AI detects cancer but it’s also reading who you are

 

At a glance:

• A new study shows that artificial intelligence systems used to diagnose cancer from pathology slides do not perform equally for all patients, with accuracy varying across different demographic groups.
• Researchers pinpointed three key reasons behind this bias and created a new approach that significantly reduced these differences.
• The results emphasize why medical AI must be routinely evaluated for bias to help ensure fair and reliable cancer care for everyone.

Pathology and the Foundations of Cancer Diagnosis

For decades, pathology has been essential to how doctors diagnose and treat cancer. A pathologist studies an extremely thin slice of human tissue under a microscope, searching for visual signs that reveal whether cancer is present and, if so, what type and stage it has reached.

To a trained specialist, examining a pink, swirling tissue sample dotted with purple cells is like grading a test without a name on it -- the slide contains vital information about the disease, but it offers no clues about who the patient is.

When AI Sees More Than Expected

That assumption does not fully apply to artificial intelligence systems now entering pathology labs. A new study led by researchers at Harvard Medical School shows that pathology AI models can infer demographic details directly from tissue slides. This unexpected ability can introduce bias into cancer diagnosis across different patient groups.

After evaluating several widely used AI models designed to identify cancer, the researchers found that these systems did not perform equally for all patients. Diagnostic accuracy varied based on patients' self-reported race, gender, and age. The team also uncovered several reasons why these disparities occur.

To address the issue, the researchers developed a framework called FAIR-Path, which significantly reduced bias in the tested models.

"Reading demographics from a pathology slide is thought of as a 'mission impossible' for a human pathologist, so the bias in pathology AI was a surprise to us," said senior author Kun-Hsing Yu, associate professor of biomedical informatics in the Blavatnik Institute at HMS and HMS assistant professor of pathology at Brigham and Women's Hospital.

Yu emphasized that recognizing and correcting bias in medical AI is critical, since it can directly influence diagnostic accuracy and patient outcomes. The success of FAIR-Path suggests that improving fairness in cancer pathology AI, and possibly other medical AI tools, may not require major changes to existing systems.

The work, which was supported in part by federal funding, is described Dec. 16 in Cell Reports Medicine.

Putting Cancer AI to the Test

Yu and his colleagues examined bias in four commonly used pathology AI models currently being developed for cancer diagnosis. These deep-learning systems were trained on large collections of labeled pathology slides, allowing them to learn biological patterns and apply that knowledge to new samples.

The team evaluated the models using a large, multi-institutional dataset that included pathology slides from 20 different types of cancer.

Across all four models, performance gaps consistently emerged. The AI systems were less accurate for certain demographic groups defined by race, gender, and age. For example, the models struggled to distinguish lung cancer subtypes in African American patients and in male patients. They also showed reduced accuracy when classifying breast cancer subtypes in younger patients. In addition, the models had difficulty detecting breast, renal, thyroid, and stomach cancers in some demographic groups. Overall, these disparities appeared in roughly 29 percent of the diagnostic tasks analyzed.

According to Yu, these errors arise because the AI systems extract demographic information from the tissue images -- and then rely on patterns linked to those demographics when making diagnostic decisions.

The findings were unexpected. "Because we would expect pathology evaluation to be objective," Yu said. "When evaluating images, we don't necessarily need to know a patient's demographics to make a diagnosis."

This led the researchers to ask a key question: Why was pathology AI failing to meet the same standard of objectivity?

Why Bias Appears in Pathology AI

The team identified three main contributors to the bias.

First, training data are often uneven. Tissue samples are easier to obtain from some demographic groups than others, resulting in imbalanced datasets. This makes it harder for AI models to accurately diagnose cancers in groups that are underrepresented, including some populations defined by race, age, or gender.

However, Yu noted that "the problem turned out to be much deeper than that." In several cases, the models performed worse for certain demographic groups even when sample sizes were similar.

Further analysis pointed to differences in disease incidence. Some cancers occur more frequently in specific populations, allowing AI models to become especially accurate for those groups. As a result, the same models may struggle to diagnose cancers in populations where those diseases are less common.

The researchers also found that AI models can detect subtle molecular differences across demographic groups. For example, the systems may identify mutations in cancer driver genes and use them as shortcuts to classify cancer type -- which can reduce accuracy in populations where those mutations are less prevalent.

"We found that because AI is so powerful, it can differentiate many obscure biological signals that cannot be detected by standard human evaluation," Yu said.

Over time, this can cause AI models to focus on signals tied more closely to demographics than to the disease itself, weakening diagnostic performance across diverse patient groups.

Taken together, Yu said, these findings show that bias in pathology AI is influenced not only by the quality and balance of training data, but also by the way the models are trained to interpret what they see.

A New Approach to Reducing Bias

After identifying the sources of bias, the researchers set out to correct them.

They developed FAIR-Path, a framework based on an existing machine-learning method known as contrastive learning. This approach modifies AI training so that models focus more strongly on critical distinctions, such as differences between cancer types, while reducing attention to less relevant differences, including demographic characteristics.

When FAIR-Path was applied to the tested models, diagnostic disparities dropped by about 88 percent.

"We show that by making this small adjustment, the models can learn robust features that make them more generalizable and fairer across different populations," Yu said.

The result is encouraging, he added, because it suggests that meaningful reductions in bias are possible even without perfectly balanced or fully representative training datasets.

Looking ahead, Yu and his team are working with institutions worldwide to study pathology AI bias in regions with different demographics, clinical practices, and laboratory settings. They are also exploring how FAIR-Path could be adapted for situations with limited data. Another area of interest is understanding how AI-driven bias contributes to broader disparities in health care and patient outcomes.

Ultimately, Yu said, the goal is to develop pathology AI systems that support human experts by delivering fast, accurate, and fair diagnoses for all patients.

"I think there's hope that if we are more aware of and careful about how we design AI systems, we can build models that perform well in every population," he said.

Authorship, funding, disclosures

Additional authors on the study include Shih-Yen Lin, Pei-Chen Tsai, Fang-Yi Su, Chun-Yen Chen, Fuchen Li, Junhan Zhao, Yuk Yeung Ho, Tsung-Lu Michael Lee, Elizabeth Healey, Po-Jen Lin, Ting-Wan Kao, Dmytro Vremenko, Thomas Roetzer-Pejrimovsky, Lynette Sholl, Deborah Dillon, Nancy U. Lin, David Meredith, Keith L. Ligon, Ying-Chun Lo, Nipon Chaisuriya, David J. Cook, Adelheid Woehrer, Jeffrey Meyerhardt, Shuji Ogino, MacLean P. Nasrallah, Jeffrey A. Golden, Sabina Signoretti, and Jung-Hsien Chiang.

Funding was provided by the National Institute of General Medical Sciences and the National Heart, Lung, and Blood Institute at the National Institutes of Health (grants R35GM142879, R01HL174679), the Department of Defense (Peer Reviewed Cancer Research Program Career Development Award HT9425-231-0523), the American Cancer Society (Research Scholar Grant RSG-24-1253761-01-ESED), a Google Research Scholar Award, a Harvard Medical School Dean's Innovation Award, the National Science and Technology Council of Taiwan (grants NSTC 113-2917-I-006-009, 112-2634-F-006-003, 113-2321-B-006-023, 114-2917-I-006-016), and a doctoral student scholarship from the Xin Miao Education Foundation.

Ligon was a consultant of Travera, Bristol Myers Squibb, Servier, IntegraGen, L.E.K. Consulting, and Blaze Bioscience; received equity from Travera; and has research funding from Bristol Myers Squibb and Lilly. Vremenko is a cofounder and shareholder of Vectorly.

The authors prepared the initial manuscript and used ChatGPT to edit selected sections to improve readability. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

Journal Reference:

1. Shih-Yen Lin, Pei-Chen Tsai, Fang-Yi Su, Chun-Yen Chen, Fuchen Li, Junhan Zhao, Yuk Yeung Ho, Tsung-Lu Michael Lee, Elizabeth Healey, Po-Jen Lin, Ting-Wan Kao, Dmytro Vremenko, Thomas Roetzer-Pejrimovsky, Lynette Sholl, Deborah Dillon, Nancy U. Lin, David Meredith, Keith L. Ligon, Ying-Chun Lo, Nipon Chaisuriya, David J. Cook, Adelheid Woehrer, Jeffrey Meyerhardt, Shuji Ogino, MacLean P. Nasrallah, Jeffrey A. Golden, Sabina Signoretti, Jung-Hsien Chiang, Kun-Hsing Yu. Contrastive learning enhances fairness in pathology artificial intelligence systems. Cell Reports Medicine, 2025; 6 (12): 102527 DOI: 10.1016/j.xcrm.2025.102527 

Courtesy:

Harvard Medical School. "AI detects cancer but it’s also reading who you are." ScienceDaily. ScienceDaily, 17 December 2025. <www.sciencedaily.com/releases/2025/12/251217231230.htm>. 

 

 

 

 

 

 

A hidden T cell switch could make cancer immunotherapy work for more people

Over the past ten years, T cell immunotherapy has emerged as one of the most promising developments in cancer treatment. These therapies work by training a patient's own immune system to detect and destroy dangerous cells. Despite their success, scientists have struggled to fully explain how these treatments function at a molecular level. This lack of understanding has slowed progress, especially since T cell therapies work well for only a small number of cancer types and fail in most others, for reasons that have remained unclear. Gaining insight into their modus operandi could help make these therapies effective for far more patients.

Scientists at The Rockefeller University have now uncovered crucial details about the T cell receptor (TCR), a protein complex embedded in the cell membrane that plays a central role in T cell therapies. Using cryo-EM, researchers from the Laboratory of Molecular Electron Microscopy studied the receptor in a biochemical setting designed to closely resemble its native milieu. They discovered that the TCR behaves like a jack-in-the-box, staying compact until it encounters an antigen or another suspicious particle, at which point it rapidly opens. This behavior contradicts what earlier cryo-EM studies of the receptor had shown.

The findings, published in Nature Communications, could help researchers improve and expand the use of T cell immunotherapies.

"This new fundamental understanding of how the signaling system works may help re-engineer that next generation of treatments," says first author Ryan Notti, an instructor in clinical investigation in Walz's lab and a special fellow in the Department of Medicine at Memorial Sloan Kettering Cancer Center, where he treats patients with sarcomas, or cancers that arise in soft tissue or bone.

"The T cell receptor is really the basis of virtually all oncological immunotherapies, so it's remarkable that we use the system but really have had no idea how it actually works -- and that's where basic science steps in," says Walz, a world expert in cryo-EM imaging. "This is some of the most important work to ever come out of my lab."

How T Cells Detect Threats

Walz's lab focuses on producing detailed images of macromolecular complexes, especially proteins found in cell membranes that help cells communicate with their surroundings. The TCR is one such complex. Made up of multiple proteins, it enables T cells to recognize antigens displayed by human leukocyte antigen (HLA) complexes on other cells. This recognition process is what T cell therapies rely on to mobilize the immune system against cancer.

Although scientists have known the individual parts of the TCR for many years, the earliest steps that trigger its activation have remained elusive. Notti, who works as both a physician and a researcher, found this gap especially troubling because many of his sarcoma patients were not benefiting from T cell immunotherapies.

"Determining that would help us understand how the information gets from outside the cell, where those antigens are being presented by HLAs, to the inside of the cell, where signaling turns on the T cell," he says.

Notti earned his Ph.D. in structural microbiology at Rockefeller before moving into oncology, and he suggested to Walz that they investigate this unanswered question together.

Rebuilding the TCR's Natural Environment

Walz's team is known for creating custom membrane environments that closely mimic the natural surroundings of membrane proteins. "We can change the biochemical composition, the thickness of the membrane, the tension and curvature, the size -- all kinds of parameters that we know have an influence on the embedded protein," Walz says.

For this study, the researchers set out to observe the TCR in conditions that closely resemble those inside a living cell. They placed the receptor into a nanodisc, a tiny disc-shaped section of membrane held in solution by a scaffold protein wrapped around its edge. Assembling the full receptor was difficult, and "getting all eight of these proteins properly assembled into the nanodisc was challenging," Notti says.

Previous structural studies of the TCR had relied on detergent, which often strips away the surrounding membrane. Walz notes that this was the first time the receptor complex had been restored to a membrane environment for detailed imaging.

Seeing the Receptor Switch On

Once the TCR was embedded in the nanodisc, the researchers used cryo-EM to visualize it. The images showed that the receptor remains closed and compact when inactive. When it encounters an antigen-presenting molecule, however, the structure opens and extends outward, resembling a wide-reaching motion.

The result surprised the team. "The data that were available when we began this research depicted this complex as being open and extended in its dormant state," Notti explains. "As far as anyone knew, the T cell receptor didn't undergo any conformational changes when binding to these antigens. But we found that it does, springing open like a sort of jack-in-the-box."

The researchers believe two factors made this discovery possible. First, they carefully recreated the TCR's in vivo membrane environment using the right lipid mixture. Second, they reinserted the receptor into a membrane using nanodiscs before conducting cryo-EM imaging. They found that an intact membrane keeps the receptor in a closed position until activation occurs. In earlier studies, detergent may have removed this restraint, allowing the receptor to open prematurely.

"It was important that we used a lipid mixture that resembled that of the native T cell membrane," says Walz. "If we had just used a model lipid, we wouldn't have seen this closed dormant state either."

Implications for Cancer Therapies and Vaccines

The team believes their findings could help improve treatments that rely on T cell receptors. "Re-engineering the next generation of immunotherapies tops the charts in terms of unmet clinical needs," Notti says. "For example, adoptive T cell therapies are being used successfully to treat certain very rare sarcomas, so one could imagine using our insights to re-engineer the sensitivity of those receptors by tuning their activation threshold."

Walz also sees potential applications beyond cancer therapy. "This information may be used for vaccine design as well," he adds. "People in the field can now use our structures to see refined details about the interactions between different antigens presented by HLA and T cell receptors. Those different modes of interaction might have some implication for how the receptor functions -- and ways to optimize it."

Journal Reference:

1. Ryan Q. Notti, Fei Yi, Søren Heissel, Martin W. Bush, Zaki Molvi, Pujita Das, Henrik Molina, Christopher A. Klebanoff, Thomas Walz. The resting and ligand-bound states of the membrane-embedded human T-cell receptor–CD3 complex. Nature Communications, 2025; 16 (1) DOI: 10.1038/s41467-025-66939-7 

Courtesy: Rockefeller University. "A hidden T cell switch could make cancer immunotherapy work for more people." ScienceDaily. ScienceDaily, 18 December 2025. <www.sciencedaily.com/releases/2025/12/251218074429.htm>.