3D Super-Cell Graphs Enhance Prostate Cancer Risk Stratification

Prostate cancer treatment decisions rely heavily on the examination of 2D histology sections. However, because these sections provide limited sampling of tissue volumes without 3D glandular and cellular context, interpretation can be misleading and ambiguous. We have developed a slide-free 3D pathology method that generates volumetric datasets that mimic the appearance of H&E staining. Using these 3D pathology datasets, we have shown that computational analysis of 3D histomorphometric features describing glandular and nuclear morphology offer better prognostic value than analogous 2D features. However, given that the spatial organization of cells in 3D remains largely unexplored, we investigate whether quantifying 3D cell–cell interactions can further enhance prognostication. To analyze 3D cell–cell interactions, we first applied a deep-learning segmentation model (Cellpose) to identify all cell nuclei from H&E-analog 3D pathology datasets, and then extracted 3D histomorphometric features from those nuclei, including shape descriptors, texture metrics, and contextual features. Since directly modeling the interactions between millions of cells in 3D is both computationally prohibitive and prone to overfitting, we grouped spatially adjacent and morphologically similar cells into “supercells”, reducing the data to approximately 10% of the original cell count (with exponentially less cell connections) while preserving the overall structure of the cell interactions. Leiden clustering based on nuclear histomorphometric features was used to group these supercells into broad cell types. We then constructed 3D supercell graphs using k-nearest neighbors, where each supercell is represented as a node and edges indicate supercell interactions. To evaluate the prognostic value of these 3D supercell graphs, we generated graphs for a cohort of 76 archived prostatectomy specimens with known 5-year biochemical recurrence (BCR) outcomes. Quantitative features extracted from the 3D supercell graphs were used as inputs to a LASSO model with 5-fold cross-validation, achieving an AUC of 0.745 ± 0.06 for predicting 5-year BCR. 3D histomorphometric features previously derived from cancer glands and nuclei masks yielded an AUC of 0.751 ± 0.08. Combining all of these features yielded an AUC of 0.837 ± 0.05. In parallel, we are assessing whether end-to-end graph neural network (GNN) learning provides additional prognostic value compared with traditional machine classifiers based on 3D supercell graph features, especially in a low-shot regime with limited cohort sizes. In summary, based on H&E-analog 3D pathology datasets, we have introduced 3D supercell graphs as a data-efficient method to capture broad cell interactions within the tumor microenvironment, showing that these 3D supercell graph features are of prognostic value. Citation Format: Yujie Zhao, Sarah Chow, Renao Yan, Rob Serafin, Elena Baraznenok, Lawrence D. True, Priti Lal, Anant Madabhushi, Jonathan TC Liu. Elucidating broad cell-cell interactions via 3D super-cell graphs for prostate cancer risk stratification [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 3997.