Pancreatic ductal adenocarcinoma (PDAC) is one of the most lethal cancers, with a 5-year survival rate below 10%. Its aggressive behavior is driven not only by the tumor cells themselves but also by the surrounding tumor microenvironment, which typically suppresses anti-tumor immune activity and promotes cancer stemness.
Three cellular populations are particularly important in shaping PDAC outcomes. CD8+ cytotoxic T lymphocytes (CTLs) are the primary anti-tumor immune effectors, capable of directly killing cancer cells when present in sufficient numbers and proximity to the tumor. Their abundance in the tumor microenvironment correlates with better patient outcomes.
In opposition to CTLs, cancer stem cells (CSCs) expressing CD133 drive tumor self-renewal, metastasis, and resistance to therapy. Similarly, tumor budding (TB), the detachment of small clusters of dedifferentiated cancer cells at the invasion front, is a histological marker of aggressive behavior associated with lymph node involvement and worse survival.
The balance between these anti-tumor and pro-tumor cellular populations has never been quantitatively integrated into a single spatial, AI-assisted prognostic framework. Developing such a framework could move beyond subjective pathologist evaluation toward reproducible, objective measurement of the immune-stemness-tumor budding axis.
The study used Tissue Gnostics, an automated AI image analysis platform, to quantify cellular populations on multiplexed immunofluorescence stained tissue sections. This system allowed simultaneous detection of multiple biomarkers within the same tissue section, enabling spatial analysis of cell-to-cell relationships.
Three tissue markers were analyzed: CD8 for cytotoxic T lymphocytes, CD133 for cancer stem cells, and CK19 (cytokeratin 19) for tumor budding identification. CK19 marks the epithelial origin of tumor cells and highlights individual budding cells or small clusters that have separated from the main tumor mass.
Spatial proximity analysis was performed using a 20-micron radius threshold to define neighborhood relationships between cell types. This generated four spatially-informed metrics: tumor bud-adjacent CD8+ T cells, CSC-adjacent CD8+ T cells, the CD8/TB index, and the CD8/CD133+ CSC index.
This approach transformed conventional immunohistochemistry into a quantitative, spatially resolved cellular map, capturing not just how many of each cell type were present but where they were located relative to each other, which is biologically more meaningful than density counts alone.
The study analyzed four patient cohorts: a primary training cohort of 160 patients, a retrospective validation cohort of 108 patients, a prospective validation cohort of 63 patients, and an external validation cohort of 95 patients. This multi-cohort design strengthens the generalizability of the findings.
Two primary prognostic indices were derived from the spatial cellular analysis. The CD8/TB index represents the ratio of CD8+ T cells to tumor buds, capturing the balance between immune attack and invasive cell detachment. The CD8/CD133+ CSC index represents the ratio of cytotoxic T lymphocytes to cancer stem cells, reflecting the balance between adaptive immunity and stemness-driven recurrence.
Both indices were tested as independent predictors of overall survival using multivariate Cox proportional hazards regression, adjusting for standard clinicopathological variables including tumor size, lymph node status, resection margin, and TNM stage.
A simplified nomogram was constructed by combining the CD8/TB and CD8/CD133 indices with selected clinical variables. The nomogram was designed to translate complex spatial biomarker measurements into a practical bedside tool for individual patient risk stratification after surgery.
Both the CD8/TB index and the CD8/CD133+ CSC index were confirmed as independent prognostic factors for overall survival in multivariate analysis. Patients with high CD8/TB ratios (more immune cells relative to tumor buds) had significantly longer survival than patients with low ratios.
The simplified nomogram achieved a concordance index (C-index) of 0.746 in the training cohort and 0.755 in the retrospective validation cohort. These values exceed the C-index achievable with TNM staging alone, demonstrating that the spatial immune-stemness profile adds meaningful prognostic information beyond standard pathological assessment.
In the prospective and external validation cohorts, the nomogram maintained robust discrimination between high-risk and low-risk patients, confirming that the prognostic signals captured by CD8/TB and CD8/CD133 indices are reproducible across independent patient populations and different staining and imaging conditions.
Survival curve analysis showed that the difference in outcomes between the highest and lowest risk groups stratified by the nomogram was substantial and clinically meaningful, suggesting the tool could meaningfully influence postoperative management decisions about adjuvant chemotherapy intensity or surveillance frequency.
Beyond survival prediction, the immune-stemness-tumor budding profile was tested as a predictor of response to immune checkpoint blockade (ICB). Patients with high CD8/TB and CD8/CD133 ratios showed characteristics consistent with an immune-responsive tumor microenvironment.
Analysis of publicly available immunotherapy datasets confirmed that tumors with higher CD8+ T-cell infiltration relative to CSC and tumor budding burden showed significantly higher rates of objective response to checkpoint inhibitors, suggesting the indices could serve as predictive as well as prognostic biomarkers.
A humanized patient-derived xenograft (PDX) mouse model was used to validate these findings in vivo. PDX mice whose tumors had high CD8/TB ratios responded more favorably to anti-PD-1 therapy, with greater tumor volume reduction compared to mice with low CD8/TB ratios.
These results suggest that the AI-derived spatial index does not merely reflect existing immune activity but may also help identify which patients are most likely to benefit from adding immunotherapy to standard surgical and chemotherapy regimens, addressing one of the most pressing unanswered questions in PDAC management.
This study establishes a novel AI-assisted precision immunopathology framework that integrates three spatially-resolved biomarkers into a clinically actionable prognostic tool. By quantifying the interplay between cytotoxic immunity, stemness, and invasiveness, the model captures the full immunosuppressive architecture of PDAC.
The validation across four independent cohorts including a prospective cohort and an external institutional cohort demonstrates that the immune-stemness-tumor budding profile is robust, reproducible, and translatable to different clinical settings. This is a critical advantage over biomarkers validated only in retrospective single-institution studies.
The addition of spatial proximity metrics to conventional marker counts represents a methodological advancement that better reflects the biological reality of tumor immunology. Cell proximity matters: CD8+ T cells that are physically adjacent to cancer cells exert different effects than those sequestered at the tumor periphery.
Future studies should evaluate whether the nomogram can guide treatment allocation decisions in prospective randomized trials, and whether spatial cellular quantification using AI platforms can be standardized across institutions for routine clinical pathology reporting. This framework represents a blueprint for AI-enabled precision pathology in pancreatic cancer.