The maturation state, spatial location and composition tell us critical information
Inside tumors, the immune system sometimes builds its own command centers — clusters of cells that organize to fight cancer. Researchers at MD Anderson have now mapped these structures across twelve cancer types, discovering that their maturity and position within the tumor matter far more than their mere presence. The work, powered by artificial intelligence and grounded in tens of thousands of pathology images, offers a new vocabulary for reading the immune landscape of cancer — one that may ultimately guide more precise and effective treatment.
- For years, oncologists treated the presence of immune hubs in tumors as a good sign, but that binary thinking left critical clinical information on the table.
- A new AI-powered atlas of 25,088 immune structures across 3,000+ pathology images reveals that maturation state and location — not just existence — determine how well a tumor's immune response will perform.
- The AI framework works with standard pathology slides already in clinical use, making this level of immune analysis scalable across hospitals without requiring exotic new equipment.
- A composite scoring system built from this data significantly outperforms older TLS measures in predicting patient prognosis and likely immunotherapy response.
- Prospective clinical trials are now needed to validate the approach before it can be woven into routine pathology workflows and used to guide treatment decisions.
Tumors are not solitary battlegrounds. Inside them, the immune system sometimes organizes itself into clusters of B cells, T cells, and supporting cells — local command centers called tertiary lymphoid structures, or TLSs — that work to fight cancer from within. Researchers at MD Anderson Cancer Center have now mapped these structures across twelve cancer types, and their findings suggest that how these immune hubs are organized matters far more than whether they exist at all.
For years, clinicians knew that TLSs were generally a good sign. But that knowledge was surface-level. The new research, led by genomic medicine professor Linghua Wang and published in Science, reveals that two tumors might both contain TLSs yet differ dramatically in how mature those structures are, where they sit relative to cancer cells, and what cells they contain. Those differences carry real clinical weight.
To reach this understanding, Wang's team built an AI framework capable of detecting and classifying TLSs from high-resolution spatial omics data, then applied it to 340 tumor samples across twelve cancer types. As these structures mature, they become more organized — with coordinated shifts in immune cells, stromal cells, and blood vessels. The team also found that a TLS's proximity to tumor cells creates spatial gradients of signaling, suggesting location is functionally important, not incidental.
The clinical innovation came when the researchers trained their AI on routine pathology slides — the kind already used in everyday practice — and evaluated more than 25,000 individual TLSs across 3,000 whole-slide images from ten independent patient cohorts. From this dataset, they developed a composite score capturing not just how many TLSs a tumor contains, but their maturation states. This score significantly outperformed older methods that simply noted presence or absence.
The next step is prospective clinical trials to validate the scoring approach. If successful, TLS profiling could become a routine part of pathology workflows. The findings also raise deeper biological questions: many TLSs in tumors remain immature, and some sit too far from cancer cells to engage effectively. Future research may explore how to push these structures toward maturity and better positioning — potentially opening new therapeutic strategies that actively promote immune function rather than simply measuring it.
Tumors are not solitary battlegrounds. Inside them, the immune system sometimes organizes itself into distinct structures—clusters of B cells, T cells, and supporting cells that function like local command centers for fighting cancer. Researchers at The University of Texas MD Anderson Cancer Center have now mapped these structures, called tertiary lymphoid structures or TLSs, across twelve different cancer types, and what they found suggests that how these immune hubs are organized matters far more than whether they exist at all.
The work, published in Science and led by Linghua Wang, a professor of genomic medicine at MD Anderson, represents a significant shift in how oncologists might think about the immune landscape within tumors. For years, clinicians knew that TLSs were generally a good sign—their presence correlated with better outcomes and stronger responses to immunotherapy. But that knowledge was surface-level. The new research reveals that TLSs are far more nuanced. Two tumors might both contain these immune structures, yet differ dramatically in how mature those structures are, where they sit relative to cancer cells, and what cellular components they contain. Those differences, it turns out, carry real clinical weight.
To reach this understanding, Wang's team built what amounts to an encyclopedia of immune architecture. They developed artificial intelligence frameworks capable of detecting and classifying TLSs from spatial omics data—high-resolution maps showing where different molecules and cells sit within tissue samples. They then applied these tools to 340 tumor samples spanning twelve cancer types, creating a pan-cancer atlas that revealed how TLSs vary across tissues. As these structures mature, they become more organized, with coordinated shifts in immune cells, supporting stromal cells, and blood vessels. The team also discovered that a TLS's proximity to tumor cells creates spatial gradients of tumor signaling, suggesting that location is not incidental but functionally important.
The real clinical innovation came next. The researchers trained an AI model to identify and classify TLSs from routine pathology slides—the kind of images already used in everyday clinical practice. This made the analysis scalable in a way that spatial omics data alone could never be. Using this approach, they evaluated more than 25,000 individual TLSs across 3,000 whole-slide images from ten independent patient cohorts. From this massive dataset, they developed a composite scoring system that captures not just how many TLSs a patient's tumor contains, but their maturation states. This composition score significantly outperformed older methods that simply counted whether TLSs were present or absent.
Linghua Wang described the shift in thinking this way: the field had been asking a binary question—are TLSs there or not?—when the real story was far more complex. The maturation state, spatial location, and cellular makeup of these structures reveal critical information about the tumor's immune microenvironment and how it will respond to treatment. A tumor might harbor many TLSs that remain immature, or structures positioned far from cancer cells where they cannot effectively engage. These details matter for prognosis and for predicting which patients will benefit from immunotherapy.
The next phase is validation. The composite scoring approach must be tested in prospective clinical trials to confirm that it genuinely improves patient stratification and treatment planning. If those trials succeed, TLS profiling could become a routine part of pathology workflows, since the AI framework works with standard imaging that pathologists already generate. Beyond clinical validation, the findings raise deeper biological questions. Many TLSs in tumors remain immature, and some are positioned away from cancer cells entirely. This suggests that future research should explore how to push TLSs toward maturity and how to position them where they can most effectively attack tumors. Such work might eventually lead to new therapeutic strategies that not only harness existing immune structures but actively promote their formation and function.
For now, the atlas stands as a proof of concept: that artificial intelligence can reveal hidden complexity in cancer's immune landscape, and that this complexity, once mapped, becomes actionable. The immune system's response to cancer is not a simple presence or absence. It is a landscape of structures at different stages of development, positioned in different locations, composed of different cells—and understanding that landscape may be the key to better predicting and improving how patients respond to treatment.
Citas Notables
TLSs in tumor tissues are much more complex than that. Their maturation state, spatial location and composition within tumors can tell us critical information about the tumor immune microenvironment, treatment response and clinical outcomes.— Linghua Wang, Professor of Genomic Medicine, UT MD Anderson
La Conversación del Hearth Otra perspectiva de la historia
Why does it matter whether a tertiary lymphoid structure is mature or immature? Isn't an immune hub an immune hub?
Not quite. An immature TLS might have the right cells present, but they're not yet organized into a coordinated fighting force. A mature one has undergone changes in how its immune cells, blood vessels, and supporting tissue are arranged. That organization determines whether it can actually mount an effective attack on tumor cells.
And location matters too?
Significantly. A TLS positioned far from tumor cells is like having a well-trained army stationed miles away from the battlefield. The proximity affects how tumor signaling reaches the immune structure and how effectively immune responses can reach back to cancer cells. The study found that spatial gradients of tumor signaling are tied to where TLSs sit.
So the old way of thinking—just checking whether TLSs are present—was missing most of the story?
Entirely. It's like asking whether a city has hospitals without asking whether they're staffed, equipped, or located where patients can reach them. The presence alone doesn't tell you much about actual clinical outcomes.
How does the AI fit in?
The researchers needed to analyze thousands of TLSs across hundreds of patient samples to see these patterns. AI let them do that at scale using routine pathology images—the kind already taken in clinics. Without automation, this would have been impractical.
What happens if the validation trials work?
Then pathologists could get a TLS composition score as part of standard tumor analysis, helping oncologists predict which patients will respond to immunotherapy and which might need different approaches. It becomes a practical tool, not just research.
What's the biggest remaining question?
Why so many TLSs remain immature and poorly positioned. If we understood how to mature them and move them closer to tumor cells, we might be able to engineer better immune responses—not just detect them, but actively improve them.