Topological Tumor Graphs: A Graph-Based Spatial Model to Infer Stromal Recruitment for Immunosuppression in Melanoma Histology

Cancer Res 2020 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Mapping the Stromal Barrier in Melanoma with Graph Theory

The Problem of Stromal Immunosuppression Metastatic melanoma remains deadly despite immunotherapy, partly because stromal cells in the tumor microenvironment block immune cells from reaching cancer cells. Understanding and measuring this stromal barrier requires methods that go beyond simply counting cell types.

What TTGs Are Topological Tumor Graphs (TTGs) are graph-based representations of tumor tissue where each cell becomes a node and edges connect cells within 35 micrometers of one another. This transforms chaotic histology slides into structured networks that can be analyzed with social network tools.

Study Scale The researchers applied TTGs to 400 melanoma H&E-stained whole-tumor sections from The Cancer Genome Atlas (TCGA), covering primary tumors, regional lymph node metastases, and distant metastases. This is one of the largest spatial analyses of the melanoma tumor microenvironment to date.

TL;DR: Topological Tumor Graphs convert melanoma histology slides into cell networks, enabling quantitative measurement of stromal organization across 400 TCGA specimens.
Pages 3-5
Building TTGs: Cell Classification, Graph Construction, and Network Metrics

Automated Cell Classification The CRImage computational pathology pipeline used watershed segmentation and a support vector machine with 97 morphological features to classify every nucleus as cancer, lymphocyte, stromal, or artifact. Validation against 3,230 manual annotations achieved 84.9% balanced accuracy.

Graph Construction Cancer cells that were closely packed were merged into supernode representations. Lymphocytes and stromal cells each remained as individual nodes. Edges were drawn between any two cells closer than 35 micrometers - a threshold set by measuring actual cancer-to-stroma distances in the slides.

Two Key Network Features Stromal clustering was calculated as the average clustering coefficient of stromal cells - measuring how tightly stromal cells group together. Stromal barrier counted how many stromal cells a lymphocyte would have to cross along the shortest path to reach a cancer cluster, capturing the physical impediment to immune infiltration.

CNx Deep Learning Integration An unsupervised encoder-decoder network called CNx integrated genome-wide copy number alterations with transcriptomic data to find genomic drivers of stromal phenotypes. The bottleneck layer compressed 15,667 genes into 200 nodes representing copy-number-driven gene expression patterns.

TL;DR: TTGs use automated cell classification and graph theory to compute stromal clustering and stromal barrier scores from H&E slides, with CNx linking these scores to genomic data.
Pages 7-8
Stromal Features Are Independent Predictors of Melanoma Survival

Stromal Clustering and Barrier Associate with Lymphocyte Exclusion Both stromal clustering and stromal barrier were significantly and inversely correlated with lymphocyte percentage across all tumor types. Tumors with high stromal clustering had fewer CD8 T cells; high barrier was associated with low-cytotoxicity immune phenotypes.

Survival Impact Is Independent of Known Factors In univariate analysis, high stromal clustering carried a hazard ratio of 1.97 and high stromal barrier a hazard ratio of 1.92 for 10-year overall survival (both P less than 0.05). Crucially, these associations held after adjusting for Breslow depth, ulceration, stromal percentage, and lymphocyte percentage.

Combined Phenotype Has the Strongest Effect Patients with high clustering plus high barrier had significantly worse survival than low clustering plus low barrier patients. Importantly, even among tumors with high lymphocyte counts, high stromal features still predicted worse outcomes, suggesting the spatial organization of stroma matters beyond mere quantity.

TL;DR: High stromal clustering and barrier independently predict poor melanoma survival, even after accounting for standard prognostic factors and immune cell content.
Pages 9-10
CNx Reveals Molecular Programs Underlying Stromal Recruitment

CNx Links Genotype to Stromal Phenotype CNx identified 578 genes associated with stromal barrier and 633 genes associated with stromal clustering. Critically, these genes showed significantly stronger correlations between copy number and gene expression compared to genes found by conventional differential expression, validating the deep learning approach.

Immune Pathways Are Suppressed in High-Stroma Tumors Genes downregulated in high-barrier and high-clustering tumors were enriched for TCR signaling in naive CD4 T cells, IL signaling, MAPK, and PI3K-Akt pathways. This suggests cancer genomic alterations actively suppress the immune-activating transcriptome.

Distinct Cytokine Networks for Each Feature The cytokine/chemokine genes associated with stromal clustering and barrier showed almost no overlap, sharing only GAB2, RASGRP1, and XCL1. Clustering-specific genes included JAK2, STAT3, and HIF1A while barrier-specific genes included CD80 and TNFRSF1A, suggesting the two stromal phenotypes are driven by distinct molecular mechanisms.

TL;DR: CNx identifies copy-number-driven gene expression changes underpinning stromal phenotypes, revealing suppressed immune pathways and distinct molecular networks for clustering versus barrier.
Pages 11-12
Spatial Histology as a Tool for Understanding Immune Evasion

A New Class of Prognostic Biomarker Unlike genomic markers, stromal architecture measures derive directly from standard diagnostic H&E slides that are routinely produced in clinical practice. This makes TTG-derived features potentially implementable without additional molecular testing.

Cancer-Associated Fibroblast Mechanisms The results support known CAF biology: fibroblast-derived TGF-beta inhibits CD8 T cells, dense collagen matrices physically block lymphocyte movement, and aligned ECM fibers restrict T cell entry into tumor islets. TTGs provide a way to quantify these effects at scale.

Serial Biopsy Insights In 12 patients with metastatic melanoma who underwent serial biopsies before and after treatment, stromal clustering decreased over time, suggesting dynamic remodeling of the stromal barrier during disease progression and possible response to therapy.

TL;DR: TTG-derived stromal features are clinically actionable biomarkers from standard H&E slides that quantify stromal immune exclusion, a key mechanism of immunotherapy resistance.
Page 12
Limitations and Extensions of the TTG Framework

Current Limitations The study lacked access to treatment response data, limiting direct testing of whether stromal features predict immunotherapy benefit. The H&E classification is limited to three broad cell types without specific immune subtype markers like IHC for CD8 or FoxP3.

Extensibility to Other Interactions The TTG concept can be readily applied to other cell-cell interactions beyond stromal-immune, such as tumor-lymphocyte or lymphocyte-lymphocyte networks. This opens applications across multiple cancer types where spatial immune exclusion matters.

Future Directions Validation in immunotherapy-treated cohorts, integration with multiplex immunohistochemistry for cell-type refinement, and extension of CNx to integrate additional omics layers (methylation, proteomics) could further improve the predictive power of spatial histology approaches for treatment selection.

TL;DR: While limited by available data and marker resolution, the TTG framework is extensible to other cell interactions and cancer types, with validation in immunotherapy cohorts as the key next step.
Citation: Open Access, 2020. Available at: PMC7985597.