How AI-powered computational pathology is revolutionizing precision medicine

• The TROP2 NMR device performs cell-by-cell calculations to provide diagnostic insights that are impossible to achieve through traditional methods¹

• AI-driven algorithms can identify patient populations for therapies that reader-assist algorithms and traditional interpretation may miss, validating life-saving therapies that would otherwise fail to meet clinical endpoints

• This FDA-recognized device provides a validated, reproducible blueprint for AI-powered companion diagnostics across various tumor types and therapeutic areas

As drugs become more complex, we need better tools to figure out which drug is going to work for which patient, and to make sure those drugs are actually being used for the patients they'll help. That simple idea has led to the creation of complex technology that’s changing the landscape of companion diagnostics. 

That technology, the VENTANA® TROP2 (EPR20043) RxDx Device (TROP2 NMR), is the first application of AI to identify a patient’s eligibility for a specific drug therapy. A groundbreaking companion diagnostic for non-small cell lung cancer, it received US FDA Breakthrough Device Designation in 2025 and a version of the device is currently available for research use only,* with full clinical launch planned for 2027.¹

What makes TROP2 NMR remarkable is its precision, which has the power to benefit patients as well as those who treat them. If a therapy fails a clinical trial due to standard human assessments, it can’t help anyone. We’ve seen that happen. But computational pathology is changing that. For pharmaceutical companies, this technology may meaningfully improve clinical trial outcomes.

Here’s an example of what TROP2 NMR can do: Datopotamab deruxtecan (Dato-DXd) is a targeted TROP2-directed antibody-drug conjugate designed to treat advanced solid tumors, including breast and lung cancers. In one trial, Dato-DXd failed to demonstrate efficacy in both an unselected patient population and a biomarker-selected patient population using traditional manually scored IHC staining.² When computational pathology was applied, it changed the outcome, this time enabling identification of patients who responded to Dato-DXd.³ The precision of TROP2 NMR has been shown in additional studies⁴ and is currently undergoing evaluation in clinical trials.⁵ If these clinical studies are positive, TROP2 NMR will enable patients to access a treatment that they may not have ever had access to.

Computational pathology isn’t just for IHC. It holds promise for improving precision in other areas: from blood-based diagnostics to DNA-based, and from oncology to neurology and beyond. Using it in clinical studies at the development stage could be transformative for our partners.

TROP2 NMR is much more than a single test. It’s an integrated system whose components work in sequence (see Figure 1). Those components include: 

• An antibody to make protein in patient tissue visible 

• An automated staining instrument

• A slide scanner to create high-resolution digital images 

• An AI-powered algorithm

After a pathologist reviews the data, TROP2 NMR analyzes the data and creates a normalized membrane ratio (NMR) score, which determines a patient’s likely response to therapy. What makes the device truly novel is the way all these components work together to deliver a level of precision that simply wasn't possible before.

When most people think about AI, they picture ChatGPT: ask a question, get an answer. Computational pathology is different. Instead of scraping the internet, it measures things: pulling data from an image of tissue, identifying features the human eye can’t see and performing calculations no pathologist could do alone. In the case of TROP2 NMR, the output is a diagnostic insight that’s novel because of its precision. It answers the question “Does this patient really qualify for treatment?” more accurately than ever before. This improves pathologists’ ability to find the right patients and perhaps even more patients than in the past.

TROP2 NMR performs observations and calculations beyond human capability, but it doesn’t replace humans. It frees pathologists to do what they do best: interpreting tissue morphology, applying clinical judgment and validating the algorithm’s work. Meanwhile, the algorithm does what it does best: counting, calculating and quantifying at a scale and speed no human can match.

Pathologists have a role to play in annotating the algorithm’s work and ensuring the AI is performing appropriately. It’s a fully supervised approach that requires training and trust, as the algorithm analyzes and counts hundreds of thousands of cells and scores the protein expression on each one simultaneously to compute the normalized membrane ratio (NMR) score.

To truly understand why computational pathology and TROP2 NMR are a breakthrough, it helps to know where we started. 

The tests we create at Roche Diagnostics use complex chemistry to identify proteins in tumor tissue. For decades, evaluation of the stained tissue has been a manual process. A trained pathologist observes where a protein sits within a cell, approximates how much protein is present based on color and intensity, and uses that observation to determine if a patient qualifies for a specific drug. Take PD-L1 scoring in lung cancer for example: Pathologists have to discriminate between tumor cells and immune cells that can look nearly identical under a microscope. It's not easy, and at the end of the day, it's still an approximation.

When Roche first began building digital pathology algorithms, the goal was to help pathologists use diagnostic tools to make those approximations faster and more consistently. Early AI algorithms, known as “reader-assist” tools, improved diagnostics, but the tools were trained with human-scored data and inherited inaccuracies. Sometimes pathologists have to override the results.

TROP2 NMR is different. It doesn’t replicate what pathologists do; it does things they can’t, delivering a diagnostic output that no one could arrive at just by looking through a microscope. It counts cells, scores protein expression on a cell-by-cell basis and runs calculations across hundreds of thousands of data points. You could say the leap from reader-assist tools to TROP2 NMR is like the leap from Pong to Fortnite.

When I began working in this field nearly 30 years ago as a translational academic researcher I focused on lung cancer, and I’ve been working on lung cancer ever since. I’ve seen treatments evolve from chemotherapy and radiation alone to a growing arsenal of targeted therapies and immunotherapies that give patients more options — and more hope.

Last year, my work with lung cancer became personal. My mother, Mimi, was diagnosed with lung cancer. She’s doing well. She’s lucky her cancer was diagnosed early and she carries a mutation in her tumor that qualifies her for targeted therapy. 

Patients like my mother represent a relatively small portion of the overall lung cancer population. The majority of patients, those without an actionable genomic alteration, have had limited access to targeted therapies. Computational pathology could change that by using a level of detail that can more precisely identify which patients are likely to respond to treatment. This gives me hope that there will be more options for patients, and more confidence that the treatment they’re receiving is the right one.

Recent studies show that pathologists believe the strongest appeal of computational pathology is its potential to match or exceed the performance of traditional manual pathology, but to make the shift, they need support and a better understanding of the technology.⁶

To address this need, our medical affairs team is laying groundwork: educating the pathologist community about computational pathology and making sure everyone understands their role in it. We’re not doing this in a controlled corporate lab, but in active clinical labs across the globe. We’ve given pathologists real samples to run and looked at whether TROP2 NMR can replicate consistent results from lab to lab.

What we’ve found has exceeded our expectations. Across 12 labs in the United States and Europe, including 11 external sites and involving 36 pathologists, TROP2 NMR demonstrated strong performance, with overall agreement between sites found to be 94%.⁷ Even more surprising: When we had our study pathologists separately review the exact same set of NSCLC samples, the overall agreement between the 36 pathologists was 100%.⁸

When we first presented this data at the International Association for the Study of Lung Cancer (IASLC) World Conference on Lung Cancer in September 2025, it created excitement, and word traveled. Someone actually made the comment that the data was too good to be true. The final data, which we presented in March 2026 at the European Lung Cancer Congress (ELCC) in Copenhagen,³ showed that this novel device is highly reproducible in real-world laboratories.

AI is redefining what IHC companion diagnostics can do, making them even more critical to the mission of connecting more patients with new precision therapies. With our expertise and global reach, we’re helping our pharmaceutical partners deliver in this new landscape. We are developing innovative solutions, vetting them, bringing them to market, and ultimately driving adoption.

If the clinical data continues to bear out, our work with algorithms and devices like TROP2 NMR won’t stop with lung cancer. TROP2 NMR is a template for devices that could address other tumor types and identify patient eligibility for other drugs with the same level of precision. 

That’s what I find most exciting about this moment. We’re not just building and launching a test. We’re driving education and training so pathologists and oncologists are ready to use computational pathology to make a real difference for patients. 

In a broader sense, we’re establishing what it looks like to bring AI-powered companion diagnostics to market safely and responsibly by laying the groundwork for everything that comes next.