Please note transcript has been edited for brevity and clarity.
FLG: Hi everybody. Ahead of The Festival of Genomics, Biodata and AI in Boston this summer, we're speaking to some of our expert speakers to get a feel for what they'll be discussing at the event. Today, I'm here with Tom Lanz from Pfizer, who will be speaking on day one of the Festival on multi-omic case studies in drug discovery. Thank you for making the time to speak to me today, Tom. Could you start by telling us a bit about yourself and your background?
Tom Lanz: My name is Tom Lanz. I work at Pfizer. I was trained in neuroscience, and spent the first 17 years my career working on new targets in the neuroscience space (psychiatry and neurodegenerative diseases), before transitioning over to drug safety a few years ago. My group is currently called Multi-omics and Biomarkers, and so we sit at the intersection of omic technologies, computational biology and translational questions in the drug safety realm. We work across disease areas and across the whole portfolio, from early targets to phase three and beyond.
FLG: And could you tell us a little bit about what you're going to be talking about at the event?
Tom Lanz: I've got three case studies I'm going to share for this year's Festival. The first is using RNA-Seq and laser capture microdissection, some fun tools, to try to understand in a preclinical model, an observation we saw with one of our compounds in the clinic, where patients were experiencing dysgeusia. So, we took an approach to try to understand that in rats, and we've got some interesting insights that we're following up on clinic. The second case study is using multi-omics approaches to try to find some biomarkers of checkpoint inhibitor-induced liver injury. We've got a really nice cohort, with a prospective trial [design]. We have samples before and after treatment, and we're able to use multiple different kinds of omics together to identify some potential novel biomarkers in this space. This is a nice story, bringing together multiple omics. Then the last case study is another variant of liver injury, this time in response to gene therapy. We had a very complicated study design, and we ended up using some foundation models to help us parse through the data and focus on some potential biomarkers that we think are performing better than the current gold standard.
FLG: Well, it sounds really exciting. Now, what makes multi-omics approaches particularly powerful for understanding drug safety?
Tom Lanz: Biology is multi-dimensional in nature, and very often we're left with just taking a snapshot in time and trying to dissect the chain of events that led to a finding. And sometimes it's not clear. And so the more different lenses that you can apply to a problem, I think the better odds that you can try to disentangle what happened.
FLG: What defines a good mechanistic biomarker in this context? Is it sensitivity, specificity or something else?
Tom Lanz: So, that question really depends on what is the context of use for that biomarker? In the drug safety world, usually sensitivity is paramount. You don't want something slipping through. And translation is also a big piece of it. It has to be sensitive. Ideally, it's something that we can understand well enough preclinically that could translate into the clinical trial. If you're in another realm where you're looking to diagnose something, specificity may be more important in that context. But in drug safety, really, sensitivity tends to be the most important. We don't want to miss something that could be potentially problematic.
FLG: As part of your talk, you're going to discuss the use of AI foundation models. What advantage do they offer over more traditional analytical approaches?
Tom Lanz: They're really good at certain tasks that conventional approaches fall apart at. Batch integration works really well and tends to outperform standard PCA methods when you have a complex study design, or your study is either very high dimensional but low powered, or something along those lines. Traditional stats can sometimes underperform, whereas the foundation models, they're trained on biology. And so by focusing on differential attention, rather than just simply differential expression, sometimes you can pull more out of a study than we could otherwise. And we've seen that in a couple of cases where, again, the batch integration is really great at doing that, integrating data, and then helping to sift through the noise and identify something that is biologically meaningful. And the hope is that, because they were trained on biology to begin with, the things that are floating to the top are not just the most significant statistically, but they have some kind of biological underpinnings that make it more interpretable.
FLG: And where does AI add the most value in these kind of workflows?
Tom Lanz: I’d say when your data sets get large and complicated, AI can certainly help in integrating the data and helping you to add some context to what you're seeing. If you have 1000s of things that are differentially expressed, I've seen that AI has been really useful in helping to find those patterns that are linked to the phenotype. And then as I mentioned, the batch integration performs very well. And so as studies get bigger, I think tools like this become more and more, not useful, but almost necessary in some cases.
FLG: How do you ensure that these models are robust and reliable, and how do you make sure they're trustworthy?
Tom Lanz: We're using them to try to, in many cases, find a biomarker, or understand a mechanism, and so while we benchmark them against traditional methods, we're looking for things that make sense biologically and that we can independently validate. So, if we got something out that was absolute nonsense, it didn't make sense, we're not going to just blindly trust that and employ this biomarker, we're going to try to understand where that came from. But if it comes out with potential biomarkers that make sense, we still validate those in another cohort, independently. And so we're using it as a tool to find that needle in the haystack. We're not relying on it as the decision maker.
FLG: Could these kind of methods be used to predict, say, drug side effects earlier in development?
Tom Lanz: Absolutely. I mean, that's what we're hoping. And as we learn more about these models and our data sets grow, I think that's where the field is moving. It’s more prediction, less reliance on animal studies, and the FDA has put out their guidance, and they're looking forward to seeing more data sets from new alternative methods in the future, more AI, more virtual and modelling approaches, rather than simply the routine pathology studies we've been doing in animals for the last 100 years.
FLG: What are some of the developments in this space that you're the most excited about?
Tom Lanz: The models keep evolving, and their performance is improving, but also there's newer models that are able to handle different kinds of data. So I'd love to have a great model that is trained in biology, but it can readily integrate things that are not just omic data, but other kinds of biological data that you might be collecting in your study, and it can learn those relationships in such a way to really enable more sophisticated modelling predictions. So, I think this field, the pace at which it is growing is remarkable, and so it's kind of fun watching it evolve and seeing what might be possible.
FLG: Now one aim of the Festival is to get people from industry, academia, healthcare, regulators, all in one room. Is that something that's quite important in this kind of work, and how can events like this help to facilitate that?
Tom Lanz: Oh, it is important. I mean, it's happening. We consistently have, at any given time, several academic collaborations going on. I mean, there's academic groups that are specialised in different techniques, and have different specialty areas of expertise that we need to tap into and we can learn from. Especially in the drug safety realm, we’re consistently talking to regulators. And we're educating each other on new techniques. In AI and omics, these fields are growing. There's cautious optimism that they'll be useful in transforming things in the future. But that's kind of a two-way conversation. So, we continue to have those interactions through consortium and through specific interactions that are maybe targeted to a product. Having them all in one room where we're talking about these new methods and how they can be applied, how other groups and different settings are using them, I think is useful. We can learn from each other. I think the field as a whole benefits from having this kind of cross pollination of different backgrounds, trying to understand how to harness some new technology.
FLG: You have spoken at the Festival before. What did you enjoy about the event and what has kept you coming back?
Tom Lanz: It is very technology focused. But as you mentioned, the mix of academic and industry and regulatory, it's a good mix of backgrounds interpreting these new technologies. And I think it's useful to have those conversations and to learn what some of the priorities for academics or regulators might be in the newest spatial technologies, or how they're thinking AI can be employed. And vice versa, we're sharing some examples of how we're using it right now in drug development. I think the more kinds of use cases, the more new use cases we'll see in the future. I think, with AI especially, it's only limited by our creativity. The more perspectives you have on how we could use this, the more possibilities we have that'll become reality in the future.
FLG: And is there anything that you're particularly excited about at the event this year?
Tom Lanz: AI is probably the one thing that's been changing the most dramatically and the most quickly. I am curious to see how the different vendors are using these and employing these with their novel omic methods, how different groups are thinking about using it in the omic space. So, beyond what I've seen in publications and other conferences. That's what I'm probably looking forward to most. How is everyone else reacting to this explosion of new AI methods and technologies?
FLG: And what would you say to someone who's considering attending the event for the first time?
Tom Lanz: I'd say take it all in. Try to focus on something new. It's a good opportunity to learn. Don't just go to the area that you know, explore and maybe branch out. You may learn something that is useful that you wouldn't expect. I think there's a really good diversity in the talks, so keep an open mind and view it as a learning opportunity, and certainly network.
FLG: Thank you so much for your time today. Tom, it's been a really interesting discussion, and we can't wait to hear your talk at the event.
)
)
)