Part V · Immunology and the tumor microenvironment · Chapter 24

Spatial organization and what it adds

Each platform answers one question. A chapter that lists instruments without naming the questions has failed.

1 · Inflamed, excluded, and desert phenotypes

The three-way scheme comes from checkpoint biology, and it asks one question. It asks where the T cells sit relative to the tumour cells¹. An inflamed tumour has cytotoxic T cells inside the tumour nests. An excluded tumour has them present but confined to stroma. A desert tumour has few anywhere.

The distinction is not cosmetic, because it separates two treatment failures with different causes. In an excluded tumour the effector cells arrived and cannot reach the target. In a desert tumour they were never recruited. Blocking a checkpoint addresses neither directly.

Breast data support the split. In triple-negative disease, immune cell localisation combined with laser-capture expression profiling of the localised regions separated three groups. An immunoreactive group carried granzyme B positive CD8 T cells inside the tumour and did well. An immune-cold group showed absent tumoural CD8 T cells, elevated B7-H4 and fibrotic stromal signatures. A third group showed stromal restriction of CD8 T cells with stromal PD-L1, and its outcomes were poor². In a series of 681 triple-negative tumours, excluded and ignored phenotypes were associated with resistance to anti-PD1 therapy, and the inflamed phenotype with response³.

The luminal deficit sits at the desert end of this axis, with the mechanism developed in Immune contexture by subtype and the luminal deficit and Hormone receptor driven immune evasion and antigen presentation, and fibroblast-mediated exclusion in Fibroblast-mediated immune exclusion.

2 · Multiplex imaging and spatial transcriptomics platforms

Platforms divide into two families, and the division is by question rather than by vendor.

Multiplexed imaging measures a chosen panel of proteins in situ at single-cell or subcellular resolution. Multiplexed ion beam imaging quantified 36 proteins across 41 triple-negative tumours⁴. Imaging mass cytometry quantified 35 biomarkers in 720 images from 352 patients⁵, and 37 proteins across 483 tumours from the METABRIC cohort⁶. Antibody cycling methods extend the same logic to tissue microarrays⁷. The question these answer is which cell is next to which, for markers chosen in advance. The panel is therefore the hypothesis, because nothing outside it is measured, and a negative result about an unpanelled protein is not a result at all.

Sequencing-based spatial transcriptomics measures expression without choosing genes first. The original array method placed tissue sections on positionally barcoded reverse transcription primers and was demonstrated in human breast cancer⁸. The question it answers is which transcriptional programmes vary across a tissue. Its limit is the spatial unit, because at the resolutions in common use a barcoded spot contains several cells. A spot measurement is therefore a mixture, and assigning transcripts to cell types requires deconvolution against a single-cell reference⁹.

The practical rule follows from the two questions. If you need to know which cell expresses something, use an imaging or single-cell resolved method. If you need to know what programme is active in a region, use transcriptomics. What each platform can and cannot resolve for heterogeneity specifically is set out in Single-cell and spatial platforms and what each can resolve.

3 · Single-cell atlases, SCSubtype, and breast cancer ecotypes

The reference atlas for this field combined single-cell and spatially resolved transcriptomics of human breast cancers¹⁰. Three outputs from it are used repeatedly.

The first is SCSubtype, a method that assigns an intrinsic subtype call to individual neoplastic cells. The result matters more than the method. A single tumour contains cells carrying different intrinsic subtype calls. Intrinsic subtype was defined on bulk expression, so a PAM50 call is a population average over a mixed population. That reframes the classification argument in Single-cell and spatial data as a challenge to bulk-defined subtypes and Clinical and molecular classification.

The second is a high-resolution immune and mesenchymal census, including macrophage populations associated with outcome.

The third is the ecotype framework. Single-cell signatures were used to deconvolute large bulk cohorts, which were then stratified into nine ecotypes with distinct cellular compositions and outcomes¹⁰. A parallel pan-carcinoma framework identified 69 transcriptionally defined cell states across 12 major lineages in 16 carcinoma types, and ten multicellular communities¹¹. The striking finding there is conservation, since most of those states were not specific to a tumour type.

4 · Spatial mapping of subclones onto ecological niches

This section asks whether genotype tracks with location, and whether location tracks with microenvironment. Both turn out to be true, which has consequences for how a subclone should be thought about.

An in situ single-cell method combining allele-specific amplification with fluorescence in situ hybridisation measured PIK3CA mutation and HER2 amplification in the same archived sections. The two alterations were not always present in the same cells. Chemotherapy selected for PIK3CA-mutant cells, which were a minor population in nearly all treatment-naive samples. Treatment-associated change in the spatial distribution of genetic diversity correlated with poor outcome after adjuvant trastuzumab¹².

Topographic single-cell sequencing profiled copy number in 1293 single cells from 10 patients with synchronous in situ and invasive disease, preserving each cell's position. Most mutations and copy number aberrations had evolved within the ducts before invasion, which supports a multiclonal invasion model¹³.

Whole-genome sequencing combined with base-specific in situ sequencing produced quantitative subclone maps across eight whole sections from two multifocal primary breast cancers. Subclone territories carried distinct transcriptional and histological features, and distinct cellular microenvironments¹⁴.

The conclusion is not that clones have addresses. It is that a clone's measured behaviour is partly a property of where it sits. A clonal fraction reported without position has averaged over that. The temporal counterpart is developed in Temporal heterogeneity and clonal evolution, the sampling counterpart in Spatial heterogeneity.

5 · Predicting spatial expression from histology

A deep learning model trained on 30,612 spatially resolved expression measurements matched to haematoxylin and eosin images from 23 patients with breast cancer predicted local expression of over 100 genes at 100 micrometre resolution. It generalised to The Cancer Genome Atlas without retraining¹⁵.

What this establishes is narrow and real. Morphology carries transcriptional information, which pathology has always assumed and could not quantify. A stained slide already exists for every patient, so a model of this kind can be applied at a scale no spatial assay can reach.

What it does not establish needs saying with equal clarity. Performance is a per-gene quantity. Only a subset of genes is predictable, and the predictable set is enriched for genes with large, spatially structured expression differences. A single mean correlation across all genes hides exactly the distinction that matters. A prediction is also not a measurement, and the error on an individual patient's predicted value is not the error reported on a cohort.

6 · Cellular neighborhoods, distance metrics, and niche definition

A cellular neighbourhood is a recurrent local composition. The computation is worth stating, because every step in it is a choice. Take each cell, list its nearest neighbours, describe that list as a vector of cell type frequencies, then cluster cells by that vector. Nine such neighbourhoods were derived in the original application, from 140 tissue regions across 35 patients with advanced colorectal cancer⁷.

The choices are the number of neighbours or the radius, whether distance is measured centre to centre or membrane to membrane, how segmentation assigned pixels to cells, and how many clusters were requested. Not one of those settings is biology, although every one of them changes the output.

The breast results are substantial nonetheless. Spatial enrichment analysis separated immune mixed from immune compartmentalised tumours, and ordered structures along the tumour-immune border were associated with compartmentalisation and with survival⁴. Systematic mapping of 693 breast tumours identified ten recurrent microenvironment structures that varied by vascular content, stromal quiescence or activation, and leukocyte composition. They were enriched differently across subtypes, and a structure containing co-occurring regulatory and dysfunctional T cells predicted poor outcome in oestrogen receptor positive disease¹⁶. Neighbourhood features were also linked to prognosis in the METABRIC imaging cohort⁶.

The measurement point is the one the field most often skips. Structures named by different groups, from different panels, with different clustering, are not the same objects. Calling two of them by the same name does not make them comparable. There is no agreed nomenclature here, and the fibroblast taxonomy in Cellular architecture of the tumor microenvironment shows that even a named taxonomy remains contested.

7 · The invasive front and intratumoral compartmentalization

The front is where the interesting questions concentrate, because three things change across it at once. Myoepithelial integrity, stromal composition and immune access.

A 37-plex imaging study of 79 surgical resections compared normal breast with matched in situ and invasive disease from the same patients. It resolved four microenvironment states defined by the location and function of myoepithelium, fibroblasts and immune cells¹⁷.

Compartmentalisation is the other axis, separating tumours that hold immune and tumour cells in distinct territories from tumours that interleave them. Ordered immune structures at the border tracked with compartmentalisation and with survival in triple-negative disease⁴. In colorectal cancer, coupling of tumour and immune neighbourhoods, fragmentation of T cell neighbourhoods, and disrupted communication between neighbourhoods were associated with worse outcomes at the invasive front⁷. That last result is from another disease and should be carried across as a hypothesis.

Compartmentalisation has a plausible mechanism, and it is also a scoring output whose value depends on segmentation quality and on the enrichment statistic chosen. Both readings are needed. Consequences for dissemination are developed in Metastatic dissemination and organ tropism, and for delivery and interstitial pressure in Vascular permeability, interstitial pressure, and drug delivery.

8 · What spatial data adds over bulk profiling, and at what cost

Three things are genuinely added, and they are narrower than the field's rhetoric.

Bulk profiling cannot separate an excluded tumour from an inflamed one carrying the same total immune content. Deconvolution of bulk expression recovers proportions, not positions. That distinction has been associated with response to anti-PD1 therapy in triple-negative disease^2,3.

Bulk profiling averages subtype. Individual neoplastic cells within one tumour carry different intrinsic subtype calls, which no bulk classifier can report¹⁰.

Bulk clonal inference gives fractions without coordinates. Spatial genomics shows that subclones occupy territories with their own microenvironments and their own transcriptional character¹⁴.

Against that, the honest accounting of cost has five entries.

The first is sampling, and it is the largest. A spatial platform is applied to a section, usually one. Resolution within a plane does not address the fact that a tumour is three-dimensional and a section is not. The problem set out in Spatial heterogeneity is unchanged by better microscopy.

The second is panel dependence for imaging, and reference dependence for deconvolution. Both fix what can be found before the tissue is looked at.

The third is analytic freedom. Segmentation, neighbourhood radius, cluster number and normalisation each change the output, and none of them is standardised across laboratories.

The fourth is tissue and money, since these assays consume material that may be needed for a clinical test, at a cost per case not comparable to immunohistochemistry.

The fifth is reproducibility, because no spatial readout in breast cancer has been shown reproducible across centres at the standard demanded of a clinical assay. That standard is set out in Receptor assessment and the measurement problem and applied to trial enrolment criteria in Clinical trials and evidence interpretation.

The summary should be stated plainly. Spatial data has changed what is believed about how breast cancers are organised. It has not yet changed what is prescribed. No spatial measurement currently selects therapy in this disease. The prognostic associations are real, and most have not been shown to add to established variables in prospective testing. Treating the first claim as if it implied the second is the commonest error in this literature.

References

1. Chen DS, Mellman I. Elements of cancer immunity and the cancer-immune set point. Nature 2017 541:321-330. PMID 28102259
2. Gruosso T, Gigoux M, Manem VSK, et al. Spatially distinct tumor immune microenvironments stratify triple-negative breast cancers. J Clin Invest 2019 129:1785-1800. PMID 30753167
3. Hammerl D, Martens JWM, Timmermans M, et al. Spatial immunophenotypes predict response to anti-PD1 treatment and capture distinct paths of T cell evasion in triple negative breast cancer. Nat Commun 2021 12:5668. PMID 34580291
4. Keren L, Bosse M, Marquez D, et al. A structured tumor-immune microenvironment in triple negative breast cancer revealed by multiplexed ion beam imaging. Cell 2018 174:1373-1387. PMID 30193111
5. Jackson HW, Fischer JR, Zanotelli VRT, et al. The single-cell pathology landscape of breast cancer. Nature 2020 578:615-620. PMID 31959985
6. Ali HR, Jackson HW, Zanotelli VRT, et al. Imaging mass cytometry and multiplatform genomics define the phenogenomic landscape of breast cancer. Nat Cancer 2020 1:163-175. PMID 35122013
7. Schurch CM, Bhate SS, Barlow GL, et al. Coordinated cellular neighborhoods orchestrate antitumoral immunity at the colorectal cancer invasive front. Cell 2020 182:1341-1359. PMID 32763154
8. Stahl PL, Salmen F, Vickovic S, et al. Visualization and analysis of gene expression in tissue sections by spatial transcriptomics. Science 2016 353:78-82. PMID 27365449
9. Andersson A, Larsson L, Stenbeck L, et al. Spatial deconvolution of HER2-positive breast cancer delineates tumor-associated cell type interactions. Nat Commun 2021 12:6012. PMID 34650042
10. Wu SZ, Al-Eryani G, Roden DL, et al. A single-cell and spatially resolved atlas of human breast cancers. Nat Genet 2021 53:1334-1347. PMID 34493872
11. Luca BA, Steen CB, Matusiak M, et al. Atlas of clinically distinct cell states and ecosystems across human solid tumors. Cell 2021 184:5482-5496. PMID 34597583
12. Janiszewska M, Liu L, Almendro V, et al. In situ single-cell analysis identifies heterogeneity for PIK3CA mutation and HER2 amplification in HER2-positive breast cancer. Nat Genet 2015 47:1212-1219. PMID 26301495
13. Casasent AK, Schalck A, Gao R, et al. Multiclonal invasion in breast tumors identified by topographic single cell sequencing. Cell 2018 172:205-217. PMID 29307488
14. Lomakin A, Svedlund J, Strell C, et al. Spatial genomics maps the structure, nature and evolution of cancer clones. Nature 2022 611:594-602. PMID 36352222
15. He B, Bergenstrahle L, Stenbeck L, et al. Integrating spatial gene expression and breast tumour morphology via deep learning. Nat Biomed Eng 2020 4:827-834. PMID 32572199
16. Danenberg E, Bardwell H, Zanotelli VRT, et al. Breast tumor microenvironment structures are associated with genomic features and clinical outcome. Nat Genet 2022 54:660-669. PMID 35437329
17. Risom T, Glass DR, Averbukh I, et al. Transition to invasive breast cancer is associated with progressive changes in the structure and composition of tumor stroma. Cell 2022 185:299-310. PMID 35063072