese: I think to a certain extent demonstrating that it works, like whatever experiment that you're running, being able to show over and over again replicability in the result. For a lot of collaborators, that is sufficient. For some, again that external validation, ViQi's done some work doing bioinformatic based validation of our phenotypic clusters when we don't have known targets. And we have seen validation of those phenotypic clusters. That builds trust with the biologist. That's something a biologist can recognize and understand. And so being able to do things like that and even better is using a traditional lab assay to validate the AI result. If you can do that, even a couple times, it builds a lot of trust.
Carli: How do you folks navigate that at ViQi? Do you folks collaborate in the sense of designing these experiments to be able to validate? How does that work in the company?
Reese: It depends on the assay that we use. As I said, there's this open source dataset jump that has a bunch of compounds. It's across a bunch of institutes. That was just a very large high content screen. We can use that and we have the identities of the compounds. We have the smiles for them and we're able to cluster them using our phenotypic profiling tool. And then we have our bioinformaticist, Becca, run through open source bioinformatic databases and start trying to validate the cluster by mechanism of action. We've seen good overlap. And so we know that there's at least some meaning in her phenotypic clusters. Obviously, the more we can do that, the better. In other cases, for example, with viral infectivity assay, we can infect cells at such a high MOI that we know that the cells are very greatly infected. And then we can have cells that are not infected that are healthy at an MOI of zero. And again, that bounds the phenotypic space so that you can evaluate infectivity without doing a traditional plaque assay or TCID-15.
Carli: What do you see so far as the biggest opportunities and also ethical responsibilities, if you will, for scientists that are developing these tools?
Reese: Accelerating discoveries, accelerating workloads, reducing workloads by accelerating pipelines. Personalization is huge. Being able to optimize something towards a person. This is huge in medicine, personalized medicine, but it's also big in services in general. And then optimization of overburdened systems. We have a lot of them in our country and across the world. And optimization of burden systems is a huge opportunity in AI. As far as ethical responsibilities, this is the obvious one, but AI is trained on human input and human behavior. We'll behave like humans. Humans are biased. We've seen lots of evidence of discrimination with, for example, face tracking initially. And so being able to do the same work that we're doing personally in the AI field to make sure that we are reducing discrimination and bias as we can. Transparency and explainability. Evaluation of model outcomes using traditional methods, using fancy statistics, but evaluation of models. And then equity. Who has access to these optimized systems? It's not optimizing overburdened systems if the overburdened systems never see these tools.
Carli: Absolutely. And I'm wondering then, how can emerging scientists then start preparing to work effectively across both biology and data science when considering these things?
Reese: I think emerging scientists should make a conscious effort to be cross-disciplinary. It's really easy to pith and hole yourself. And if you can collaborate broadly, you're being a self-advocate for your future career. Even if you're planning to stay in academia, I wish I had kept a better pulse on the market. And I kept a pretty good pulse on the market, but I wish I would have been better about it because understanding what is marketable right now can inform your side projects, your collaborations, that small conversation you have over coffee with another professor. And again, you're only being a self-advocate for your career by being as broad as possible and developing as big a skill set as you can.
Carli: Let's shift gears a little bit and talk about women in AI and science. I'm wondering, in such a male-saturated field, how do you mentor or encourage other women entering computational or interdisciplinary sciences?
Reese: Yeah, absolutely. I've had a lot of experience in spaces like this. I also, in my spare time, like to build things and build electronic art pieces and stuff, and that's also a very male-dominated field. And also, I'm a Latino woman in science, and so I've done a ton of mentorship for Latino women in my PhD program. And the main way that I encourage any minority group to enter a field where you're going to be a minority is that imposter syndrome is real. Also, you're not the only one experiencing it. Everybody experiences it to a certain extent, and you know more than you think you do, almost always. And that's definitely true for almost every minority, because they're continuously told that they don't know as much as they think they do. And that just lives in your head. It's a repeated thing that just exists with you, and you have to live with it. And so, whenever you enter a room, you know more than you think you do. Try to hold confidence in that and just find your community in that area, too. It helps a lot.
Carli: What mindset has helped you stay resilient and creative in tackling hard technical and scientific problems, regardless of the challenges that you may encounter?
Reese: Well, I think this comes back to something that's taught often in diversity and equity circles. You know, I think it's extremely challenging in our culture to approach things humbly and with confidence. Nobody wants to admit that they are part of the problem. Even they will say they're part of the problem, but then they feel so bad about it. They can't do anything about it. They get paralyzed. And so, I think that I've taken from that world that having a lot of humility but confidence will get you far in any field. And so, that's what I do when I'm tackling hard technical and scientific problems. That's what I do when I have imposter syndrome in the room. I might not know everything. I might be messing this up. And here's what I think I do know, and just trying to keep those balanced together.
Carli: And for listeners that are inspired by your path, I'm wondering what advice would you give about forging a career at the crossroads of computational biology in particular?
Reese: Yeah. I mean, what I would say is courses are great, but you don't need them. Hands-on experience is just as good. And I certainly didn't know anything about AI when I started my PhD and even partway through my PhD. And when I did start engaging with AI, it was because I wanted to use DeepLabCut, which was an AI-based tracking tool. And I was working with my colleague, David Wyrick, and he was a computational neuroscientist. And that's how I got introduced into it. It wasn't through coursework. And so, I think that if you can find hands-on projects, and again, broad, do different types of AIs, do different types of biology, get as broad as you can because diverse skill sets are kind of the name of the game. Everything is cross-disciplinary now.
Carli: I definitely agree. And I echo that for our listeners as well. And I'm wondering then, if folks want to get in touch with you or collaborate with ViQi, what would be a good place for them to go and find you folks?
Reese: Our website is viqiai.com. And also, you can email me or our CEO. We're pretty open, and our contact information is on the website. And you can also find us on LinkedIn.
Carli: Absolutely. And we will certainly have the information for both ViQi and Reese on the description notes so that you folks can get in touch. And yeah, this basically wraps then our episode of Science in Real Time for today. Dr. Finley, I want to thank you so much for joining us. This was incredibly insightful. And your journey from neuroscience to AI really shows how curiosity and computational biology can be together and also help how we see biology in very completely new ways, challenging the complexity that we generally encounter. So this was a pleasure, and I hope definitely that we will cross paths in the future. I will certainly be very, very attentive to what ViQi is going to be developing in the future. And yeah, this was amazing.
Reese: Thank you so much. Awesome. Thanks for your time. I really look forward to hearing it.
Carli: Before we sign off, here's a quick announcement from ViQi.
Reese: Hi, everyone, and thanks for tuning in. ViQi is developing a new protein expression assay, and we're looking for collaborators to provide us with image-based data sets. This is an opportunity for publication and for your group to be an early adopter and driver of new technology that has the potential to transform biomanufacturing pipelines. If you're interested, go to our website at viqiai.com or contact me directly at reese@viqiai.com or on LinkedIn. Thank you.
Carli: Thank you so much for tuning in to Science in Real Time. We're really glad that you were able to spend time with us. If today's conversation sparked questions or ideas, we'd love for you to keep it going. You can connect with our guests and the Science in Real Time team using the links in the description box below. And if you enjoyed this episode, don't forget to like, subscribe, and share with someone who loves science as much as you do. Until next time, thanks for watching.
Transcript source: Provided by creator in RSS feed: download file