HistoAtlas: A Pan-Cancer Morphology Atlas Linking Histomics to Molecular Programs and Clinical Outcomes

1. Problem Statement

Despite the ubiquity of Hematoxylin and Eosin (H&E) stained slides as the gold standard for cancer diagnosis, existing pan-cancer resources (e.g., TCGA, cBioPortal) largely lack quantitative morphological data. Current approaches suffer from three main limitations:

• Data Collapse: Morphological information is often reduced to categorical grades or discarded entirely.
• Lack of Interpretability: While deep learning foundation models (e.g., UNI, Virchow) can predict outcomes from whole-slide images (WSIs), their latent embeddings are "black boxes" that do not decompose into interpretable biological features (e.g., cell densities, spatial distances).
• Absence of Systematic Linkage: There is no comprehensive resource that systematically links quantitative histomic features to survival, gene expression, somatic mutations, and immune subtypes across multiple cancer types with rigorous statistical correction.

2. Methodology

The authors constructed HistoAtlas, a pan-cancer computational atlas based on 6,745 diagnostic H&E slides from 21 TCGA solid-tumor cancer types.

A. Feature Extraction Pipeline

The pipeline converts raw WSIs into 38 quantitative, interpretable histomic features through two automated segmentation stages:

1. Tissue Segmentation: A UNet-based model (inspired by CellViT with a Phikon ViT-B backbone) classifies tissue into five compartments: Tumor, Stroma, Necrosis, Blood, and Normal Epithelium.
2. Cell Detection & Classification: The HistoPLUS model detects and classifies over 4.4 billion individual cells into nine types: tumor cells, lymphocytes, fibroblasts, neutrophils, eosinophils, plasmocytes, apoptotic bodies, mitotic figures, and red blood cells.

From these segmentations, 38 histomic features are derived across five categories:

• Tissue Composition: Area fractions of compartments.
• Cell Densities: Counts per unit area in specific regions (e.g., intratumoral vs. stromal).
• Nuclear Morphology & Kinetics: Metrics like nuclear area, pleomorphism, eccentricity, mitotic index, and apoptotic index.
• Spatial Organization: Nearest-neighbor distances, interface densities, and infiltration ratios (e.g., tumor-stroma contact).
• Spatial Heterogeneity: Coefficients of variation for densities across tiles.

B. Statistical Framework

• Associations: The authors tested associations between the 38 features and:
  • Survival: Overall, disease-specific, disease-free, and progression-free survival (Cox proportional hazards models).
  • Molecular Programs: 293 targets including 133 curated genes (mRNA and CNV), 21 Hallmark pathway scores, and 6 immune cell fraction scores.
  • Immune Subtypes: Thorsson's C1–C6 immune classifications.
• Correction & Validation:
  • Covariate Adjustment: Models were adjusted for age, sex, stage, and stratified by Tissue Source Site (TSS).
  • Multiple Testing: Benjamini–Hochberg (BH) correction was applied within biologically defined families.
  • Evidence Grading: Results were assigned "evidence-strength badges" (Strong, Moderate, Suggestive, Insufficient) based on P-values, effect sizes, confidence intervals, and sample size.
  • Power Analysis: Minimum Detectable Effect Size (MDES) was calculated to define the limits of detection for each cancer type.

C. Clustering

Unsupervised K-means clustering was performed on the 38-dimensional feature vectors to identify 10 pan-cancer morphological clusters (L1) and 69 cancer-specific subclusters (L2), without using any molecular input.

3. Key Contributions

1. First Pan-Cancer Quantitative Morphology Atlas: A systematic resource linking 38 interpretable histomic features to molecular and clinical outcomes across 21 cancer types.
2. Spatially Traceable Results: Unlike bulk scoring, HistoAtlas allows users to trace statistical associations back to specific tissue compartments (e.g., intratumoral vs. stromal) and individual cells via an interactive web interface.
3. Interpretability over "Black Box" AI: The atlas uses explicit, human-interpretable features rather than deep learning embeddings, bridging the gap between morphology and biology.
4. Rigorous Statistical Calibration: The inclusion of MDES, evidence badges, and explicit handling of batch effects (TSS) ensures transparency regarding what the atlas can and cannot detect.

4. Key Results

A. Recovery of Canonical Biology

The morphological clusters aligned with known molecular programs without prior molecular input:

• Cluster 4 (Thymoma-enriched): Showed strong immune rejection pathway enrichment.
• Cluster 6 (Colorectal-enriched): Dominated by Wnt/β-catenin signaling and C1 wound-healing immune subtypes.
• Cluster 8 (Breast/Prostate-enriched): Characterized by estrogen response upregulation and proliferation suppression.
• Correlation with Molecules: Morphometric features served as proxies for molecular programs (e.g., mitotic index correlated with PLK1; invasion depth correlated with EMT markers like ZEB1).

B. Compartment-Specific Prognostic Signals

A critical finding was the differential prognostic value of immune location:

• Intratumoral Lymphocytes: Showed a strong protective association with survival (Pan-cancer HR = 0.87).
• Stromal Lymphocytes: Showed a significantly weaker protective effect (HR = 0.89).
• Implication: Bulk TIL scoring (which combines both compartments) dilutes the strong prognostic signal of intratumoral infiltration. In Breast Cancer (BRCA), intratumoral density was highly protective (HR = 0.72), while stromal density was non-significant.

C. Novel Morphological Subtypes

The atlas identified morphologically distinct clusters with divergent outcomes that transcriptomic subtyping missed:

• Cluster 2 (Quiescent-Cold): Characterized by suppressed proliferation and lymphocyte density, yet associated with favorable survival (HR = 0.54).
• Cluster 8 (Hormone-Driven): Associated with adverse survival (HR = 1.37).
• This suggests that the binary "immune-hot/cold" classification obscures biologically relevant heterogeneity (e.g., quiescent inflammation vs. active immune engagement).

D. Spatial Immune Topology

• Immune Exclusion: Tumor-lymphocyte nearest-neighbor distance at the invasive front inversely correlated with CD8A expression, validating that spatial exclusion detected by histomics corresponds to reduced cytotoxic T-cell gene expression.
• Interface Pressure: A composite measure of immune cell engagement at the tumor-stroma boundary showed protective effects in specific cancers (e.g., HNSC).

5. Significance and Limitations

Significance:

• Biomarker Discovery: Enables systematic discovery of biomarkers from routine H&E slides without the need for specialized staining or sequencing.
• Clinical Utility: Provides a transparent, queryable tool for researchers to explore morphology-molecular links, potentially guiding treatment selection (e.g., immunotherapy response prediction).
• Open Science: All data, code, and an interactive web atlas are freely available at https://histoatlas.com.

Limitations:

• Data Scope: Limited to TCGA (retrospective, predominantly European ancestry), excluding 12 cancer types with cell morphologies outside the model's training domain (e.g., lymphoid, glial).
• Feature Reliability: Performance varies for out-of-distribution cancer types (e.g., inflated kinetic indices in PRAD/LIHC due to reduced cell detection sensitivity).
• Validation: Lacks external replication in independent cohorts (e.g., CPTAC, METABRIC) and has not been benchmarked against foundation model embeddings for predictive power.
• Single Slide: Uses one slide per case, missing intratumoral heterogeneity across multiple biopsies.

In conclusion, HistoAtlas represents a major step forward in computational pathology by transforming routine H&E slides into a high-dimensional, interpretable, and statistically rigorous atlas that bridges the gap between tissue morphology and molecular oncology.