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 have the opportunity to sit down with our expert speakers to find out what they'll be discussing at the event. Today, I'm here with John Tallarico, Global Head of Discovery Sciences at Novartis, who is one of our expert keynote speakers. He'll be joining a panel discussion on the morning of day two. John, how are you today?
John: I'm doing very well. Thank you. And you?
FLG: I'm doing well. Thank you. Could you start by telling us a little bit about yourself and your background?
John: My name is John Tallarico. I lead Discovery Sciences at Novartis. Discovery Sciences is a large multidisciplinary department that is very technology forward. We work across all of our indication areas, and we collaborate with all of our, what we call, disease areas, covering everything from target discovery all the way up into late lead optimization. My personal background is I'm a chemist by training, so background spans chemistry, of course, but it also includes biology, and I've become much more interested in large scale, data driven decision making, meaning large experiments that generate large amounts of data that help facilitate decision making. But really my goal is to help focus how we're translating complex human biology into high quality targets for the company. That's the vast majority of my work.
FLG: The panel discussion you're joining at The Festival is going to be focused on multimodal data, and there's a lot of excitement around that. But where have you actually seen it make a real difference so far in precision medicine or drug discovery?
John: I'll steer it towards what we spend most of our time on, which is target discovery. We're looking at multimodal data and thinking about multimodal data as a way to help develop and increase confidence in targets through the same, for example, convergent evidence. So, obviously, I'm a big believer in human genetics helping guide drug discovery and target discovery, human genetics doesn't lie. But how do we couple genetic data with functional genomics, with proteomics, with perturbation screens? And how do we use these things to triangulate causality, coupling genotype to phenotype, as well as get a nice handle on an understanding of mechanisms so that we can then prioritise these things? This multimodal data is a confidence builder, but it also does increase the amount of signal and noise that we're hearing. So, how do we most rapidly achieve high confidence in our signal?
FLG: At what point in the pipeline does multimodal data have the biggest impact? Is it target ID validation, clinical development?
John: I'm biased here because it's the area I work in, but I love the target identification area. It's the place where the decision is irreversible and expensive downstream if you get it wrong. Applying these tools and efforts to develop systems to deal with multimodal data and make signals from it, I think, is most important, because if you pick the wrong target to work on, it doesn't matter whether you can put 1000 FTEs to work on it downstream. If it's the wrong thing, the decision is irreversible. And so you really want to have high confidence at this earliest stage of research.
FLG: That makes a lot of sense. And your panel discussion is going to touch on AI generated insights as well. What's the biggest gap today between those AI generated insights and real-world R&D decisions.
John: I think one of the biggest gaps is how to translate it into action, right? So, you can get a lot of ideas, a lot of hypotheses. Again, let me take it back to target discovery. If we're trying to really, truly be unbiased in our target discovery and let the data yield the targets, one of the biggest challenges is, a priori, you don't know the nature of the targets you're going to discover. So, let's say in Parkinson's disease, you're using all these tools to try and develop novel hypotheses to find novel targets to treat Parkinson's disease. I don't know if I'm going to discover a transcription factor, an ion channel, a kinase, some other type of enzyme. So, I don't know the nature of those targets, but regardless of what those are, if there's a diverse set of them, AI is not so helpful for actually sorting out straightforward ways of testing these hypotheses and validating these things experimentally, so that we can get high confidence in the model. Usually, it requires real deep expertise in transcription factors. It requires deep expertise in kinases. It requires deep expertise in ion channels. And so it's very hard to assemble all that expertise in a single project team in order to validate what the models are telling you. And then, of course, feeding that information back into your model in the first place to help you develop yet more confidence in it is really complicated, and the right models don't help us sort that out.
FLG: What metrics or benchmarks can you use to evaluate whether an AI model is actually improving productivity?
John: Yeah, that's a good question. Productivity means a lot of different things to a lot of different people. I think as an early stage scientist, for me, productivity means higher probability of success that we're working on the right things. It's hard for me to really be quantitative about this. I guess I would go with the counterfactual, which is, if we didn't have this model, or we didn't have these tools, what would we have done? Could we have gotten to this end point without those things? And if the answer is no, then I think that's a pretty good benchmark, saying, yeah, these things have been valuable. But I wouldn't talk about specific measurements. To me, it's really more about the narrative and going back to, could we have done this without these models or without these tools?
FLG: What are some of the most common mistakes that you see people make when trying to operationalize multimodal data?
John: This is very good question. To me, the details are really important. So, the details of any particular set of experiments you're doing. With multimodal data it’s generally large scale. You're not aiming it at one little particular tiny hypothesis. You're using it to generate a lot of different hypotheses, which we want to follow up on. But of course, as we know if you look backwards at existing medicines - which is the ultimate goal, we want to make medicines - the details of how drugs work at a molecular level turn out to be really important. And I worry that sometimes we ignore this context or this assay reality, mixing all these modalities without aligning on the details of what needs to be known downstream. In biology, I think it is hard. I think translatability is also hard. We make models, you make them as simple as possible, but not too simple. I forget which famous scientist said that, but these models are only models, and we really need to worry about translatability of these models into actual human biology. I think that's a challenge as well.
FLG: And what does it take to get a cross-functional team to trust shared data and models in the space?
John: Yeah, I think it's also a good question. I think at this moment in time, there are a lot of people talking about labs in the dark, lab in the loop sort of workflows. And I know this term is very commonly used, this human in the loop workflow. It's very important at this stage to still have a human in the loop workflow so that scientists can challenge outputs. We want multimodal data and these AI systems to give you hypotheses, things to think about. But we need a human in there to perhaps make connections between different hypotheses that current systems just won't point out to us. So how do we annotate exceptions? How do we feed these human corrections back into the loop? I think that's going to be critical, at least in this current stage of building trust in these systems.
FLG: That sounds really important, remembering that human aspect of it. How do you approach data sharing, while at the same time trying to protect any competitive advantage?
John: Yeah, this is good one, because right now, I think the feeling in the world is that data is really what's going to be proprietary. And there are a lot of organisations building these tools and models that are lacking data to help refine and build and make their tools more effective. In theory, we're an organisation that would have such data, and so we're sort of doing this dance of, how much should we be sharing versus how much we should be keeping to ourselves. I think what we could do more openly, as a field, I'm not just talking about my own organisation, is really setting up standards. This would be fantastic, right? One nice aspect about the current era that we're in, and this quick evolution that's happening in the field of life sciences and AI and drug discovery, is there are a lot of organisations putting large data sets out in the open, which I think are really powerful and useful for helping fuel other teams to do discovery. So, this is a trend that we've always had in science, people putting information out in the open, letting others build on it, this constant building on the shoulders of giants. I think, right now for us, putting standards out there to make sure that we're all working with the same sort of guardrails around the data is going to be the most important thing. But right now, it's hard for us to feel good about putting data out in the open ourselves, from our own proprietary work.
FLG: As I mentioned, you're one of our keynote speakers at The Festival of Genomics, Biodata and AI. What are you excited about, and what makes it stand out compared to other events you might attend?
John: This will be the first time I'm going, so I'm really excited. And of course, I have been doing my own digging around into the history of The Festival. I think the breadth and the cross-sector mix. You know, this is not just about genomic specialists. I know we call it Festival of Genomics, but these are not the only types of scientists that are there. The agenda includes genomics, multi omics, AI, data governance. I think there's a real breadth of different attendees - academics, industry, biotech executives, pharma executives. I think it'll be great. That really creates an awesome situation for networking and cross disciplinary exchange, which really gives me a lot of energy.
FLG: And what do you hope to take away from the event, and what do you hope that other attendees take away from the event?
John: I'm always most curious at how others are thinking about these things. I use it to benchmark our own thinking. It's not about competition or anything like that, but listening to others and engaging with them really helps me learn what we might be doing well, and what we might not be doing well. And that's what I really want to bring back to my own organisation.
FLG: Well, thank you so much, John. It's been a really interesting conversation. And to everybody listening, make sure you get to the MCEC early on day two to see John's panel discussion as part of the morning keynote session.
)
)
)