Pancreatic ductal adenocarcinoma (PDAC) is characterized by a densely immunosuppressive tumor microenvironment that physically and functionally excludes anti-tumor immune cells. This immune exclusion is one of the primary reasons checkpoint inhibitors, which work by reactivating exhausted T cells, have largely failed in PDAC clinical trials.
T cells are the central adaptive immune effectors capable of recognizing and killing cancer cells. However, in PDAC, T cells that infiltrate the tumor frequently become exhausted, lose cytotoxic function, or are physically segregated from the cancer cells by dense stromal barriers. Understanding exactly which T-cell states and spatial configurations predict outcomes could guide better therapeutic strategies.
A new class of immunotherapy agents, anti-CD40 agonist antibodies, aims to overcome PDAC immunosuppression by activating antigen-presenting cells and reprogramming the tumor stroma to allow T-cell infiltration. However, whether and how anti-CD40 therapy reshapes the T-cell compartment in PDAC was not well characterized.
This study used multiplexed immunohistochemistry (mIHC) combined with machine learning to systematically characterize T-cell functional states, spatial organization, and their associations with clinical outcomes and anti-CD40 therapy response in a PDAC patient cohort.
A 21-antibody multiplexed immunohistochemistry (mIHC) panel was applied to tissue sections from 29 PDAC patients, enabling simultaneous detection of T-cell markers, activation and exhaustion molecules, and structural landmarks within the tumor microenvironment. This generated cellular phenotype data for approximately 2.5 million individual cells across 306 tissue regions.
Single-cell phenotyping identified 18 distinct T-cell states using combinations of functional markers including CD44 (memory/activation), CD69 (early activation/tissue residency), PD-1 and TIM-3 (exhaustion), and lineage markers like CD4 and CD8. Each T cell was assigned to one of these 18 states based on its full marker expression profile.
From the 18 T-cell states, the study derived 961 T-cell functionality barcodes, representing all biologically meaningful functional combinations, and 268 cell-cell interaction pairs capturing how different T-cell subtypes co-localize with cancer cells, stromal cells, and each other. Together these generated a feature space of 1,252 tumor microenvironment features.
Tissue regions were classified into recurrent cellular neighborhoods (RCNs) using K-means clustering on cellular composition, yielding 10 spatially coherent microenvironmental zones such as tumor nests, stromal regions, and immune aggregates. This spatial framework provided the context for interpreting which T-cell states were operating in each microenvironmental compartment.
Elastic net (EN) regularized logistic regression was applied to the 1,252 tumor microenvironment features to build predictive models for anti-CD40 treatment status, disease-free survival (DFS), and overall survival. Elastic net combines LASSO and ridge penalties to perform simultaneous feature selection and regularization, making it well-suited to high-dimensional biological datasets.
Models were trained using leave-one-out cross-validation (LOOCV) to account for the limited sample size of 29 patients. This approach provides the most conservative and reliable performance estimate when external validation data is unavailable, ensuring that reported performance metrics are not inflated by overfitting.
The elastic net classifiers achieved AUC values of 0.87 to 0.90 for predicting anti-CD40 treatment status, demonstrating that the T-cell state and spatial composition of the tumor microenvironment is substantially reshaped by the therapy and that machine learning can detect these changes reliably.
SHAP (SHapley Additive exPlanations) values were calculated to interpret model predictions at the individual feature level. SHAP quantifies each feature's contribution to each prediction, revealing which specific T-cell states and spatial relationships were most important for distinguishing treated from untreated tumors and high-risk from low-risk patients.
Anti-CD40 therapy produced a significant reduction in T-cell exhaustion markers, particularly decreasing the density of PD-1+ TIM-3+ double-positive exhausted T cells within tumor nests. This suggests that anti-CD40 can partially reverse the exhausted T-cell phenotype that characterizes untreated PDAC.
The treatment also increased both the overall density and proximity of CD4+ T cells to cancer cells, particularly CD44+ memory-activated CD4+ T cells that adopt a Th1 effector phenotype. These cells were more frequently found in immune aggregate neighborhoods after treatment compared to pre-treatment biopsies.
The most important predictive feature identified by SHAP analysis was the density of CD44+CD4+ Th1 cells within immune aggregates, specifically in recurrent cellular neighborhood RCN5. Patients with higher densities of these cells in RCN5 had significantly longer disease-free survival after resection, identifying a specific spatiotemporal T-cell signature of favorable outcome.
These findings provide mechanistic insight into why some PDAC patients respond to anti-CD40 therapy while others do not. The presence of pre-existing or therapy-induced CD44+CD4+ Th1 cells in organized immune aggregates may represent a permissive microenvironmental state for durable anti-tumor immunity.
The identification of CD44+CD4+ Th1 cells in immune aggregate RCN5 as a predictor of disease-free survival suggests a potential tissue-based biomarker that could be assessed at the time of surgical resection to stratify patients for adjuvant therapy intensity or clinical trial enrollment.
If validated in larger cohorts, measuring the spatial density of Th1-phenotype CD4+ T cells in immune aggregates could serve as a complement to standard pathological staging, identifying patients at high or low risk of early recurrence who might benefit from more or less intensive postoperative treatment respectively.
The finding that anti-CD40 therapy reshapes immune aggregate composition provides a mechanistic rationale for neoadjuvant anti-CD40 treatment before surgery. If treatment can build CD44+CD4+ Th1 cell aggregates in the tumor, surgery may be performed in a more immunologically favorable context, potentially improving long-term outcomes.
The small cohort of 29 patients is a key limitation, and the clinical translation of these findings will require prospective validation in larger multicenter trials. Nevertheless, this study establishes a framework for integrating spatially-resolved T-cell phenotyping into biomarker-driven trial design for PDAC immunotherapy.
This study demonstrates that machine learning applied to high-dimensional multiplexed imaging data can extract clinically meaningful signals from the spatial architecture of the tumor immune microenvironment. The combination of mIHC, cellular neighborhood analysis, and elastic net modeling provides a scalable framework for mapping T-cell biology in solid tumors.
The use of SHAP interpretability tools transforms black-box predictions into biologically interpretable insights, connecting machine learning outputs back to specific cell types and spatial contexts that can be understood and acted upon by clinicians and researchers. This interpretability is critical for clinical adoption.
The 1,252-feature tumor microenvironment dataset generated in this study provides an extraordinarily rich resource for hypothesis generation. Beyond the CD44+CD4+ Th1 cell finding, the full dataset may harbor additional predictive signals for other therapeutic modalities including CTLA-4 blockade, KRAS-targeted therapies, or combinations of immunotherapy with chemotherapy.
Ultimately, this work contributes to the broader vision of precision immuno-oncology: moving beyond population-level treatment decisions toward patient-specific therapy selection guided by the functional and spatial architecture of each individual tumor's immune microenvironment.