Histoatlas Pan Cancer Eval

---
name: histoatlas-pan-cancer-eval
description: Evaluates the prognostic and molecular predictive value of 38 automated histomic features extracted from H&E whole-slide images across 21 solid-tumor cancer types. It probes whether purely morphological patterns can recover canonical biology, predict survival outcomes, and correlate with gene expression, pathway activity, and immune subtypes. Use when the user wants to benchmark on TCGA Pan-Cancer H&E Cohort, or asks about evaluating this task. Reports Cox proportional-hazards model.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2603.16587
  bibtex_key: bannier2026histoatlas
  confidence: high
---

# histoatlas-pan-cancer-eval

> HistoAtlas: A Pan-Cancer Morphology Atlas Linking Histomics to Molecular Programs and Clinical Outcomes — Bannier et al. (2026) (arXiv:2603.16587, 2026)

## What this evaluates

Evaluates the prognostic and molecular predictive value of 38 automated histomic features extracted from H&E whole-slide images across 21 solid-tumor cancer types. It probes whether purely morphological patterns can recover canonical biology, predict survival outcomes, and correlate with gene expression, pathway activity, and immune subtypes.

## Datasets

- **TCGA Pan-Cancer H&E Cohort** — total 6745; splits: full (6745); repo https://github.com/histoatlas/histoatlas

## Metrics

- `Cox proportional-hazards model` **(primary)** — range: [0, ∞)
  - Ratio of hazard rates from Cox proportional-hazards models linking histomic features to survival endpoints (overall, disease-specific, disease-free, progression-free). Adjusted for age, sex, stage, and tissue source site. Significance determined via Benjamini-Hochberg FDR < 0.05.
- `Spearman rank correlation ($\rho$)` — range: [-1, 1]
  - Non-parametric measure of monotonic association between histomic features and molecular targets (gene expression, pathway scores, immune fractions). Computed per cohort with Benjamini-Hochberg correction.
- `Cliff's delta ($\delta$)` — range: [-1, 1]
  - Effect size measure for pathway enrichment across morphological clusters. Represents the probability that a randomly selected observation from one group is greater than one from another, minus the reverse probability.

## Input / output format

**Input**: Whole-slide H&E images (6,745 slides across 21 cancer types) processed through UNet tissue segmentation and HistoPLUS cell detection to yield 38 quantitative histomic features per slide (tissue composition, cell densities, nuclear morphology/kinetics, spatial organization, spatial heterogeneity).

**Output**: Statistical association results: hazard ratios with 95% CIs and adjusted P-values for survival endpoints; Spearman $\rho$ coefficients with adjusted P-values for molecular targets; effect sizes (Cliff's $\delta$, $\eta^2$) for cluster enrichment; cluster assignments (L1/L2) with stability metrics (ARI, Jaccard).

## Scoring recipe

```python
def evaluate_histoatlas(features, clinical_data, molecular_data):
    results = []
    for feature in features:  # 38 features
        for cohort in cohorts:  # 22 cohorts
            # 1. Fit CoxPH model: hazard ~ feature + age + sex + stage + tissue_source_site
            model = fit_coxph(clinical_data[cohort], feature)
            hr, ci, p_val = model.summary()
            # 2. Check proportional hazards assumption
            if schoenfeld_test(model).p < 0.05:
                continue  # Exclude if PH violated
            # 3. Apply Benjamini-Hochberg FDR correction within cohort/endpoint family
            adj_p = benjamini_hochberg([p for p in all_p_values], family=cohort)
            if adj_p < 0.05:
                results.append({'feature': feature, 'cohort': cohort, 'HR': hr, 'adj_P': adj_p})
    return results
```

## Common pitfalls

- Ignoring the proportional hazards assumption check: associations violating PH (Schoenfeld P < 0.05) are excluded before FDR correction, invalidating their Cox P-values.
- Confusing unadjusted vs. covariate-adjusted models: many associations lose significance after adjusting for age, sex, stage, and tissue source site.
- Assuming morphological clusters are trivially driven by cancer-type identity: the paper explicitly tests this by examining mixed-type clusters and within-cancer subclusters.

## Evidence (verbatim from paper)

> We then tested each feature for associations with survival and molecular programs across all 22 cohorts. For survival, we fitted Cox proportional-hazards models for each combination of 38 features, 22 cohorts (21 cancer types plus a pan-cancer cohort), and four endpoints (overall, disease-specific, disease-free, and progression-free survival), yielding $5,623$ evaluable associations (of a theoretical maximum of $6,688$; the remainder were excluded for insufficient sample size or events) under two adjustment tiers, unadjusted and adjusted for age, sex, stage, and tissue source site (§[4.4](#S4.SS4 "4.4 Survival analysis ‣ 4 Methods")). After Benjamini–Hochberg correction within predefined correction families (§[4.9](#S4.SS9 "4.9 Multiple testing correction ‣ 4 Methods"); Supplementary Table 6), 260 associations were significant at a false discovery rate of 0.05.

## Citation

```bibtex
@misc{bannier2026histoatlas,
  title={HistoAtlas: A Pan-Cancer Morphology Atlas Linking Histomics to Molecular Programs and Clinical Outcomes},
  author={Bannier et al. (2026)},
  year={2026},
  note={arXiv:2603.16587}
}
```

- arXiv: 2603.16587