Please welcome all these brainiacs to the stage, please. Thank you.
So, obviously there's enormous interest in healthcare and AI right now to state the obvious, but in an industry with long timelines and high stakes, hype isn't enough as we were talking about earlier. So, let's talk about first and foremost what's driving investment. Renee, over the course of your career at multiple companies, you've managed capital and transactions totaling over $100 billion. So, what kind of opportunities, Renee, are you most excited about focused on right now and how are you looking at things differently because of the role AI is playing in healthcare?
Yeah, well, thank you for the question and just before I get started, thank you to Nat and Recursion for this amazing event. Thrilled to be a part of it. But in terms of how we're focused, we're a biopharma company and so we're focused on R&D, commercial, we're focused on corporate development, but at our core, we are focused on investing in innovation for patients. And we also acknowledge we're a relatively small company. We're less than 3,000 employees. And so that means we can't do it all ourselves. We invest in internal innovation as well as external innovation collaborations. We talked earlier about a collaboration with Stand Up To Cancer and those are really important for our work. With the emergence of AI, it has changed how we think about where we're investing and what we're investing in. So internally, several years ago, we made the decision to start investing in our data environment. It was more to be able to bring all of our data together, invest in data tools, and we're now realizing that having made these investments several years ago, it's putting us in an excellent position today. And then for external innovation, we're continuing to invest in collaborations, but also investing in tools with other companies that are more AI forward, AI focused. So tools that for example we can get access to in drug discovery to be able to move faster whether that's better understanding toxicology or pharmacology just getting drug candidates to the clinic faster and also investing in our people because if we invest in our people capabilities then they'll understand how to use the tools that we have. So every single Jazz employee by the end of this year will be trained at a foundational level of AI and then we have other groups that are much more advanced. So it's pretty much across the board.
Are there specific diseases that you're specific issues that you all are focused on as a company?
So we focus on primarily rare disease and then life-threatening diseases. So rare sleep, rare epilepsy, rare cancers. We just had a drug approved last year which is the first drug approved for this indication at all which is H3K27M mutant DMG which is a devastating and fatal brain cancer that there's been no meaningful innovation for in 60 years. No drug approved. I mean this is back in the time of rotary phones and there's been nothing until now. So those are the things we focus on. Reed, meanwhile, you have built and backed companies that have scaled incredibly quickly and health care, as everyone knows, is a very complicated, fragmented, slow to adapt industry at times and slow to change. So, what convinced you that this is a place where transformational change is possible and the time is now?
So as probably you and a number of people know I am by default a software technologist but it goes from LinkedIn, you know, Airbnb, a whole bunch of other things and it's entirely because of AI. So what happened is I was sitting with my partners at Greylock and I was saying look there's going to be a bunch of coding agents there going to be a bunch of productivity but there's a bunch of areas that I actually think that the classic Silicon Valley software technology are massively transformational that is extremely important to do that won't be the usual oh just CS will solve the whole problem but it can be a really strong amplifier and you know I had the fortune of being you know kind of leading the first commercial investment in OpenAI and being on the board and other things. And so I was seeing what was coming and so I flew out to New York, had dinner with Siddhartha Mukherjee who's here in the audience and said, 'Hey, I've got this idea.' And he said, 'Oh, well that idea is better than you think, right? I've done a bunch of stuff in cardio and all the rest. And here's kind of the way to think about it.' And we said, 'Okay, let's sit down.' And we looked at kind of thin slicing the entire process by which you start with an idea. You kind of have a target. You have kind of what how the whole drug discovery process works all the way to clinical trials and FDA. And it was like okay which places could AI be at minimum a 10x preferably 100 plus x if you kind of reinvent the process. And to kind of give a taste of that also in software like we've announced the deal with Schrödinger because part of what my theory is these large language models are great. They have a large compression of a bunch of different information but actually in fact when you begin to put in the kind of interesting constraints about how you look at it. So for example, Schrödinger brings you know physics and chemistry knowledge and that should be part of the model you're training the space you're exploring as part of what you're doing in order to create an AI generative factory. So fundamentally I'm a software guy and I'm here as an AI guy, right? And the rest of it all just strikes me as kind of a little bit of a horror show from the tech side because it's kind of like the how does that FDA process work? Oh, right. But you know, Sid, Sid covers all that stuff very well.
And Sid is here and he's a friend of mine and he wrote The Emperor of All Maladies, won a Pulitzer Prize and is a brilliant scientist. So, yay, Sid. But are there any particular So, you're kind of combining your brain power. Are there any specific cancers that you're focused on? Are you doing these models to address anything specific or are you kind of throwing out a wide net?
So part of it is if you actually start with a software as a platform and you think about platforms as a way of doing it, we actually already and even our early prototyping have got some interesting INDs and other things just coming out of hey let's try to build the model this way. Let's factor in the data this way. So we're already seeing some pretty interesting early promising signals. Now and all this path is you know everyone else knows better than I do it's a long path but the basic idea is say when you have a target not say ah Eureka I have one idea that I'm going to go spend 10 years on it works it's how do you have AI help you actually force rank targets vis-a-vis feasibility give you lots of variation of answers that you then cycle on very quickly you use the data process by which you're going into them all the way into the clinical trials to refactor your models to make the generativity a lot better for your specific therapeutic cases. So yes, we have some specifics, but it's actually more of a massive acceleration factory is what we're targeting.
And this is Menashe, your company. How much of your time are you spending on this?
Well, you know, I spend call it a day or two a week. And you know, Sid, who's the CEO, does the usual startup CEO thing, which is eight days a week, you know.
Meanwhile, Eric Lefkofsky. I suddenly feel like I'm on Severance, by the way. Tempus was the outgrowth of a very personal experience for you. I think that's how a lot of people really get interested in this field. Your wife was diagnosed with breast cancer in 2014 and I understand you grew extremely frustrated at the time with the lack of data that was being leveraged for her care. Can you talk about what you observed, what your reaction was and how that led you to start Tempus?
Yeah, I mean I think in my case I also came from technology and never like probably like Reed never thought in a million years I would get into healthcare or be in this space. But when she was diagnosed with breast cancer we had kind of tremendous access to some of the best care in the world. And yet in accessing that care I felt like I was being teleported back into time. Like I would drive my Tesla and pull out my iPhone and then go into this hospital where it felt like I was going back in time. And it just seemed maddening that we were giving, you know, truck drivers trying to figure out where to move pallets of water more technology than we're giving physicians who are making life and death decisions. And so at the time I really got fixated on this very small sliver of the problem which was genomic profiling had just kind of gotten to some scale. This is 11 years ago and it was very hard to understand what to do with all this information. And so the tests themselves weren't helping physicians point patients in the right direction. And we started thinking well why can't we pull data out of the EHR and contextualize it so you can marry it up with these genomic results? And that was harder than we thought, but here we are now some 10 years later and that's what we do for a living across about 5,000 hospitals in the United States and now touching almost one of every two cancer patients in the United States. We're amassing a lot of data to help physicians contextualize these unbelievably complex now molecular reports so they can make sure their patients on the right therapy.
Especially when they don't have a ton of time, right, to deal with every patient and to be able to synthesize that quickly and effectively for people is so critically important. And you describe what you do as a flywheel, a powerful data flywheel. Can you explain that to me?
Well, I mean, I think first of all, yes, you're right. About 85% of all cancer patients are treated in the community setting where doctors are bouncing between breast cancer and colorectal cancer. But long before AI was in vogue, we were using natural language processing and optical character recognition and all kinds of just general machine learning to try to create an environment that was self-learning. Even the best of kind of targeted insights only work on a small population typically. So you might have, for example, if I'm a non-small cell lung cancer patient and I'm EGFR positive, I might have 20 or 30% of patients that have a durable response and I'll have 20 or 30 that have no response basically in some group in between. And many targeted therapies work like that. And so it seemed crazy to us that you wouldn't be collecting data to eventually figure out how to refine those insights. And so we've just built this giant self-learning machine that collects longitudinal clinical data again on about half the United States that has cancer and it's become like almost 500 petabytes of data. It's a huge amount of data and we're sequencing as many patients as we can and then trying to figure out over time how to move from really targeted therapeutics to precision medicine which can only exist if you know who the patient is.
How come only half? Why not everyone?
Well, to connect to hospitals is miserable. So you have to convince a hospital to connect to you. You have to convince them to give you their data. You have to get through legal, which takes forever. You have to get through IT, which makes legal seem fast. And then you have to, you know, you build these connections. And so there's roughly 8,000 hospitals in the United States. So the fact that we've done that for five out of every eight hospitals is kind of crazy. And we'll eventually get to all 8,000.
When you think about all the information we don't know and that doctors don't have access to other than like helping them with longitudinal studies that will kind of guide how they give care or whatever. Are there other areas of technology that you feel are ripe for developing that will help a patient population?
Well, I think leaving aside all the benefits of genomic profiling which is extraordinary because molecules are responsible for most disease. One of the single biggest challenges we have in this particular system is error, mistake and waste which candidly probably accounts for maybe a trillion or a trillion and a half dollars of the problem.
Error, mistake and waste essentially not that anyone wants to make a mistake but that they make a mistake. And that mistake can exist on many many levels. And so one of the benefits of being connected to so many hospitals is we're basically reading both structured text and unstructured text in real time. So physician progress notes, pathology reports, radiology reports, and we have technology that essentially flags these mistakes as they occur. No one wants to make a mistake. But the system is so overridden with demand relative to supply that we just make errors all the time. We don't order the right diagnostic. We don't get the right result. We don't make sure the patient's on the right drug. We fail to note that this drug has this adverse event and oh my god that patient five years ago had this. So I believe over time one of the best benefits of these large language models and large multimodal models is that they will create an infrastructure to ensure that that doesn't occur. And if you look at the money that's made in the system we spend all our time complaining about drug profits or payers I think the total drug profits are like 125 billion and payers make maybe 50 billion. So when you think about a $6 trillion system, mistake, error, waste, that's where all the money is.
Meanwhile, Eric Topol. I love Dr. Topol.
I set him up for that, by the way. I gave that.
I mean, let's talk about I mean, you've been practicing medicine for how many years, Dr. Topol?
Yeah. And you've seen this evolution of medicine just unfold before your eyes. And as somebody who's been in the biz, as they say, for a very long time, I just would love to hear your perspective on how the practice of medicine and healthcare at large has changed.
Well, Eric did set up the problem we have that the medical community doesn't like to admit, which is rife with medical errors, diagnostic errors especially, but also treatment errors. But the way things have evolved, unfortunately, Katie is that we've seen this erosion of the patient doctor relationship. And you alluded to it, not enough time. We don't have this gift of time. Physicians are squeezed. All clinicians are heading towards a burnout and disenchantment why they went into the profession and that's been a steady erosion that was exacerbated with electronic health records and various reimbursement tactics. But we're starting to see this flip and that's what's so exciting right now because we're seeing for the first time this is where AI the near term is really kicking in these ambient conversations that are not just made into better notes than what exist in electronic records today but also they drive all the other things like pre-authorization, prescriptions, next visits, labs and even nudge the patient subsequent to the visit about things that were discussed. So this is giving the gift of time getting away from data clerk function which has become such a burden and that's the first step of AI starting to change the way medicine's practiced and starting to restore this critical essential patient doctor relationship.
What when you think about AI and medicine, you know, you're saying we're sort of turning a corner. What are you most worried about?
Yeah. Well, there's plenty. There's a long list that AI has.
Well, we have a fair amount of time, so go for it.
No, no, no. But it's a net benefit, but the worries include not just things like privacy, security, security of the data. The biggest worry to change the medical community, you have to have compelling data. And there hasn't been enough diligent effort to get that data. And sometimes that's randomized trials. Sometimes that's prospective cohorts, but that's really important. But we have issues with regulatory, not knowing how to handle this stuff, reimbursement, transparency of the companies that are doing this work, because they tend to keep things in a very opaque, proprietary locked up way. So there's a long list, but I think the hallucination confabulation story is one that there will be a net benefit. People held clinicians accountable for their errors. They don't realize that that rate is quite high and even though AI may make some mistakes along the way, it's probably got a better chance of bringing that net mistake rate way down from where it is today. So, I think we have to look at that overriding benefit. And what we're seeing now, which is so exciting, is this big jump in accuracy. And that comes with medical images. It could be any type of medical scan. And interpretation of virtually any diagnostic image has soared with particularly mammography and colonoscopy. They've been the subject of massive clinical trials, randomized trials. So we are starting to see that kind of compelling evidence that's going to be requisite to change medicine to make it a whole lot better. And if we get to the point where there's not a lot of data burden, data clerk burden for clinicians, those two things are kind of in the immediate phase that should make a big difference.
Do you all have any thoughts on what Dr. Topol is talking about about some of the challenges and how they're going to be addressed for clinicians and for even scientists and researchers?
Anybody?
Yeah. I mean, I think it's a very real challenge. I mean, there's no question that the winning answer, I think, is not AI as a replacement to physicians or to nurses, care teams in every form and facet, but as an enhancement to the work they do. And I think one of the challenges we have is that this system has just not been very embracing of technology for a very long time. And the technologists have kind of stayed away from it for all kinds of reasons, regulatory complexity, all kinds of stuff. Now this marriage, this shotgun wedding is occurring and it can't not. And so it's going to take I think it's going to take people and maybe Reed and I are two of them that can build some kind of bridge because we're going to live in both worlds. We understand how systems get built and where technology matters and doesn't matter. But we also spend the time to figure out the nature of health care and its complexity. And ultimately if we don't find a way to kind of marry these two the tsunami that's coming is that patients aren't going to kind of sit on the sideline. They will go into ChatGPT or Claude or Gemini or whatever. They'll ask increasingly complex questions and they will show up unbelievably knowledgeable about the disease they have and you're now going to have this crazy imbalance where physicians aren't armed with those same tools. And so I think we just need to kind of catch up fast.
There's two things I'd add. So, one is I think it's a nearly certain future that since we already have frontier models, I suspect everybody on this phone has a frontier model on their phone. And already today, if you're not using your frontier model of choice or frontier models of choice as a second opinion or anything serious, you're making a mistake either as a patient or as a physician. It's a quick easy cross check, right? Then part of it is say for example you're a person who doesn't have access to elite medical care doesn't have a concierge doctor they can call at any hour of the day and get that they say I've got some concern myself my spouse my child my parent you can talk through it part of the future that I see in this is it'll say oh I think you might have something serious maybe you want to consider going to the emergency room or a medical clinic and then do you mind if I share a precis of the conversation we just had with the physician with the nurse to say okay these are the questions I asked these are the answers I got here's what I think is possibly going on think about the acceleration of the aid that enables and it's 24 by 7 we cannot afford 24 by 7 medical care for other than a very small number of people that's the kind of acceleration that we are capable of already today we just need the human systems to function the right way and regulatory is part you know weird ass medical liability is another one etc etc.
I'm curious about the role of the government in officiating the so-called shotgun wedding or I mean what is the government doing or what should the government be doing to sort of accelerate and elevate artificial intelligence in drug development and medical settings and a whole host of arenas. Are they doing enough?
I mean, Marty Makary and FDA have made it very strong, known that they're going to be trying to accelerate approvals, whether that's 510K or actual approvals for algorithms that are used in medicine. And that's also involving patient, as Eric alluded to. I mean, just this past week, we saw ChatGPT Health and Claude Health come out as well as Doctronic Utah getting refills with AI with no doctor. So, we're seeing things that without the government that is the companies know that this convergence is not only inevitable, but it's probably the most important application of AI across all the different domains.
Well, what's really happening is the government isn't in the way and we're starting to see some things that are being done to foster this. Some will argue that they're not doing enough to put guardrails on it, but that's not what this administration looks like right now.
What about you, Renee? Do you have any thoughts on that?
Yeah, I would say I think the FDA is reacting to strong data in terms of different accelerated approval paths that in particular in oncology leaning into things like real-time oncology review where you can send them your data right away and they can start processing it so that the timelines can accelerate because ultimately every moment matters for these patients. We have to accelerate, but we also need to see more data innovation at the FDA because we can all take advantage of all of these tools, but if the FDA can't keep up and they need to ensure that they are updating all of their systems and they can take advantage of these tools as well, then they will end up being somewhat of a bottleneck.
I'm curious about access. You know, the medical system in this country, as you were mentioning, Reed, you know, only a handful of people can afford like the super charged, you know, AI-driven medical care. And I'm curious to hear your thoughts on how this is going to make health care more accessible to so many people who either are underinsured or lack insurance or can't go to the fanciest doctors and the most expensive hospitals or most Americans in other words. How do you see it impacting what I find is really upsetting inequities in healthcare in this country? Really it's basically a caste system.
