Jonathan D. Grinstein, PhD, the North American Editor of Inside Precision Medicine, hosts a new series called Behind the Breakthroughs that features the people shaping the future of medicine. With each episode, Jonathan gives listeners access to their motivational tales and visions for this emerging, game-changing field.
For years, Tempus has operated with an air of mystery. Their website provided little detail about their work developing AI-enabled solutions for personalized healthcare. Yet headlines frequently announced massive funding rounds and high-profile partnerships. Beyond that, details were sparse—a fact that drove my curiosity to no end.
Today, we pull back the curtain. I had the privilege of speaking with Dr. Kate Sasser, Chief Scientific Officer at Tempus. Kate’s journey is a testament to the power of combining passion with purpose. Inspired by the groundbreaking approval of Herceptin in 1998, she set out to bridge biology and medicine to address real-world challenges. From early roles in clinical trial coordination to leadership positions at biotech pioneers like Genmab, Kate has carved out a career at the forefront of innovation. Now, she leads research and development at Tempus.
In our conversation, Kate shares the principles driving Tempus and the programs she’s most excited about, from leveraging AI to extract insights from unstructured clinical data to creating patient-derived organoid models for refining therapeutic targets. One of the most compelling aspects of our discussion is Tempus’ vision for an interconnected “lab-in-a-loop” system—a continuous cycle of real-world data collection, biological modeling, and AI-driven analysis.
Highlights of this interview have been edited for length and clarity.
IPM: What patient data is at the core of developing AI-enabled solutions for personalized healthcare?
Sasser: We want to get the most data from any patient at any time, not just the clinical data and a single DNA panel. Firstly, this means we’re offering an expansive range of NGS tests across the patient’s journey, not just DNA panels, but also liquid biopsy and whole transcriptome. So we’ve been running the entire RNA transcriptome for several years. We now have over 300,000 patients profiled with DNA, RNA, and often a matched normal sample, which gives us a very robust molecular profile of each patient that can be combined with clinical data. Often this includes CT scans and pathology data, such as hematoxylin and eosin (H&E) stain or immunohistochemistry (IHC) stains. We also have imaging data. We also incorporate other modalities, broader proteomics profiling, spatial transcriptomics, and single-cell sequencing. You can envision a future where we continue adding various modalities to this data. The data will enable us to leverage AI and machine learning for pattern recognition, uncovering new biological insights. These insights can be translated into practical applications, such as developing new diagnostic tests or drug modalities in collaboration with our pharma or biotech partners. There is a wealth of information to analyze and consider.
IPM: How does Tempus define RWD? What potential does Tempus see in the role of RWD in precision medicine? How is Tempus leveraging RWD?
Sasser: When we talk about RWD, we’re talking about the data that comes directly from patients, physicians, and healthcare systems. We discussed being integrated with electronic medical records and pulling RWD to structure, clean, organize, and ask questions about.
Clinical notes are a rich source of really interesting connections because you think about how they enter the system. Often, a doctor sees a patient, takes notes, and describes the visit, including observations, the patient’s journey, and recommendations. And they can use this spot text to capture whatever the physician or care provider is thinking about in that moment. There are many instances where you can gain fascinating insights, and where you must be cautious with data cleaning. Occasionally, a patient’s history is just a note, and you may have to dig to find that they have a comorbidity. The doctor may say the patient has lung cancer and that their brother and mother have diabetes. If you don’t clean up the data, you may have inaccurate or misleading information in the database. For example, in this scenario, the brother has diabetes, and the mother has heart disease. Sometimes, systems may mistakenly include the patient’s information instead of the family history if not carefully managed. So, those are just nuances that come into the work that have to happen to clean up something like an unstructured note. Now that AI is entering this space, we can do that work much faster.
You can envision AI agents that can now crawl across these notes when they’re not structured yet, pull out associations, and start to catalog some of these things together. We can find interesting patterns and insights there, especially at scale, which we don’t often discuss in detail. When we’re thinking about it, we’re discussing multimodal data and other approaches. All of those things get so interesting when you start operating at a truly massive scale. When you go from a hundred patients to a thousand to a hundred thousand, or hopefully to a million, the types of insights and patterns you can uncover in notes or multimodal data become increasingly interesting. At such scale, you can identify patterns even for very rare alterations or patient groups.
IPM: How does Tempus test and then incorporate the identified emergent patterns into the patient experience to improve healthcare?
Sasser: It’s a big challenge because it doesn’t have an official start—where’s the head and where’s the tail? We conceptualize it as a continuous loop or circle.
We have an interesting loop between RWD and multimodal data that incorporates genomics, clinical data, and imaging, along with these outcomes in the real world. We can leverage that data to ask interesting questions, whether you’re a drug developer, a physician, or a researcher.
We’re super excited about a systems biology approach that we’ve leveraged utilizing patient-derived organoids. We’ve created biological models in the lab that we think better represent the patients. We can go from RWD to the biological models we have and refine the question. This approach allows us to dive deeper into these model systems, examine the functional outputs from those models, and then take the functional outputs observed in a more controlled setting and return to RWD.
Instead of having a distinct beginning and end, we establish a cycle in which we pose intriguing questions about the data model, refine these questions—sometimes in collaboration with partners or researchers—using our own model systems, and then return to the data system itself. With this lab-in-the-loop approach, we aim to accelerate the discovery and insight generation process, enabling us to move more swiftly. Whether that’s drug development or building a new diagnostic tool, we can move along much faster and uncover derived insights, generating products in a much shorter period that are also more relevant to the patients we’re studying and developing products for.
We’re coming off of a project where we’ve shown that by leveraging this lab-in-the-loop, starting with Tempus RWD, focusing on one particular indication, non-small cell lung cancer, and being able to uncover subtypes of patients that have very poor prognoses, regardless of standard of care therapy, and then being able to refine what we know about those patients in terms of their lack of response, be able to characterize them more deeply with some of the multimodal data, the molecular data that we have, and then be able to take that insight and go into our biological models. In this case, we’re using patient-organoids, and we were able to show that.
We have a good representation of patients who are poor responders to standard care therapy. We can then go into the organoids and test a wide range of potential targets based on what we know about the biology of those patients. We can start to screen those targets using functional screens and CRISPR screens, then narrow down to find potential drugs against those targets, and finally go back into the RWD to ask: if we were able to target that particular gene of interest, what would happen to those patients? Do we see evidence in the RWD that targeting this gene changes patient outcomes?
We were able to run that example that I just talked through in a matter of months: find a cohort of patients that are not doing standard of care therapy, go after particular targets, identify drugs against those targets, ask the question, do those drugs actually change output in the functional screens, and then go back to the data. We were able to do that in six to nine months, a timeline that is significantly shorter than the typical drug development process. This accelerated timeline is exciting. As we collect more data, we’ll be able to uncover even more novel subtypes of interest and develop unique targets against that biology.
Another area we are excited about is enhancing our physician-focused efforts by developing new diagnostics and focusing on highly sensitive methods for measuring minimal residual disease. We’re spending a lot of time and energy on building out both a tumor-informed platform and a tumor-naive platform so we can leverage both options for patients and physicians, primarily in oncology, where it’s most relevant today.
We’re talking about DNA, RNA, imaging, sometimes DigPath, H&E, and CT scans. As that continues to grow, we’re also envisioning being able to go from single biomarkers to diagnostics. Typically, oncology uses biomarkers such as PD-L1 or specific mutations like EGFR. To transition from single to multimodal measurements, we’re developing algorithmic diagnostics that combine multiple biomarkers or signatures to refine patient recommendations.
This is just one example of an approach we think will be used more often, not just at Tempus, but in the field as a whole, where we can start to combine measurements and biomarkers to refine predictions for patients ultimately. As the robustness of the data expands, we’ll see these multimodal, multiparametric biomarkers come into play—a super exciting development.
