Osteopontin Fingerprints in Tissue Scans Predict Pancreatic Cancer Survival

Pancreatic ductal adenocarcinoma remains one of medicine’s most unforgiving diagnoses. Despite decades of progress in surgical technique and chemotherapy, the five-year survival rate sits below 12 percent, and for patients whose disease has spread to distant organs, it falls under 5 percent. The reasons are familiar to oncologists: the disease is usually found late, it metastasizes aggressively, and it builds an extraordinarily immunosuppressive tumor microenvironment that shields it from both drugs and immune attack. Adding to the challenge, pancreatic tumors vary enormously from patient to patient, which undermines one-size-fits-all chemo- and immunotherapies and makes reliable prognostic biomarkers urgently needed. A new study from researchers at Shandong Cancer Hospital and Institute, published in Frontiers in Oncology, now reports a way to link a single immune-related molecule to measurable patterns in routine tissue images, producing a prognostic model that is both quantitative and interpretable.

The molecule at the center of the work is secreted phosphoprotein 1, better known as osteopontin, a multifunctional secreted glycoprotein that has attracted attention as a candidate prognostic biomarker in several cancers. In pancreatic ductal adenocarcinoma, SPP1 is upregulated in tumor tissue compared with normal pancreas and correlates with tumor progression and adverse outcomes. Its expression is controlled at multiple levels, including genetic variants, epigenetic modifications, and long non-coding RNAs. Functionally, SPP1 has been implicated in tumor-intrinsic processes such as mesenchymal cell fate determination and epithelial-mesenchymal transition, as well as in tumor-extrinsic immune modulation. Of particular importance, SPP1-positive tumor-associated macrophages have emerged as drivers of immunosuppression and therapeutic resistance in pancreatic cancer: they interact with PRRX2-positive epithelial cells through the TGFB1/TGFBR2 axis to promote resistance to gemcitabine, and their CXCL9-low/SPP1-high phenotype is associated with a non-inflamed tumor microenvironment that may be reversed by myeloid-targeting strategies.

The research team, led by Jinming Zhao, Mingjun Li, Jialiang Xing, Yunsong Chen, and Xingguo Song, began by mapping SPP1 expression across public datasets. Analyzing the GSE226840 dataset, which contains 14 pancreatic cancer-associated tertiary lymphoid structures and 5 adjacent normal tissues, they identified 85 differentially expressed genes, and SPP1 stood out among the downregulated genes in TLS-enriched regions. This was a striking counterpoint to the broader picture: systematic validation across four independent GEO cohorts, GSE15471, GSE62165, GSE28735, and GSE62452, confirmed consistent tumor-specific elevation of SPP1 mRNA, a pattern corroborated at the protein level using the CPTAC dataset. A pan-cancer analysis integrating The Cancer Genome Atlas and GTEx datasets showed significant SPP1 overexpression across multiple solid tumors, with pancreatic adenocarcinoma displaying particularly pronounced transcriptional upregulation. In other words, SPP1 is high in the tumor overall, yet comparatively low in the very immune structures associated with better outcomes.

That dichotomy carried prognostic weight. In the TCGA-PAAD cohort, patients with low SPP1 expression showed significantly better overall survival, with a hazard ratio of 2.65 (95 percent CI: 1.49 to 4.71, log-rank P = 0.00059), and better recurrence-free survival, with a hazard ratio of 5.13 (95 percent CI: 1.70 to 15.46, log-rank P = 0.0014), compared with patients whose expression was high. The finding was reproduced in the independent GSE85916 cohort, where low SPP1 expression was again associated with favorable prognosis (hazard ratio 1.97, 95 percent CI: 1.16 to 3.36, P = 0.0124). To confirm the signal at the protein level, the team performed immunohistochemistry on a commercial tissue microarray of patients who underwent curative-intent resection between January 2009 and August 2013. After excluding two specimens with poor slide quality, 63 evaluable cases remained, and whole-slide digital pathology quantification of the SPP1-positive area ratio showed that patients with low protein expression had significantly better overall survival (hazard ratio 2.15, 95 percent CI: 1.12 to 4.11, P = 0.0312).

The truly novel step was translating these molecular observations into measurable features of tissue architecture. Using a CellProfiler-based image processing pipeline, the researchers separated hematoxylin and DAB staining channels by color deconvolution, segmented nuclei and cytoplasm automatically, and extracted 361 handcrafted quantitative pathomics features from each tissue core, covering object shape, size, texture, and pixel intensity distribution. To distill this high-dimensional feature space, they deployed an ensemble feature selection strategy combining LASSO regression with 10-fold cross-validation, random forest, and decision tree, the latter two constrained to shallow trees with a maximum depth of three to prevent overfitting. LASSO initially identified five candidate pathomics features, which were then pooled with three clinical variables: sex, age, and tumor volume. Feature importance ranking revealed that Zernike_2_2, Hematoxylin_4of4, Zernike_1_1, and age were consistently selected by all three algorithms, while sex, tumor volume, Zernike_9_5, and Zernike_6_2 contributed minimally.

Zernike_1_1 proved the weakest contributor, with a mean absolute SHAP value of 0.054 and no significant association with overall survival (hazard ratio 0.936, P = 0.749), so it was dropped to keep the model parsimonious. The final Cox proportional-hazards model retained three features: Zernike_2_2, a Zernike moment descriptor capturing the shape characteristics of primary objects; Hematoxylin_4of4, a radial distribution measure of hematoxylin intensity variation; and age. Each patient received a risk score as a linear combination of the three features weighted by their Cox coefficients: Risk Score equals negative 0.6189 times Zernike_2_2 plus negative 0.6415 times Hematoxylin_4of4 plus 0.0390 times age. Splitting the 63 patients at the median risk score produced a high-risk group of 31 and a low-risk group of 32, and Kaplan-Meier analysis revealed a highly significant survival difference between the two groups (log-rank P < 0.0001), with an apparent C-index of 0.738.

Performance metrics held up under scrutiny. Time-dependent ROC analysis yielded an area under the curve of 0.866 at 12 months, and the calibration curve closely followed the 45-degree reference line with a slope of 0.881, indicating well-calibrated predictions. Decision curve analysis suggested potential net benefit, with positive net benefits across threshold probabilities from 0 to 0.6 and outperformance of both treat-all and treat-none strategies, particularly between 0.1 and 0.5. SHAP analysis clarified each predictor’s contribution: Zernike_2_2 was the most influential (mean absolute SHAP 0.475), followed by Hematoxylin_4of4 (0.445) and age (0.322), with directional effects consistent with the Cox regression results. Schoenfeld residual tests found no strong evidence of violation of the proportional-hazards assumption, although the global test P value of 0.055 was borderline and the authors urged cautious interpretation. Most importantly, internal validation with 1,000 bootstrap resampling iterations produced an optimism-corrected C-index of 0.715 (95 percent CI: 0.626 to 0.801) and a corrected 12-month AUC of 0.766 (95 percent CI: 0.645 to 0.877), indicating minimal overfitting.

Where does SPP1 come from inside the tumor? Single-cell transcriptomic profiling using the TISCH2 platform detected SPP1 expression across 13 major cell types, with prominent expression in myeloid-lineage immune cells, particularly macrophages and monocytes. Independent re-analysis of the CRA001160 dataset showed elevated SPP1 expression in macrophages and malignant cells, and visualization via the scCancerExplorer platform indicated predominant expression in macrophages and ductal cells. Immune deconvolution analyses using TIMER 2.0 with multiple algorithms, including quanTIseq, CIBERSORT, and TIDE, showed that SPP1 expression was positively associated with both M1 and M2 macrophage infiltration estimates in some algorithms, whereas negative associations with specific M2 subclusters emerged using TIDE. The authors interpret this algorithm-dependent variability as evidence that SPP1 is linked to macrophage phenotypic heterogeneity rather than a fixed polarization state, and they caution that correlation-based analyses cannot establish causation.

The team is candid about the study’s boundaries. The single-center retrospective design and the small sample of 63 patients limit statistical power, as reflected in the wide bootstrap confidence interval for the C-index, and multicenter external validation in larger independent cohorts will be essential before the model can be considered generalizable. The feature set excluded established serum biomarkers such as CA19-9, whose inclusion could improve accuracy, and the absence of spatial transcriptomics or multiplex immunohistochemistry means SPP1 localization relative to tertiary lymphoid structures and macrophage subsets could not be directly verified in the same tissue sections. The authors also note that they deliberately chose handcrafted, interpretable features over black-box deep learning or multiple-instance learning approaches, prioritizing traceability of feature selection and decision-making in line with the TRIPOD reporting guidelines and PROBAST risk-of-bias framework. Even so, the work offers a compelling proof of concept: a molecule long associated with macrophage-driven immune suppression and poor outcomes in pancreatic cancer leaves quantifiable fingerprints in routine stained tissue, and those fingerprints, captured by transparent computational pathology, can stratify patients by survival with striking statistical separation. If larger studies confirm the signal, SPP1-associated pathomics could become a practical bridge between tumor biology and the everyday pathology slide.

Subject of Research: Osteopontin Fingerprints in Tissue Scans Predict Pancreatic Cancer Survival

Article Title: SPP1-associated pathomics features for prognostic modeling in pancreatic ductal adenocarcinoma

Article References: Zhao, J., Li, M., Xing, J., Chen, Y., & Song, X. (2026). SPP1-associated pathomics features for prognostic modeling in pancreatic ductal adenocarcinoma. Frontiers in Oncology, 16, Article 1923860. https://doi.org/10.3389/fonc.2026.1923860

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1923860

Keywords: pancreatic cancer, SPP1, osteopontin, pathomics, prognostic model, tumor microenvironment, computational pathology, SPP1-associated, features, prognostic, modeling, pancreatic