Breast cancer · spatial transcriptomics · histology · study 3 of a series

Where are the candidates expressed in tissue, and can histology see them?

Study 2 nominated breast-cancer drug-target candidates from tumour omics and found that subtype over-expression is a weak guide to dependency. Bulk tumours mix cancer cells with stroma and immune infiltrate, so a gene can look "up" because the tissue's composition differs, not its cancer cells. This study takes the candidates to intact tissue: which compartment each sits in, whether that explains which nominations replicated and which genes cells depend on, and whether a pathology foundation model can see any of it in the H&E image.

pre-registered 2026-10-07no data downloaded yetresults pending

Design (pre-registered) · Literature · Repository · Study 2: does the protein layer pick better targets? · Study 1: cell states and survival

The questions

Compartment. For each candidate, is its expression in tumour cells, stroma or immune infiltrate, and does nomination enrich for tumour-cell genes beyond random genes of the same abundance?
Compartment and the study-2 outcomes. Does a candidate's tumour-cell compartment predict whether its nomination replicated between cohorts, or how strongly subtype-matched cell lines depend on it?
Histology. Can a frozen pathology foundation model predict the candidates' local expression from the H&E image alone, and is what it sees simply the compartment structure?

Expected answers, stated before any data: nomination does enrich for tumour-cell genes, most of all in the DNA-informed arm; compartment explains part of replication and little of dependency; histology predicts the candidates poorly except where they are compartment markers, and a pathology encoder adds a little over generic image features and nothing over knowing the compartment.

Why this study

Open question

No published study scores a candidate's tumour-cell compartment in spatial data and tests whether it predicts replication between bulk cohorts or CRISPR dependency. Compartment methods exist and are benchmarked; nobody has pointed them at nominated targets.

Histology and targets

Predicting spatial expression from H&E is an active benchmark field (HEST-1k reports 0.47–0.60 mean correlation on breast with frozen encoders, with an ImageNet baseline close behind). None of it asks about nominated drug targets specifically, and the field's own 2026 critique shows aggregate correlations are inflated by between-slide means.

What it adds to the series

Study 2's candidates get a tissue-level rung they could not have from bulk data, and the series gains two data types (spatial transcriptomics, histology images) under the same discipline: pre-registration, nulls, held-out sections.

Design

Data

Role	Dataset	Sections	Subtypes	Compartment labels
Discovery	Wu et al. 2021, Visium	6	TNBC, ER+, HER2+	pathologist annotation per spot
Replication	Li et al. 2025, Visium	23	TNBC, HER2+	tumour / immune / stroma cell counts per spot
Per-cell validation	Janesick et al. 2023, Xenium	2 blocks	HER2+	authors' cell types; 313-gene panel
Candidates	study 2	150 genes	50 per subtype	with replication status and dependency

Compartment score

Per section, spots are tumour, stroma, immune or other from the shipped labels. A gene's tumour log-fold-change is its mean log expression in tumour spots minus the mean in all other spots; its score is the median across the sections of its subtype. Two nulls: spot labels permuted within each section, and 1,000 random gene sets matched on abundance and detection rate.

Histology

One 224-px tile at 0.5 µm/px per spot; CLS features from a frozen Phikon encoder and from an ImageNet ViT-B; PCA to 256 and ridge per gene, trained on all sections but one and scored on the held-out one. Correlations are reported raw and after centring per section, against a within-section feature permutation, and beside a model that knows only the spot's compartment.

The ladder

Rung	What it establishes
Matched-random floor	what any genes of that abundance score
Label-permutation null	the compartment score of a gene with no spatial structure
Section hold-out	what transfers to tissue the model never saw
Within-section feature permutation	the histology correlation with no image information
ImageNet ViT-B; compartment-composition model	whether a pathology encoder adds anything to generic features and to knowing the compartment
Control genes	EPCAM, KRT8, KRT18 (tumour); PTPRC (immune); COL1A1, DCN (stroma); ACTB, GAPDH (housekeeping)

Predictions, fixed before any data

	Prediction	Note	Outcome
P1	Control genes land in their compartments in every section	pipeline check before anything else	pending
P2	The candidates' mean tumour log-fold-change beats the matched-random floor in each subtype; at least half are called tumour		pending
P3	Candidates from the DNA-informed arm have the highest median tumour score, by at least 0.3 over RNA alone		pending
P4	Replicated candidates score higher than non-replicated ones by at least 0.3; odds ratio per SD at least 1.5		pending
P5	Tumour score and subtype-matched dependency correlate weakly, \|ρ\| < 0.2		pending
P6	Histology predicts the candidates poorly: median held-out correlation < 0.3 raw, < 0.15 section-centred; control compartment genes > 0.5		pending
P7	Phikon beats the ImageNet ViT-B by at least 0.05, and the composition baseline by less than 0.05		pending
P8	A candidate's histology correlation is explained by its compartment score: Spearman ≥ 0.5		pending

Committed in DESIGN.md §8 on 2026-10-07. Any later change to the design is logged in its §11.

Results

pending Compartment calls for the 150 candidates, per subtype, against both nulls (P1–P3) — fills from results/ once stage compartment has run.

pending Compartment score against replication status and dependency (P4, P5) — fills from results/ once stage compartment has run.

pending Held-out histology correlations: candidates, controls, both encoders, the composition baseline (P6–P8) — fills from results/ once stage histology has run.

Data and licences

Dataset	Route	Licence
Wu et al. 2021 Visium, 6 sections, with pathologist annotation	Zenodo 4739739	CC BY 4.0
Li et al. 2025 Visium, 23 sections, with tissue annotation	Zenodo 15211538	CC BY 4.0
Janesick et al. 2023 Xenium + Visium CytAssist	Zenodo 10076046	CC BY 4.0
Study 2 candidate tables (pinned commit)	brca-target-evidence	MIT
Phikon encoder weights	Hugging Face hub, not redistributed	non-commercial
ImageNet ViT-B/16 (timm)	Hugging Face hub	Apache-2.0

Every file is recorded with URL and sha256 in results/MANIFEST.sha256 when stage data runs. Model weights are fetched by the user's own hub token and never stored in the repository.

Literature

The survey behind the design, with a strength tag on every entry, is in research/literature.md.

Wu SZ et al.. A single-cell and spatially resolved atlas of human breast cancers. Nat Genet 2021. doi
Li et al.. Spatial transcriptomics of breast cancer with computational tissue annotation. npj Precis Oncol 2025. doi
Janesick A et al.. High resolution mapping of the tumor microenvironment using integrated single-cell, spatial and in situ analysis. Nat Commun 2023. doi
He B et al.. Integrating spatial gene expression and breast tumour morphology via deep learning. Nat Biomed Eng 2020. doi
Jaume G et al.. HEST-1k: a dataset for spatial transcriptomics and histology image analysis. NeurIPS 2024.
Nguyen et al.. Is H&E image-to-spatial transcriptomics simpler than it looks?. arXiv 2026.
Cable DM et al.. Robust decomposition of cell type mixtures in spatial transcriptomics. Nat Biotechnol 2022. doi
Filiot A et al.. Scaling self-supervised learning for histopathology with masked image modeling (Phikon). medRxiv 2023. doi

Limits, stated in advance

Few sections. Six discovery sections across three subtypes; the replication cohort has two subtypes, so the luminal arm rests on the ER+ sections of Wu et al. alone.
Spots are mixtures. A 55 µm spot can hold tumour and stroma; the per-cell Xenium check covers only candidates on a 313-gene panel.
Modest correlations by construction. A frozen encoder with a few thousand spots per section; what matters is the comparison with the baselines and nulls, not the absolute value.
Licences. Phikon is non-commercial; the repository depends on it through the hub and never redistributes weights.