Foundation models in pathology: How Virchow is accelerating precision drug development

Executive summary

 

• Unlocking Underutilized Data: H&E images contain rich biological information that remains massively underutilized today; our pathology foundation models drive a paradigm shift by unlocking H&E as a powerful, routine modality to directly inform drug development.
• Trained at Scale: Virchow, part of the pFM suite of Tempus’ pathology foundation models, is trained on millions of real-world, diverse tissue images, then adapts to many downstream tasks with far less labeled data, including in the rare and biomarker-defined cohorts where data is scarcest.
• Accelerated Impact: For life sciences teams, the payoff is speed and reach: faster biomarker discovery, patient stratification, and companion diagnostic development, backed by the validation and regulatory rigor required to move from research to the clinic.

 

Introduction

 

Artificial intelligence is inseparable from the future of drug development, and nowhere is its potential more concentrated than in pathology. The hematoxylin and eosin (H&E) slide is one of the richest, most universal records of disease biology in medicine, yet historically the information inside those slides has been read one question at a time, by one model trained for one task. Furthermore, the pathology modality was only a domain for basic diagnosis, and underutilized for deeper biological understanding.

 

That model-by-model approach is now the rate limiter. As modern advances in precision medicine are enabling sponsors to target new biomarker-defined indications and expand investigational programs into increasingly rare populations, building a bespoke model for every question has become untenable. Foundation models offer a different path, and Tempus, with our Virchow pathology foundation model (part of our pFM suite), is helping define what that path looks like for life sciences teams designing programs that depend on extracting more signal, faster, from the tissue they already have.

 

Context and challenges

 

The core challenge in computational pathology is not a shortage of images; it is the cost of making images usable for drug development. Traditional, task-specific models depend on large volumes of expertly annotated data, and that dependency creates three persistent gaps.

 

• Annotation does not scale. Expert labeling is slow and expensive, and a model built for one task rarely transfers to the next, so every program effectively starts from zero.
• Rare conditions are systematically underserved. A model trained on the handful of slides available for a rare cancer or biomarker lacks the signal to perform well, yet these are often exactly the subpopulations where unmet need and development risk concentrate.
• Real-world variability breaks brittle models, and limits their clinical applicability. Staining, scanners, and tissue-preparation methods differ across institutions, so a model tuned to one site’s data often degrades on another’s, a serious problem for multi-site studies and for evidence that must withstand regulatory scrutiny.

 

The net effect: turning tissue into evidence at the pace modern drug development demands has been the slowest, most expensive part of translational pathology. As a result, this modality is dramatically underused in drug development, and valuable information contained in the H&E image remains largely untapped today.

 

Innovative solutions and emerging trends

 

Foundation models close these gaps. Trained once on a very large, diverse image dataset through self-supervised learning, a model learns useful representations of tissue directly from the images themselves, without requiring a human annotation on every slide.

 

Because Virchow carries general features of tissue learned across its full corpus, those patterns transfer to new applications, even when task-specific examples are few. In practice, that generalization supports uses that matter most to drug development:

 

• Accelerating Diagnostic tasks, such as cancer detection, classification, and subtyping across both common and rare tissue types.
• Phenotype-genotype mapping, enabling biomarker detection and characterization of expression and regulation across genomic, transcriptomic, and protein levels. Using only a single system, it identifies key molecular markers—such as MSI-H, BRAF, and FGFR—directly from standard H&E images across cancer types.
• Novel biomarker discovery, including biomarkers that predict payload resistance or overall therapeutic response.
• Tumor microenvironment characterization, associating Virchow’s tissue representations with spatial transcriptomic data to map how cellular architecture and gene expression jointly shape biomarker expression and therapeutic response.

 

“Instead of assembling and annotating a large dataset for every question, you start from a model that has already learned what tissue looks like across many cancers and has a fundamental understanding of tissue morphology. By unlocking these structural patterns, we can directly link morphology to patient outcomes and therapeutic benefits. That strategic advantage allows you to scale high-performance pathology models across your entire portfolio of programs.”

– Siqi Liu, Vice President, Machine Learning, Tempus

 

Leverage comes from breadth: Virchow2, the foundational building block of our pathology AI portfolio, was trained on roughly 3.1 million H&E and 400,000 IHC whole slide images spanning more than 40 tissue types, drawn from a wide breadth of labs globally (see the full model specifications for the Virchow family). The model produces tile-level embeddings, compact digital fingerprints of each slide region, and teams build lightweight models on top of them for detection, grading, subtyping, quantification, and segmentation—work that is faster and far less data-hungry because the foundation has already been laid.

 

Practical implications for Life Sciences organizations

 

For sponsors, the implications are concrete. A pretrained pathology backbone enables a team to rapidly build multiple downstream models with less labeled data and shorter timelines per program, turning translational pathology from one-model-at-a-time into build-once, reuse-everywhere. That makes previously marginal questions, particularly in rare and biomarker-defined cohorts, worth pursuing.

 

It also strengthens the kind of evidence that holds up. Because Virchow is trained across a diversity of labs and countries, models built on it are better positioned for the multi-site reproducibility that is a necessity for robust validation, regulatory engagement, and clinical deployment. This foundation is further strengthened by our rigorous data governance framework, which supports end-to-end data provenance and security. Crucially, a foundation model is a starting line, not a finish line. Any application intended to inform clinical decisions still requires its own task-specific, stratified validation and appropriate regulatory authorization.

 

This is where partnership matters: Tempus brings proven expertise in taking models from proof-of-concept to robust, validated, and documented products such as Paige Predict and PanCancer Detect, alongside a dedicated regulatory team experienced across software and traditional diagnostics.

 

Future outlook

 

Building on the pioneering work of Tempus and Paige on Virchow and PRISM, and following its acquisition of Paige, Tempus is now expanding the portfolio with multimodal foundation models like oFM that integrate pathology images, clinical notes, and other modalities, unlocking a more comprehensive understanding of disease. That work draws on one of the largest de-identified multimodal datasets in the world, with millions of unique cases across therapeutic areas.

 

Virchow enables life sciences teams to stop rebuilding from scratch for every study and instead build on a foundation that compounds in value across their portfolio. The model unlocks H&E as a powerful, versatile modality across every stage of the drug development lifecycle.

 

To explore how Tempus’ foundation models could accelerate your pipeline, contact your Tempus life sciences team or learn more about our pathology foundation models.

 

This article describes capabilities of the Virchow foundation model and Tempus’ foundation-model program, including statements about potential applications in drug development and precision medicine. Such forward-looking statements reflect current expectations and involve risks and uncertainties; actual results and the performance of any model or product built using these technologies may differ materially. Capabilities described reflect model performance in research settings and do not represent the validated performance of any specific commercial product or clinical diagnostic. Nothing in this article constitutes a claim of clinical or diagnostic performance for any Tempus product.

 

References

 

1. Vorontsov E, Bozkurt A, Casson A, et al. Virchow: A Million-Slide Digital Pathology Foundation Model. arXiv:2309.07778. https://arxiv.org/abs/2309.07778

 

2. Paige. The Virchow Foundation Model, Explained: A Q&A with an AI Scientist. https://www.paige.ai/blog/the-virchow-foundation-model-explained-a-qa-with-an-ai-scientist

 

3. Paige Unveils Game-Changing AI That Revolutionizes Cancer Detection Across Multiple Tissue Types