Based on Cho et al., Science 2026
Upload a hematoxylin & eosin slide and receive automatic detection and maturation classification of tertiary lymphoid structures — a validated prognostic biomarker across 12 cancer types. Free, open-access, and backed by a pan-cancer spatial atlas of 25,088 TLS profiled from 3,071 whole-slide images.
The science
Tertiary lymphoid structures (TLS) are ectopic lymphoid aggregates that arise within tumors under chronic inflammatory conditions. They closely resemble secondary lymphoid organs — containing segregated B-cell and T-cell zones, follicular dendritic cell (FDC) networks, and high endothelial venules — and are critical regulators of anti-tumor immunity.
TLS presence has been associated with enhanced immune activity, improved responses to immune checkpoint blockade (ICB), and favorable survival outcomes across multiple cancer types. However, their maturation states, spatial locations relative to tumors, and context-dependent associations had not been systematically characterized at a pan-cancer scale — until now.
The pan-cancer TLS atlas by Cho et al. (Science 2026) combined spatial transcriptomics from 340 tissue sections with AI-enabled analysis of 3,071 H&E WSIs to build the first comprehensive, deployable framework for TLS profiling and patient stratification.
Cho et al., Science 392, eadz2742 (2026)TLS maturation and spatial context are linked to distinct local immune programs and tumor signaling gradients, varying substantially across IT, PT, and DT tumor regions.
A maturation-aware composite TLS score — capturing within-tumor E-, P-, and S-TLS composition — robustly stratifies patients by survival and ICB treatment response across cancer types.
YOLOv8-based deep learning detects and classifies TLS maturation states directly from routine H&E slides with 94.7% overall accuracy, validated across 10 independent cohorts.
Findings were validated using ultrahigh-plex imaging, single-cell spatial profiling, and independent therapy cohorts spanning chemotherapy, targeted therapy, and ICB.
Classification
TLS are classified by unsupervised clustering of canonical follicular dendritic cell (FDC) and germinal center (GC) B-cell markers, consistent with established models of TLS development.
Lymphocyte aggregates without organized follicular structure. Lack FDC networks and germinal center (GC) markers. Associated with proliferative and stress-adaptive tumor programs in nearby tissue.
CD3D — T cell markerMS4A1 — B cell markerSegregated B- and T-cell zones with FDC presence but without a fully developed germinal center. Intermediate maturation with increased immune activation and complement pathway activity.
CR2, FCER2 — FDC markersWell-organized B- and T-cell zones with a localized germinal center. The most mature state — associated with improved ICB response, antibody production, and favorable overall survival across cancers.
BCL6, MKI67 — GC B cell markersCD4, ICOS, PD-1)IL10, CD24) enrichmentTLS located within two spots of tumor tissue. Most frequent in LUAD.
TLS within two spots of tumor tissue boundary.
TLS greater than two spots from tumor. Higher fractions in LIHC, CRC, BLCA, and KIRC.
The platform
Our pipeline reproduces the AI framework from Cho et al. — adapted for routine clinical H&E histopathology without spatial transcriptomics infrastructure.
Upload any hematoxylin & eosin whole-slide image in standard formats. Patient data is de-identified and handled securely. Supported cancer types span all 12 validated in the atlas.
A HookNet-TLS deep learning model tiles the WSI at multiple resolutions and identifies candidate TLS regions, yielding individual TLS instances with spatial coordinates.
A YOLOv8-based classifier — trained on 956 H&E images from TCGA LUSC, BLCA, and KIRC cohorts — assigns each TLS to E-, P-, or S-TLS with 94.7% overall accuracy (95% CI 91.4–97.0).
A maturation-aware composite TLS score (PCA of log-transformed, z-score–normalised E/P/S-TLS counts) stratifies your sample into C1 (mature, favorable) or C2 groups with survival context.
Coverage
The atlas spans 340 spatial transcriptomics samples from 255 patients across 12 solid tumor types, validated in 10 independent cohorts.
Open science
This platform implements the computational framework described in the pan-cancer TLS atlas study published in Science in 2026. The underlying AI model and analysis pipeline are open-source and available on GitHub.
The composite TLS score derived from E-, P-, and S-TLS state composition consistently outperformed conventional TLS metrics — including simple TLS presence/absence and dominant-state classification — in prognostic prediction across all six validated cancer cohorts (BLCA, LUSC, LUAD, STAD, COAD, KIRC).
For research use only. Not validated for clinical diagnosis or treatment decisions.
Pan-cancer spatial atlas of tertiary lymphoid structures
Kyung Serk Cho, Yunhe Liu, Guangsheng Pei, Jianfeng Chen, Yibo Dai et al.
Science 392, eadz2742 (2026)
Free, open-access analysis for research use. No infrastructure required.
For research use only — not validated for clinical diagnosis.