Tertiary lymphoid structures (TLS) are organized clusters of immune cells that form within tumors and surrounding tissue. Unlike lymph nodes - which are fixed organs that filter lymph - TLS arise spontaneously in response to chronic inflammation, including the immune response to cancer.
TLS are rich in B cells (antibody-producing immune cells) and T cells, and evidence from multiple cancer types suggests that patients whose tumors contain TLS tend to have better outcomes and respond more strongly to immunotherapy. However, manually identifying and measuring TLS in tissue slides is laborious and inconsistent.
A critical unanswered question is whether the location of TLS relative to the tumor matters, not just their presence. This study developed an AI model to automatically detect TLS and classify them as either close to (proximal) or far from (distal) the tumor, then tested whether location predicted survival and immunotherapy response.
The researchers trained a DeepLabV3_resnet101 model - a deep learning architecture designed for image segmentation - to identify TLS in H&E-stained whole-slide images. The model achieved a Dice coefficient of 0.945, meaning its TLS boundaries matched pathologist annotations with 94.5% overlap, indicating very high accuracy.
Once TLS were located, they were classified based on distance from the tumor edge: proximal TLS (pTLS) within 500 micrometers of the tumor, and distal TLS (dTLS) between 500 and 5,000 micrometers away. This spatial separation allowed testing whether proximity to the tumor changes TLS function and clinical value.
The cohort included endometrial cancer patients who received immune checkpoint inhibitor (ICI) therapy, allowing direct measurement of whether TLS location predicted actual immunotherapy response. B cell receptor (BCR) repertoire sequencing was also performed to compare B cell populations inside TLS versus within tumor-infiltrating lymphocytes.
TLS were detected in 69% of patients, but their presence alone did not predict outcomes. The key finding was that location mattered: distal TLS (dTLS) - those farther from the tumor - were significantly associated with better overall survival (HR=0.56, p=0.01) and better progression-free survival (HR=0.58, p=0.004). In contrast, proximal TLS were not significantly associated with improved outcomes.
The difference in immunotherapy response was striking: patients with distal TLS had an 87.5% response rate to immune checkpoint inhibitors, compared to only 41.7% in patients without distal TLS. This suggests that dTLS may be functioning as active immune hubs that support anti-tumor immunity.
The B cell receptor (BCR) repertoire analysis revealed that B cell clones within distal TLS showed significantly more overlap with B cell clones in tumor-infiltrating lymphocytes than proximal TLS did. This means the B cells in dTLS are more likely to be specifically responding to tumor antigens - they are actively recognizing and attacking the cancer.
The study proposes a biological explanation for why distal TLS outperform proximal ones: tumor microenvironments close to the cancer are rich in immunosuppressive signals. Proximal TLS may be exposed to factors like TGF-beta and adenosine that suppress B and T cell activity, limiting their effectiveness even when structurally present.
Distal TLS, being further from these suppressive signals, may maintain a more functional immune environment. The BCR repertoire data supports this - the higher overlap between dTLS B cell clones and tumor-infiltrating B cells suggests dTLS are generating tumor-specific immune responses that then traffic into the tumor.
These findings have implications for patient selection for immunotherapy. Rather than using TLS presence as a binary biomarker, spatial quantification - now automated by the AI model - could provide a more precise predictor of who will respond to checkpoint inhibitors.
The AI model developed here could serve as a biomarker tool in routine clinical pathology workflows. After scanning a tumor slide, the model automatically identifies TLS, measures distances to the tumor boundary, and generates a dTLS score - all without manual pathologist annotation.
This has immediate relevance for immunotherapy decision-making in endometrial cancer, where immune checkpoint inhibitors are increasingly used but not all patients respond. Current biomarkers like mismatch repair deficiency (MMR) status miss a subset of responders; dTLS status could complement MMR testing to identify additional patients likely to benefit.
The next steps include validating the model in larger, prospective cohorts and testing whether it generalizes across different EC subtypes and treatment settings. The researchers also suggest that interventions designed to promote TLS formation at specific distances from tumors - potentially through neoadjuvant therapy - could be a novel therapeutic strategy.