Q&A: From proof of concept to clinical deployment: How foundation models are accelerating drug development

05/01/2026

Q&A: From proof of concept to clinical deployment: How foundation models are accelerating drug development

Tempus leaders discuss the practical application of multimodal foundation models, from de-risking programs and uncovering novel biomarkers to deploying AI-driven insights at the point of care.

Authors Razik Yousfi
Senior Vice President, AI Products, Tempus

Arpita Saha
Vice President, Applied AI and Research, Tempus

Siqi Liu
Vice President, Machine Learning, Tempus

The path to bringing novel therapies to market is challenged by high attrition rates and the increasing complexity of patient biology. Foundation models offer a new approach to addressing challenges in drug development, particularly when trained on diverse multimodal, de-identified real-world data.

In a recent webinar, Tempus leaders Razik Yousfi, SVP and GM of AI Products, Arpita Saha, VP of Applied AI and Research, and Siqi Liu, VP of Machine Learning, discussed the practical application of foundation models in drug development. They explored how these technologies can help address key research questions and support program development through real-world deployment.

This article synthesizes their discussion, highlighting how Tempus creates a unique pathway from research to reality, enabling life sciences partners to build and deploy novel AI models.

How does Tempus define and approach the development of foundation models for oncology?

Siqi Liu: Our approach to building foundation models is driven by the nature of the data itself. A patient’s cancer journey is not a single snapshot in time; it’s a temporal and multimodal story. Our goal is to build a unified representation of that journey by integrating data from critical anchor points, like the start of therapy or disease progression.

We focus on two core signals: the tumor tissue biology, derived from pathology and genomics, and the patient journey, captured in longitudinal EHR data and clinical notes. By combining these, we create a much richer understanding of a patient’s state at any given time. This allows us to move beyond traditional AI, which often requires building a new model from scratch for every question. A foundation model, trained on this comprehensive data, can serve as a “pre-trained biological engine” to help answer questions like: Given everything we know, will a patient respond better to a specific therapy versus other options?

With so many AI tools available, what makes foundation models an essential shift for drug development?

Arpita Saha: The simplest way to think about it is that a traditional AI model is a single-purpose tool, whereas a foundation model is a biological engine. You don’t necessarily need a foundation model to solve one static problem, but you do to run an entire R&D pipeline efficiently.

The key economic shift is that you can stop relearning biology from scratch for every new trial or research question. Instead, you start with a pre-trained understanding. This is especially powerful in two common scenarios: developing therapies for rare indications where labeled data is scarce, and answering complex questions that span multiple data types like genomics, imaging, and clinical history. The result is higher model performance, faster development with less training data, and lower overall costs compared to building task-specific AI for every need.

“The key shift is we should not be relearning the biology every time from scratch, for every trial, for every question, but start from a trained understanding of biology.”

– Arpita Saha, VP, Applied AI and Research, Tempus

Publicly available models are common. What is the unique advantage of a model trained on Tempus’ proprietary, multimodal, de-identified data?

Siqi Liu: Public models are a great baseline for fast iteration, and we use them for benchmarking. However, they are often limited to a single modality and lack the scale and diversity of data needed to capture the full complexity of cancer biology.

The advantage of the Tempus dataset is its combination of scale, multimodality, and longitudinality. We have a large and growing library of paired data, including DNA, RNA, and pathology images, linked to treatment outcomes and the patient journey. But data is only part of the story. We match that data scale with the necessary compute power and team expertise to unlock its full potential. Simply feeding this data into a generic large language model won’t work. Our approach is designed to capture predictive signals that go beyond simple prognosis, which requires the rich clinical context that only comes from multimodal, longitudinal, de-identified data.

What are some tangible use cases for life sciences partners using these models to de-risk a program or accelerate a timeline?

Arpita Saha: We are already seeing foundation models create significant value. A few key use cases include:

Biomarker detection from H&E: We can predict IHC and molecular biomarker status directly from a standard H&E-stained pathology slide, with no additional assay needed. This unlocks molecular signals from tissue you already have, which can expand cohort sizes for retrospective analysis and help replicate biomarker strategies across indications.
H&E-based response prediction: The model can identify morphological features in a tumor that are predictive of treatment response but are not visible to the human eye or captured by a single marker. This can help identify responders earlier and discover novel resistance signatures.

A model is powerful, but how does Tempus make these AI-driven discoveries actionable in a clinical setting?

Arpita Saha: The real value comes from our integrated, end-to-end ecosystem that takes a discovery from research to deployment. This includes three layers:

Build: You can build custom algorithmic tests for your asset or indication on top of our foundation models.
Validate: You can validate your biomarker in Tempus’ CLIA-certified lab to generate the necessary evidence for regulatory review.
Deploy: We can deliver the final, regulated test through our extensive hospital and lab network at the point of care.

A great example is Paige Predict, launched in January 2026.¹ Paige Predict is an AI-enabled digital biomarker panel that can predict over 1,600 biomarkers across 505 genes from H&E slides with a single AI-enabled test. It is currently available as an off-the-shelf research tool to support biomarker discovery and research in tissue-scarce cases.

What is your advice for a life sciences leader who wants to get started with this technology?

Arpita Saha: Start with the question, not the technology. Identify the single moment in your program where uncertainty is costing you the most, whether it’s patient selection, biomarker strategy, or a go/no-go decision. Then, ask what data you already have and what data you are missing. That clarity is what turns a foundation model from an interesting capability into a program-level asset.

Siqi Liu: Precision medicine AI is advancing, and pilot projects can be used to research and validate signals as part of a broader scalable biomarker strategy. The technology will only get better as we accumulate more data and continue to push the algorithmic boundaries.

Footnotes

For Research Use Only (RUO). Not for use in diagnostic procedures.

Note: Content edited for clarity and conciseness. Some language and formatting may have been adjusted to enhance readability while preserving the original meaning and intent of the discussion.

To gain comprehensive insights into Tempus’ role in advancing precision medicine, we invite you to watch the webinar recording. For in-depth demonstrations of our AI-enabled applications, contact us here.