Lefkofsky spoke with Paul Goldberg, editor and publisher of The Cancer Letter.
Paul Goldberg: How is the CancerLinQ deal structured? What is happening with this?
Eric Lefkofsky: As you know, ASCO formed CancerLinQ some time ago with ambition of trying to aggregate data, the real raw-source data, meaning patient records, from across a broad part of the oncology community. The goal being, if they could get these data sets that were locked inside electronic medical record systems and electronic hospital records systems, they could look for patterns that were clinically relevant, look for patterns that would improve quality of care, and look for patterns that would ultimately lead to new research.
So, they established CancerLinQ, and they were getting all this source data, which was really amazing—their penetration of the market has been extraordinary—but as the data began to accumulate, they found themselves with the challenge of what do we do with all this data?
How do we structure all this unstructured data? And the data was actually coming in so fast that I think it was a bit overwhelming. So they began to search for partners who could help them structure this data. And at the same time they began to look for partners to structure the data, it also became apparent that once the data was structured there would be multiple avenues that it could be put to use to help make it useful. Obviously, one of them is how do you improve decision support for clinicians and how do you improve research?
But there are other implications, including how do you get some of the insights from this data into the hands of biotech companies and pharmaceutical companies to make better drugs? How do you get this data into the hands of people who run clinical trials so they can run more efficient clinical trials? How do you get this data into the hands of people who make reimbursement decisions to determine which drugs to pay for?
So, they began a journey to look at a variety of companies that might help them, not only structure this data, but also help them bring it to practical use. They met with a bunch of companies, who obviously were interested, and at the end of that process they ended up selecting Tempus and a company called Precision Health AI, as their two partners to help them both structure and bring the data to market.
In terms of Precision Health AI, how does that work, in terms of your relationship with them?
EL: We are two separate companies. We actually met during this process. We were both independently working with CancerLinQ. We were both interested in helping them with this data, and we decided at some point several months ago that instead of going at it independently, we would go at it collectively, and we would both approach CancerLinQ and ASCO and say “we want to work together with you collaboratively.”
We are still two separate companies, but right now we believe it’s better to have two separate companies than one company curating and abstracting and cleansing the data and improving its use. So, they ultimately agreed with that approach and selected us as their partner.
How far is this from being monetized?
EL: CancerLinQ announced some time ago that they already have a collaboration with AstraZeneca. They have other collaborations that they have announced as well. So, it’s less about how the data is monetized… I can only tell you our approach; right?
That’s what I was asking for.
EL: This gets a bit into Tempus’s business model and at least how we view the world. This data has some commercial applicability today. But the real value in this data is in how do you build much larger, much more comprehensive datasets that actually can be used to usher in an era of precision medicine.
The data is interesting today, it absolutely has some practical application today. But these kinds of datasets at scale have yet to be used to truly impact care in the ways we all envision. What you are really thinking about when you are talking about structuring clinical data is how do you build a sustainable model that’s going to allow you to structure this data and ultimately use this data.
It’s less about how do I make money in the short term. It’s more about how do I create sustainability.
And I will tell you what I mean by that: twenty years ago, when we migrated to electronic medical record systems (big EHR and EMR systems) there was a movement to do that very quickly, in large part based on reimbursement, and what happened in the creation of these systems is that we basically took a lot of the complexity—the key phenotypic, therapeutic and outcome response data—and we just left it to the side, essentially in free text and images.
We left it inside physician progress notes, left it inside pathology reports, left it inside radiology reports, or it sat inside scans, or it sat inside pathology slides. And nobody tried to structure that data. It was just too big of a process and too expensive.
And so here we are today, with 15 million people with cancer in North America and this enormous amount of data that’s unstructured. And so what CancerLinQ set out to do and what Tempus has set out to do is to create a model that would allow you to pool data from the people who have it and begin to structure all that unstructured data in a way that would allow you to look for clinically relevant patterns or patterns that could lead to better research.
There is a cost to basically structure that data, depending on the patient record, whether it costs you $25 or $50 or $100 or $250, whatever the number is. And since there is a cost to structuring that patient’s data, you will ultimately want to create a sustainable model that allows you to do that not just for 1,000 patients, but for 10 million patients.
To me, this is more about how do you get that data structured at a large enough scale that you can get the various constituents inside the healthcare industry interested in helping you build something that’s sustainable.
And that’s the part of the journey we are all on right now; which is, okay, we need to have a million patients, or two million patients or five million patients with structured clinical data in order for people within the industry to say, “All right, I want some of that de-identified data for this purpose or some of that de-identified data for that purpose, and I am willing to pay you and help you create a data ecosystem that can scale.”
I think when we talk about commercialization or monetizing the data, and it’s still early days. We are really just starting down the journey to make these datasets valuable, and just beginning to have the earliest conversations with people to say “how do we create a new model where data flows freely, and it’s adding real value”.
How much do you think you need to be investing in this right now?
EL: I can only speak for Tempus—PH.AI, of course, has made its own investments that are significant in this space—but at Tempus, we put $130 million into Tempus, and we expect to put hundreds of millions more over time.
This is a significant endeavor for us. And it’s long-term, and one that we believe is invaluable, not just based on the impact it has on patients, but also it has value in the market, and in ushering in an era of precision medicine.
The $130 million is all of Tempus; it’s not just CancerLinQ?
EL: It’s Tempus and, as I said, Precision Health has invested lots of money as well. I don’t know the exact amount, but they have invested significant money in the formation and capitalization of PH.AI.
These are two formidable, well-capitalized businesses that are coming together as part of this collaboration, trying to make this data more valuable and improve patient outcomes.
Are you and PH.AI doing the same things or different things? Are you complementing each other?
EL: I thing we complement each other. They are more focused on machine learning and analyzing large datasets and combining these datasets with other datasets to look for patterns that are interesting.
Our focus on this has been more on combining clinical data with molecular data. We have a very strong molecular data orientation, which is why we built a lab to sequence patients. We believe that the patterns that are most interesting in cancer in the short-term are based on combinations of clinical and molecular data, predominantly, how do you amass large datasets to answer the following questions: who are these patients, how are they being treated, how are they responding to treatment, and what’s their molecular profile and composition, and can we see something in that molecular profile that we think is indicative of how a patient might respond to a given therapy.
That’s been our mission since we launched Tempus.
But as far as, do you foresee a situation where you’re sitting there and reading or having people reading records—just humans?
EL: We do that today. We have a significant operation today; we have very large teams structuring clinical data today, even unrelated to CancerLinQ, and we use a combination of optical character recognition technology and natural language processing, and humans. We have something like 120 people today abstracting and curating clinical records, and that team grows by something on the order of about 25 to 30 a month, so it’s a rapidly scaling team that curates and abstracts clinical records for cancer patients.
I see. It sounds like such an enormous thing. It’s probably bigger than all of ASCO in terms of the amounts of money that needs to be committed to this.
EL: You know, it’s not small.
You know, it’s sort of interesting how it grew as an undertaking for ASCO. It’s probably larger than the organization itself, to do it right.
EL: ASCO is a very large organization, and you have to talk to, obviously, Cliff and Kevin [Fitzpatrick, CancerLinQ CEO] about that, but I do think—I can speak to the comment you just made. I’ve been in technology for 20 years, building companies that all kind of do the same thing.
We structure unstructured messy data and try to bring technology to industries that have not had a lot of technology, whether that’s printing or logistics or manufacturing or local commerce, and I’ve never seen anything like what’s happening in health care, and in particular, in cancer care.
You have these massive technology paradigm shifts hitting oncologists and pathologists and radiologists and surgeons all at one time. One is the revolution in our ability to collect and analyze genomic and transcriptomic and proteomic data—in other words, molecular data—at very low prices relative to what they were just 10 years.
There’s been a million-fold reduction in the cost of generating genomic data in about 10 years, which is just staggering. At the same time, you have equal advancements in machine learning and artificial intelligence, especially on the image recognition side of this, impacting our ability to read pathology slides or read radiology scans and draw important clinical distinctions.
So, you have these two incredible technology movements hitting physicians that are treating cancer patients all at one time, and I do think it’s massive, and I think organizations like ASCO and their commitment to CancerLinQ and their commitment to getting ahead of this, is really extraordinary. I think it’s a model for how all associations should be thinking about how to have an impact in their respective diseases.
Is the decision support system still a possibility or was that something that was science fiction, based on CancerLinQ?
EL: I think when you have datasets that are large—at some point, we’d have to figure out and I mean, we can log in to WebEx as long as it’s confidential, and we can show you our system today, which might make sense for five minutes at some point.
I’d love to do that in a different story. I can also hop on a plane and come see you.
EL: When you see what we built, unrelated to CancerLinQ, I think it will make more sense. We have agreements now with the majority of NCI cancer centers in North America and a significant and growing percentage of the market outside NCI cancer centers. We’ve built, in a very short period of time, a large footprint where we’re both sequencing for oncologists and producing genomic data and also helping them analyze their data and make sense of it.
We attempt to use that data today to help answer a series of questions, again, largely from a molecular lens, because we tend to be more focused there. But it is real-time decision support, so you’ve got a pancreatic patient and they failed first-line therapy—and you’re considering between a second-line therapy or maybe you’re considering between Folfirinox and gemcitabine/Abraxane, or whatever the chemotherapy options are—and if we can see that if the particular molecular profile of that patient, maybe the patient has a CDKN2A mutation or some other mutation that appears based on the totality of evidence we’ve been able to collect, to be driving better responses for one chemotherapy vs. another or a targeted therapy vs. chemotherapy, or radiotherapy vs. chemotherapy—that’s data that you as a physician want to have.
Because, ultimately, there are times when the guidelines allow you as an oncologist to consider multiple therapies. There is no clear path that’s going to lead to a good outcome, so you as a doctor want to have all the data that you can have at your fingertips. What people like CancerLinQ are trying to do and what Tempus is trying to do and what PHI is trying to do, is get that data and put it in one place—in hands of people that are making those types of decisions.
I think we’re actually, in cancer, much closer to real-time true decision support than people think. I don’t think it’s a decade away; I think you’re going to see really significant changes in the field in the next couple of years.
Can you think of other research questions that you will be able to answer with CancerLinQ that you might not be able to answer quite as quickly, or at all, right now?
EL: Yes. What ends up happening is, let’s just think about immunotherapy for a minute. There are roughly 8,000 clinical trials in North America, at least there were the last time I looked at the data. Well over 1,000 of those trials were immunotherapy-based, and so, you have an enormous number of trials and you have an enormous amount of oncologists today that are looking at immunotherapy as an option, especially for metastatic patients or late-stage patients or high mortality rate patients.
There’s very little data on whether or not some of these known therapies are actually producing durable responses or non-responses. We talk all the time about MSI status or tumor mutational burden or mismatch repair genes that may be mutated or immune infiltration or HLA typing, or any one of these different markers that seem to be leading to a particular response.
The data is very sparse, so when somebody like CancerLinQ can aggregate this data from the field and start to collect insights on which patients have taken a checkpoint inhibitor or some other immunotherapy drug and had a positive response, or which patients haven’t. That information is super powerful in that it can be a really big flashlight leading the way not just to new trials, but also to a refinement of existing practice.
I think you’re going to see, as we start to cleanse and structure data at really unparalleled rates, because remember what’s also really interesting here is the technology backdrop that’s allowing us to scale. We talked about imaging and genomics, but there’s an equally powerful technology paradigm shift in our ability to store data, to structure data.
These tools are really just a few years old, where we now can use natural language processing and we can use other techniques to structure data. You’re going to have these massive new datasets that are going to arrive that are providing incredible insight that there just not cost-effective to build even ten years ago.
Maybe it cost $50 to abstract a record today whereas it would’ve cost you $500 or $5,000 five or 10 years ago. It just opens up the door to new possibilities.
