Annotation-Free Prediction of Cancer Cells and Glands and Spatial Analysis of Immune Cells

Kyeong Joo Jung, Soumya Ghose, Sanghee Cho, Elizabeth McDonough, Chrystal Chadwick, Robert B. West, James D. Brooks, Dongjun Chung, Fiona Ginty, Raghu Machiraju, Parag MallickbioRxiv 2025 · DOI

TL;DR

This paper introduces TOPAZ, a novel computational framework that integrates single-cell protein data with histology images to automatically classify cancerous and non-cancerous cells and glands in prostate tissue. This tool enables detailed spatial analysis of immune cell interactions with cancer, revealing new insights into disease progression and predicting biochemical recurrence risk.

What Problem Does This Paper Address?

Prostate cancer is often described as "immune-cold," but immune cells are present, and their spatial relationship with cancer cells and glands is poorly understood. This knowledge gap is partly due to the lack of automated tools that can accurately identify and classify both cancerous and non-cancerous cells and glandular structures within tissue samples. Understanding these spatial relationships is crucial for deciphering cancer progression and identifying prognostic markers.

Methods

The core of this work is the TOPAZ (Tissue Organization identification using sPAtial proteomics) pipeline. It combines two key approaches:

  • Annotation-Free Cell Classification: TOPAZ uses a Gaussian Mixture Model (GMM) to classify epithelial cells as cancerous (CE) or non-cancerous (N-CE) based on protein expression profiles (specifically p63 and CK5 markers). This method avoids the need for manual expert annotation, making it scalable.
  • Gland Shape Modeling: For gland classification, TOPAZ employs a principal curve estimation method. This approach models the intrinsic shape of N-CE cells to define gland contours and then assigns cell types based on their proximity to these curves. This allows for the prediction of gland boundaries and the extraction of geometric features.

After classification, TOPAZ extracts multiscale spatial features, such as distances between cancerous glands and immune cell regions, and correlates these with clinical outcomes.

Datasets

The study utilized tissue microarrays (TMAs) from 217 prostate cancer patients. These TMAs contained 753 cores and were analyzed using multiplexed immunofluorescence imaging (Cell DIVE, Leica). The imaging captured various markers, including:

  • Epithelial cell markers (panCK26, S6, NaKATPase)
  • Basal cell markers (p63, CK5)
  • A cancer cell marker (AMACR)
  • Immune cell markers (CD3, CD4, CD8, FOXP3, CD68)

The TOPAZ tool itself is available as an open-source Shiny app.

Key Insights and Ideas

  • Automated, Annotation-Free Classification: The ability to classify cells and glands without manual expert input is a significant step towards automating digital pathology workflows. This is crucial for analyzing large patient cohorts.
  • Linking Spatial Features to Prognosis: The study demonstrates that the spatial arrangement of immune cells relative to cancerous glands is not just descriptive but directly predictive of patient outcomes, specifically biochemical recurrence.
  • Immune Cell Dynamics with Grade: The research reveals how the proportions of different immune cell types (T-helper, regulatory T-cells, macrophages) change with increasing cancer grade, offering insights into the tumor microenvironment's evolution.
  • Macrophage-Tumor Spatial Coupling: The findings highlight a strong correlation between macrophage distribution (area, proximity to glands, infiltration) and cancer grade, suggesting a significant role for macrophages in disease progression.

Key Results

  • Cell Classification Accuracy: TOPAZ achieved 82% sensitivity and 99% specificity for cancerous/non-cancerous cell classification against expert annotations.
  • Gland Classification: The pipeline successfully predicted cancerous/non-cancerous glands without manual threshold tuning.
  • Immune Cell Trends: Helper T-cell and macrophage proportions increased with cancer grade (P < 0.001), while regulatory T-cell proportions remained stable (P = 0.065).
  • Recurrence Risk Prediction: Patients with greater distances between cancerous gland centroids and regulatory or helper T-cells (around 50 µm) had a doubled hazard of biochemical recurrence (log-rank P < 0.01).
  • AMACR Validation: Cancerous epithelial cells consistently showed elevated AMACR expression compared to non-cancerous cells across all grades, validating the cell classification.

Limitations and Future Work

  • Sample Size for Specific Grades: While the overall study is large, some specific analyses, particularly those stratifying by grade for certain immune cell types, might have been limited by sample size, leading to non-significant trends.
  • Future Improvements: The authors envision future developments for TOPAZ including adaptive biomarker-threshold learning, seamless integration with existing clinical platforms like QuPath, and extending the principal-curve strategy to other gland-forming cancers. The goal is to create a foundation for next-generation spatial pathology that links tissue organization to disease trajectory, treatment response, and precision oncology.

Summary updated

Abstract

Prostate cancer is classified as "immune-cold" due to limited infiltration of immune cells and no clear correlation between immune cells and clinical outcomes. However, immune cells are found in prostate cancers and the spatial relationships between these immune cells and cancer cells/glands have not been investigated, partly due to a lack of automated tools that classify both cancerous cells/glands. In this paper, we have developed an end-to-end tool (TOPAZ: <b>T</b>issue <b>O</b>rganization identification using s<b>PA</b>tial proteomics) that combines multiplexed single-cell protein data with histology images to: 1) predict cancerous versus non-cancerous epithelial cells using a Gaussian-mixture model; 2) predict cancerous/non-cancerous gland type using a principal curve estimation. Using TOPAZ to assign cancer and non-cancerous labels to cells and glands, we extracted multiscale spatial features from the classification results-including immune dense-region geometrical features and cell-to-gland distances-and correlated the features with risk of biochemical recurrence and cancer grade. Tissue-microarrays containing 753 cores from 217 prostate cancer patients underwent multiplexed immunofluorescent imaging (Cell DIVE, Leica) for epithelial cell markers (panCK26, S6, NaKATPase), basal cell markers (p63, CK5), a cancer cell marker (AMACR), and T cell markers (CD3, CD4, CD8, FOXP3, CD68). Cancerous/non-cancerous cell classification from TOPAZ achieved 82% sensitivity and 99% specificity against expert annotation, and the pipeline further predicted cancerous/non-cancerous glands without manual threshold tuning. Regulatory-T-cell and helper-T-cell percentages decreased, and macrophage percentage increased with grade increase (P < 0.05). When the median distance from cancerous gland centroids to the nearest regulatory or helper T-cell exceeded approximately 50 μm, the hazard of biochemical recurrence doubled (log-rank P < 0.01). The open-source Shiny app TOPAZ (https://chunglab.bmi.osumc.edu/TOPAZ) packages the workflow, predicting individual cell types and gland shapes. By combining probabilistic cell typing with gland-shape modeling, TOPAZ yields interpretable multiscale spatial features linked to prognosis and is released as an open web app for unrestricted use.
BibTeX
@misc{jung2025annotation,
  author = {Kyeong Joo Jung and Soumya Ghose and Sanghee Cho and Elizabeth McDonough and Chrystal Chadwick and Robert B. West and James D. Brooks and Dongjun Chung and Fiona Ginty and Raghu Machiraju and Parag Mallick},
  title = {{Annotation-Free Prediction of Cancer Cells and Glands and Spatial Analysis of Immune Cells}},
  year = {2025},
  doi = {10.1101/2025.11.09.687528},
  howpublished = {bioRxiv},
}
Figure