Precision Medicine Is (Almost) Here | Tempus AI CEO Eric Lefkosky

Thanks, thanks for having me.

I mean, I think most of what I learned in the first part of my career in technology has found its way into Tempus. I had started four companies before Tempus that all got to some scale. And all four of those businesses were really focused on applying technology to an antiquated industry or kind of a non-technology industry. And you know, one of them was applying technology to printing, and one of them was logistics, and one of them was media buying, and one of them was pizza parlors and burger shops and spas. And so I had developed this kind of proficiency for understanding how industries that don't have a lot of technology could be reinvented with technology. And I used to say to people, the only two industries I'll never go into are healthcare and banking. But about 11 years ago, my wife is diagnosed with breast cancer. And so I found myself in the middle of that journey with her, and it was just kind of maddening how little technology was a part of her care, how little data there was. And so I kind of did a left turn and stopped doing everything I was doing and focused on bringing technology to cancer care and diagnostics. But almost everything that I learned in that early part of my career is a part of Tempus. I mean, you know, Tempus had its heart as a tech company. We think of ourselves as a tech company. We have the DNA of a tech company. And so we have 700 software engineers, 100 folks in AI, 400 PhDs that are working on these complicated systems. We spend a lot of time thinking about how to build software and solutions. So all that came with me. The part that clearly doesn't come with you if you have a tech background and you enter healthcare is you kind of, I think, have to really appreciate the complexity, the legal complexity, the regulatory complexity, the kind of scientific complexity of healthcare, and how slow it can move at times, how insane it can seem at times, how insane it is at times. You have to really digest that, otherwise you're going to be spinning.

I mean, I've said this before. So like, if you had said to me as a guy who spent building tech companies for a long time, if you said to me, I want you to design a system that at its heart would impede progress, design an industry and then see if you could kill all progress, like innately within the architecture of the industry, you would almost create this industry. You could imagine all the characters. Like literally, you'd be like, we'll create a world where people run these really small randomized trials and we'll silo data and money and no one will share anything and no one will know if anything's really true and we'll create layers of payers and we'll lock all the data in these massive EHR systems. I mean, just all of it, you know, and the government will pay different prices for different things. You'll never know why or how and it'll take forever to get paid. And every once in a while they'll stop paying you and then maybe they'll try to get their money back and just all of it, you know, you'd create this system of just total craziness. So, I don't know how we got here. I mean, at its heart, the best part of the US economy and the best part of the US healthcare system is that it's a free system. It encourages innovation and ideas and it is the only real playground at scale where you can build a Tempus from scratch. But other than that freedom, it's a really, really broken industry. And I think it just starts with the fact that we spend something like $6 trillion a year and don't produce outcomes that are really any better than the rest of the world despite the fact that we spend, I think, as much as something like the next largest 18 or 19 largest countries in the world.

Yeah. So when my wife was diagnosed with cancer, one of the first things that we did was we had her sequenced, which at the time was kind of not that common. Next generation sequencing 11 years ago, you said, right? Sequencing was really kind of invented when we sequenced the human genome some 22 or 23 years ago, 25 years ago now maybe. And it was very expensive in the first decade, but by the time she was diagnosed, it had come down in price and some companies were gaining a little bit of scale. Foundation Medicine maybe had just gone public or something like that. So they were just gaining a little bit of scale. And so we had her sequenced and I remember when we had her sequenced my first reaction was this report that's coming back that we're all waiting for and relying on as this critical piece of information has no idea who she is. It doesn't know if she's a man or woman. Doesn't know if she's old or young. Doesn't know what drug she's taken. Doesn't know what clinical trial she's actually eligible for. And I thought it's crazy that these things aren't connected. You know what I mean? Like why would you not be able to kind of take whatever clinical data existed about her and feed that to the company that was doing the sequencing so that when the report came back it was personalized. The joke of the whole thing is that they'd call this personalized medicine and the only thing it knew about my wife is she wasn't a monkey or a cat or a dog. It knew she was human. Literally beyond that it knew nothing else. And I thought that's really more targeted than personalized. So someone should try to fix that. So when we started Tempus, or when I started Tempus 10 years ago, it was really designed around could you pull data out of these electronic healthcare record systems or EHRs or EMRs. Could you pull that data out and find a way to package it up for these sequencing companies so that they could put the two together? And that's what we thought the business was going to be back then. We help these big sequencing companies contextualize these reports and really try to make precision medicine a reality. Unfortunately, the sequencing companies back then wouldn't give us their data. We had to do the marrying up of these two things and they wouldn't give us their data. So we had to open up a lab. And so about maybe nine years ago we opened up a lab and began sequencing patients and now fast forward almost a decade and we become the largest sequencer of cancer patients I think in the United States. So we sequence a lot of patients and we generate these very personalized reports. So we don't recommend drugs a patient's taken in a prior line of therapy and failed. We don't recommend clinical trials they're not eligible for. We try to really contextualize the molecular findings and help get them on the right path. And so that's what Tempus started doing. And the only kind of change is that as we began building the technologies to do that, we realized that we could not only make cancer diagnostics or cancer sequencing smart, we could make all diagnostics smart. You know, a blood test, a CAT scan, an MRI, an EKG. And so probably maybe four or five years ago, we began entering other disease areas like cardiology and neuropsych. And so today we're broad.

I mean, it was the hardest part of getting this thing off the ground. We would go to people and say, we want to, you know, sequence your patients, but we want these tests to be intelligent. We want to create a self-learning system. And so, if you don't give us the clinical data, we can't get smarter. And if we can't get smarter, then there's no point working. And at the time, that was a very hard sell 10 years ago. I mean, you know, you'd go talk to five hospitals and four of them would say, you know, I'm not interested. But it just, you know, there was an inflection point where some people said yes. You know, you have early adopters and we start delivering these reports and doctors are like, oh my god, these things are great and they're actually helping me and they're helping patients and it's like a snowball and at some point you get more and more. Today we have, I want to say, more than 5,000 hospitals in the United States connected to us giving us data. We have data from more than half of all cancer patients in the United States. Like all or most of their data and the database is so big the way we measure these databases is we say in petabytes. And so we have over 450 petabytes of data, which is an enormous amount of data. I think we have something like 50 million patients worth of data in totality. We have maybe approaching 10 million in cancer alone and millions with unbelievably rich multimodal data meaning we have clinical data and imaging data and sequencing data and all that. And so now the number one question I get asked when I go to hospitals, including I think we're the largest sequencer across the NCI cancer centers, is can you help me license this data? Can I make money off of this data? It's literally 10 years ago, it's no way am I giving you my data. And now it's how do I make money off my data.

You know, maybe we're, I would say 10 years ago we were a zero or whatever, a one and I think we're probably today about a four. Interesting. I think the challenge is you can only get to about a six or seven and then you hit a wall because when we talk about the totality of a patient's healthcare data being sharable, the totality of their data is not in Epic nor is it in Cerner. If Epic wanted to give you everything they had, they won't get to more than like 60 or 70% of the totality of the data. For example, just look at Tempus. So they don't have our data, right? So if we sequence a patient, they don't have those BAM files. To the extent we digitize an H&E, they don't have the digitized pathology slide. There are all kinds of radiology images that aren't stored in Epic. They're stored in different PACS systems. So patients order tests outside the hospital system that they go to genomic providers or whatever. They got a function now, function, yeah, I was going to say 23andMe, but not say all this data is in different places. So, you know, so one of the things we've got to do and I think the big LLM, the kind of frontier model folks are trying to tackle this, although I think they're going to run into walls. But one of the things that we certainly are going to need to do over the next 5 years is make it so that all that data can be uploaded into these large models so that patients can query the totality of their data and so can physicians.

Well, I think the businesses that are most interesting are businesses where somebody has access to vast amounts of proprietary data. They can leverage these large multimodal models to generate insights and they own the distribution of the insight back to the client. So this notion that somehow the people that make the models, by the way all of which become largely commoditized within 6 to 12 months, are going to displace all the people that have proprietary data and proprietary distribution which a lot of these software companies have. I mean, Salesforce has enormous amounts of data and connectivity is probably way over. It's, you know, one could argue that the displacement could easily be the other way where people like Salesforce might displace the need for some of these frontier model folks that are basically generating a lot of revenue because the fact right now that's where you have to go to get some of these applications. So, I don't know which way this goes. It's not so clear to me that it's going to be Anthropic eating Salesforce and Adobe and whatever else. It could easily go in other directions. It's too new. The scale of the growth of the companies that have built these models and the scale of the investments into developing these models is so extreme that it makes it very hard for bystanders watching the space because you know, I've been building tech companies since basically 99. So I've seen the birth of the internet, the birth of mobile, the birth of social and so you could watch these things. They kind of moved slower, right? The insanity of the internet in 1999 was hard to wrap your head around, but it was nothing like this, right? This would be like compacting maybe 1994 to the year 2000 into like three months.

Yeah. So, it's just very hard to watch. But I don't think, I think there's going to be a lot of overreaction in this period every which direction. But I think long-term the value will reside with people that have proprietary data and proprietary distribution. I think the notion of interoperability is going to change dramatically from asking people like Epic or Yale to please send someone my data to empowering patients with applications where they have their data and it will be instantaneous for them to send it to somebody. So, for example, if I've loaded all my data into Chat because I love Chat and I'm using Chat Health and that's what I'm doing, I think it'll be quite easy for me to ask Chat to send my data somewhere else or the data will reside within a device I have and I can do it myself. So, I think interoperability will kind of go away. It'll be like what you'll ask somebody in 20 years what that word means and as it relates to healthcare data, they'll be like what are you talking about? Like it's my data and I send it anywhere I want and it's in my eyes.

Sorry, but one of the, I've been saying I'm a little bit like the boy who cried wolf. I started telling big hospital systems seven or eight years ago this was coming. We didn't refer to it at that time as foundation models or large language models. Back then we referred to it as machine learning or AI but we would say to people this is coming. The most profound part of it is that really for the totality of medical care delivery the physician had expertise the patient had none and so the patient was largely placing all their faith in the hands of the physician. But what's happening now is patients are basically uploading their data into Claude or Chat or Gemini or whatever and they're showing up in doctor's offices unbelievably educated as if the top 100 doctors in the world spent a month looking at their case and had three things to say and doctors aren't prepared for it. So I think that's a whole another profound change.

Well, I think in healthcare, I feel like that has, I feel having come from consumer and dealt with those privacies and data issues into healthcare. It's one of, I think healthcare actually did this remarkably well. It's one of the few times you can actually credit the government for like good legislation. I think HIPAA basically works and I think at the end of the day, we've resolved that issue. We've said healthcare data singularly belongs, the ownership entitled to that data is with the patient and no one can ever take that away. That said, copies of that data can also be owned by institutions that hold that data and they're free to do whatever they want with it so long as they de-identify. They can't essentially commercialize identified data. So I think that's probably the way. Now that same kind of protection doesn't exist on the internet, right? But it does exist in healthcare. So I think that's the way it's going to be. Now there's some states or some regulation may try to block companies from commercializing that data given that it's owned by the patient. But I think it gets complicated because in order to really block it you have to kind of do away with HIPAA and this notion of de-identification. And once you do away with that you grind the whole system to a halt. You know we're in the very early innings of this. And it's not just that we're doing this. So I think this is also the work that people like Anthropic and OpenAI are doing and others. I think you're going to get very powerful tools for patients in the next several years. I'd said to Google years ago when I was out there, you know, I suspect this will hit everybody including them. You can imagine a world where you have three kinds of searches, you know, normal search, incognito, and fully identified. Fully identified would be like I'm going to upload all of my stuff to you and when I search on Google I want you to know everything about me like you know and that's super powerful for example in the case of medical records where I don't want to ask you a question and you not know what drug I'm on or not know what's going on with me I want you to actually give me the best result.

I think AI is already unlocking and will completely unlock the relationship a patient has with a technology company and their physician. In other words, historically it was patient direct to physician. It's 100% going to be like literally 100% of the time patient to technology to physician or it's going to be a triad now. It's going to be a triad. Yeah, it's an AI triangle that already exists at scale and will exist seamlessly all the time and it's actually one of the biggest parts of the problem. And so whereas the system is entirely broken today, I'm hopeful that it can be fixed because AI is designed to do the thing that the system does so badly today. And what AI can do is essentially look at multiple variables and kind of tell you the right path. No different than GPS tells you where to go or radar tells you what weather to avoid. And we don't do that in healthcare. We give people the wrong drugs and the wrong dose at the wrong time all the time. And when you look at the big pockets of money, you know, we talk about like we got to go after drug companies or we got to go after payers, you know, I think the total drug company profits like 125 billion. And maybe payers are like 75 billion. So it's just minuscule in a $6 trillion system. If we passed a law tomorrow that no drug company can make a penny and no payer can make a penny, we're still stuck with a $5.8 trillion problem. The biggest chunk of the spend that can be removed is error and waste. Essentially the wrong next step and I think AI will do a great job at getting at that. And so there you're going to have this relationship. Does that mean we're going to dispense care without physicians? I don't think so for a long time. Does it mean that companies can skirt the basic payment infrastructure of the US healthcare system? I don't think so. For a long time, the business models that have gotten to some scale are kind of living in these small pockets of the system. Like, you know, oh yeah, like we, it's not hard to prescribe us, you know, erectile dysfunction drug or a GLP-1 and so like I'll build a massive business doing this little thing, but it's so big I can actually get to some scale. But it's like a very, very, very tiny sliver of the totality of dispensing care. So I don't think there's any real danger of like a new paradigm of how patients basically get medical care without physicians. Not for a long time.

Yeah, I mean it's probably almost entirely that. I mean, you know, we saw this with sequencing 10 years ago. There was a significant percentage of oncologists when we started who thought sequencing was a waste and that it didn't matter and that it was like dumb and it was almost entirely correlated with age. You know, we sequenced the human genome again maybe 25 years ago. So if you're a 60-year-old physician or 65-year-old physician, you graduated medical school before we even sequenced the human genome. So I mean it's just and so what kind of wasn't taught or you didn't really need to understand and now is commonly taught. So I think with the next generation coming on they will embrace these tools with much less skepticism. The other challenge you have in terms of any industry, this holds true in healthcare or any other industry, is that typically the people that are the most successful, which often can be, you know, they're older in their career, they've spent more time, are resistant to change because they've essentially achieved this high degree of success by doing something. And so now you're going to walk in and be like, okay, everything that you've been doing for 25 years that's made you successful no longer works. It's not an easy sell. You know, they're very resistant to that. So I think it just takes time but AI is coming and it can't be stopped and physicians are going to be accessing these kind of technologies, you know, I would say in the next 5 years routinely throughout their day and in 10 years it will be entirely embedded in their workflow and within 20 years physicians won't know how to dispense medical care without AI. It would literally be like a pilot flying with a radar.

Tempus has gone further and I think had more success in bringing AI to the clinic than anyone. And so, even though we're not enormous, we at least I think are a roadmap for what that success looks like. And I think the roadmap is you need the great tech companies to embrace the complexity of the medical community and not try and shove their solution kind of down the throats of the medical community. You have to build it with them. They have to be part of the solution and part of the journey. And I think too often if you look at the companies that have failed, they've walked in and said, you know, I'm Microsoft or I'm Apple or I'm whoever Google and I obviously know how the world works because I'm dominating the world and so you need to do what I tell you. And I have had this conversation with some of the largest tech companies in the world because they also find it very confusing when they walk into these places. They're so used to everyone like wanting to be in their world and take their money and whatever. And you know, you can't walk into Yale or Harvard or MIT and they just don't care. They've been around for hundreds of years and they're there to preserve this institution and they're like, yeah, the fact that you're from Google or Facebook means nothing to me. And so you have to kind of, I think, and but we've had a lot of success at trying to build that bridge. And that's also why we have a very large medical team, a very large team of PhDs and scientists. And so we spend a lot of time honoring the complexity of biology and chemistry.

Well, I think the big AI frontier model folks have become convinced that one of the best use cases for their models is healthcare. That if you want someone to pay you money for in perpetuity to access something that's smart, if that smart thing can tell you whether or not you're on the right antidepressant or whether or not you should take 5 milligrams or 10 milligrams of Crestor, that's a really powerful use case to keep you connected and engaged because once you upload all your data and have that kind of relationship, you're unlikely to start over again with another model. So I think one of the big frontier folks will seek to crack that code and I suspect that they will get far enough like ChatGPT 1 or whatever that it causes a panic in the rest of the players to become fast followers. And what you saw, I think, with OpenAI and Google is probably the best example of that, which is, you know, Google was obviously incredibly protective of its search engine, which effectively rendered ads in the most critical real estate of the homepage below the search box until OpenAI showed up and said like, I'm just going to give you the answer. And now Google's like, okay, move everything down. Now, by it's all worked out fine. But I think if I would have said to you in Google's headquarters, we're going to do that five years ago, you'd have fired me. That's the single dumbest idea ever. Go work on Tempus. Yeah, go work. So, I think it's going to be that somebody is bold enough to say, I'm not afraid. And because the real thing that holds all the big tech companies back from being successful in this space is they're so frozen by the privacy and data issue that they're incapable of taking a step. And I'm not saying it's wrong. I'm just saying I think Google's like I don't want to hold two billion patients worth of PHI. You know, I don't want to even hold that. And so I think one day they're going to realize they have to and they will. But until they have to, I think there's too many lawyers show up not to attack the law.

Yeah, it's going to happen. I mean, I think it's just unstoppable at this point. So, it's just a matter of when and how and again, I think they're, you know, one year ago, two years ago, I couldn't get any, we're the largest licenser of de-identified cancer data in the world. I mean, we think we signed like two or three billion dollars worth of licensing deals to have our data out there to advance drug discovery and development. And a year ago, you know, I couldn't get any big tech company to like even engage. Their eyes would gloss over. And now we're in conversations with half of them and my guess is a year from now all of them to try to as they're trying to understand what data they would need to build models. And so I think it's going to be a bit of a lab.

In oncology or in cancer, I think we're a six or seven heading to a 9 or 10. It's a function of time and money. It'll happen. There's nothing structural that needs to happen. It will happen. Tempus and other companies like Tempus will get there. You know, it just takes time to, you know, it's a complex system. There's lots of drugs and lots of heterogeneity and lots of response and lots of adverse events and lots of everything. It's have to build tools that make sense of all it. But we will get there. In the next decade, we'll be a true precision medicine in oncology. Outside of oncology, we're a zero to one everywhere.

Zero to one.

Zero to one. So, the biggest unlock is in order to have precision medicine, you know, all effectively and by the way the fact that you have very good targeted therapies or advancements in surgery or even novel drugs is great but if you want true precision medicine that by definition that means I understand what are the unique attributes of this patient's disease and I'm going to target or tailor my therapies based on that. So given that we are made up of molecules, almost all disease is molecular in orientation, right? Like something's going on that's causing you to have this disease. And the challenge with most therapeutics is they work some of the time, not all the time. They work for different durations. And the answer to that question is going to be molecular in nature. We will know why. You can take a drug and be on it for three years and somebody else takes a drug and they're only on it for three months. Is it something about your DNA, your RNA, your epigenome, like the things that surround that, the environment? Is it something at a protein level? Is it your immune microenvironment? Like what's going on that's driving it? All those questions are essentially sequencing oriented or molecularly oriented. The challenge outside of oncology is that when Nixon declared a war on cancer and when we set up the NCI as a branch of NIH, we began funding a bunch of cancer research. And so when we began sequencing patients, one of the first kind of avenues of all that data or the first outlets was in cancer. And so you just have a couple of decades now of all these drugs that have shown up that treat EGFR or ALK or whatever. You don't have that in other disease areas. So we're in this weird kind of chicken and an egg. We didn't fund those disease areas with lots of research that turned into lots of drugs that needed a biomarker that mandated the need for sequencing which someone would pay for to usher in precision medicine. We just were stuck. And so if you'd have said to me 10 years ago, like 10 years from now, where will we be in terms of sequencing in cardiology and Alzheimer's disease and inflammatory bowel syndrome? Like I would have been like, oh my god, we'll sequence half a million people a year, not 12 or whatever. And so that's the big miss. And I don't know how we, I don't have an answer for how that unlocks it. It unlocked recently in rare in pediatrics. It was the first non-cancer moment where sequencing was done. It was generating some insights and we got Medicare to pay for it at a rate that was sufficient to propel a diagnostic company to get to scale which has allowed the leader in that space they're called GeneDx to get to some scale. And it's a really awesome moment because it at least gives you the playbook for how this can happen in other disease areas. So we have to kind of seed some research, get some momentum in cardiology or whatever. And then I think hopefully that's how it happens. But if the government isn't paying for these kind of or private payers aren't paying for this diagnostic data, it doesn't get generated.

It gets back to my last comment, which is you have to create the economic incentive to usher in the change you want. So if I was the head of HHS or if I was basically running the system, I would find a way to responsibly pay for AI. And so I would now, I would probably force AI companies like Tempus to go through the FDA. I think it's a very good gate. I would provide provisional reimbursement for the AI. Like you tell me what it's going to do. Tell me it works. Prove to me it works and it's effective. It's safe. Tell me what you think it's going to do in terms of outcome response, value based care, whatever. I'm going to pay for it for two or three years and then I'm going to make you prove to me doing some real world study that it did the intended thing. But in that ecosystem, I would try to really invest a ton of capital in getting AI companies to solve this core problem of waste and mistake and error. And I would pay for it. And that's the core problem I think the government has to find a way to solve. And in fact, it's going the other way by the way. The AMA because they don't like AI. They're going the other way. They're like trying to make it harder for these kind of technologies to get in the market.

Yeah, they have a framework for it by the way. This was the framework people used to get FDA approval for Foundation. Like I didn't invent this. It's just a regurgitation of a program that existed. It's gone now. But for a very short period of time, they bundled FDA approval with CMS reimbursement to encourage labs that didn't need FDA approval to get FDA approval. And Foundation Medicine took advantage of it. It's the first company in our space that got FDA approval. It came with payment and that led the way to all of us.

Thanks for having me.