Part VI · Heterogeneity, evolution, and metastatic biology · Chapter 34

Measuring heterogeneity in practice

Every platform reports a different kind of heterogeneity. Only a few of them change a decision.

1 · Multi-region and multi-site sampling, yield against feasibility

A single core answers a question about one place. Multi-region sampling answers a question about a tumour. The difference is not a refinement. It is a change in what the measurement is about.

The yield is quantifiable. Whole-genome and targeted sequencing of multiple samples from each of 50 primary breast cancers, 303 samples in total, found that potentially targetable mutations were subclonal in 13 of those 50 cancers¹. A single sample from one of those tumours can report a subclonal alteration, and nothing about the report says the alteration is subclonal. That is the argument for multi-region work stated at its strongest.

Multi-region and multi-site sampling are different designs answering different questions. Multi-region sampling of one mass estimates variation within a lesion. Sampling several lesions estimates variation between deposits, which is the problem set out in Receptor discordance and conversion. A study that does one and reports the other has overreached.

Feasibility is the constraint, and it is not only cost. Every additional region consumes tissue, pathologist time and turnaround. The number of regions sampled sets the resolution of the answer, and it is almost never stated in a report. A claim about intratumoral heterogeneity derived from one block is a claim about one block.

Proteomic measurement scales differently from genomic measurement. Mass-spectrometry proteomics across 280 tumour regions found that proteomic heterogeneity increased with tumour progression, independently of genomic heterogeneity, and tracked with differences in the microenvironment². The axes described in A framework for heterogeneity are therefore separable in practice and not only in principle.

2 · Single-cell and spatial platforms and what each can resolve

The two platform families trade the same two properties against each other. Single-cell methods resolve individual cells and discard where those cells were. Spatial methods preserve position and give up either per-cell resolution or transcriptome breadth, depending on whether the platform is spot-based or imaging-based.

Single-nucleus sequencing established the case. Sequencing 100 single cells from one polygenomic breast tumour resolved three distinct clonal subpopulations, a structure that bulk sequencing reports as a single averaged genome³. The same work argued for punctuated clonal expansion rather than gradual accumulation, which is a claim about evolutionary mode that only per-cell data can support.

Spatially resolved transcriptomics adds the coordinate. A single-cell and spatial atlas of human breast cancers defined recurrent neoplastic cell states and resolved stromal and immune populations. It also showed that stromal and immune niches are spatially organised rather than uniformly distributed⁴. The platform comparison itself belongs to Spatial organization and what it adds and is not repeated here.

Two limitations deserve to be stated as measurement facts rather than caveats. Dissociation selects cells, so a single-cell census counts what survived the protocol. And every spatial platform has a resolution floor, below which two adjacent cells are reported as one measurement. A heterogeneity estimate is bounded by both.

3 · Circulating tumor DNA as a heterogeneity integrator and its blind spots

Plasma is the only sample that draws from every deposit at once. That is the whole case for it, and it is a strong one. It makes circulating tumour DNA good at detecting convergence and poor at detecting divergence, because a variant found in plasma carries no lesion of origin.

The integrating property has been shown to matter. In 55 patients with early breast cancer receiving neoadjuvant chemotherapy, sequencing of molecular residual disease predicted the genetic events of the subsequent metastatic relapse more accurately than sequencing of the primary tumour did. The median lead time over clinical relapse was 7.9 months⁵. The primary is a record of what was resected. The plasma is a record of what survived.

Four blind spots are worth naming separately.

Shedding is not uniform. Low-volume disease, and disease behind the blood-brain barrier, contribute less to the pool than their size implies. Cerebrospinal fluid analysis in leptomeningeal disease covers the cerebrospinal fluid alternative.

Non-tumour clones contaminate the signal. Plasma sequencing with matched white blood cell DNA was performed in 47 controls without cancer and in 124 patients with metastatic cancer. Of the cell-free DNA mutations found, 81.6% in the controls and 53.2% in the patients had features consistent with clonal haematopoiesis⁶. Without matched white cell sequencing, a proportion of what is called tumour is not. Clonal hematopoiesis and other false positives develops this.

Variant allele fraction is not clonal fraction. It is confounded by total shedding, by copy number at the locus and by the fraction of the pool that is tumour-derived. Reading it as the proportion of cancer cells carrying the variant is the commonest error in this area.

Plasma reports DNA. A receptor conversion that is transcriptional or post-translational is invisible to it, which is why Receptor discordance and conversion still requires tissue.

4 · Imaging-based heterogeneity, radiomics, PET metrics, and HER2 PET

Imaging has one property no tissue method has. It samples every lesion, non-invasively, and it can be repeated. Where heterogeneity between deposits is the question, that advantage is decisive.

Target imaging is the clearest worked example. In 56 patients with HER2-positive metastatic breast cancer scheduled for T-DM1, pretreatment zirconium-89 trastuzumab positron emission tomography was negative in 29%. A mixed pattern within a single patient was found in 46%⁷. Combining that scan with an early metabolic response assessment separated two groups. Median time to treatment failure was 2.8 months in 12 patients and 15 months in 25. Nearly half of a population defined as HER2-positive by tissue assay carried tumour load that did not bind the antibody.

Receptor imaging extends the same logic to the oestrogen receptor, where fluoroestradiol uptake has been examined as a predictor of benefit from endocrine therapy⁸. The shared principle is that a tissue assay reports one lesion and a scan reports all of them.

Radiomics measures something different again. Texture features quantify variation in voxel intensity, which is a property of the reconstructed image. That variation depends on scanner, acquisition protocol, reconstruction kernel, voxel size and the segmentation that defined the region. Cross-site reproducibility is the limiting problem, and it is a property of the pipeline rather than of the tumour.

5 · Computational deconvolution and clonal inference from bulk data

Bulk sequencing measures a mixture. Deconvolution infers the components of that mixture from the mixture itself, which is an inverse problem and is solved under assumptions.

The standard approach clusters somatic mutations by their estimated cellular prevalence, correcting for copy number change at each locus and for contamination by normal cells, and returns putative clonal clusters⁹. Single-cell sequencing has been used to check that the inferred clusters correspond to real populations. The same logic applied to expression data uses single-cell reference signatures to estimate cell type composition in bulk cohorts, which is how large cohorts were stratified into recurrent cellular ecotypes⁴.

What the inference cannot do is worth stating explicitly. It cannot recover a population that was not in the sample. It cannot separate two populations that happen to sit at the same cellular prevalence. It degrades when copy number is inaccurate, and breast cancer genomes are copy number driven. And it assumes that each mutation arose once, which structural rearrangement and extrachromosomal amplification violate. The mechanism is set out in Amplicon topology, extrachromosomal DNA, and the mechanistic origin of instability.

6 · What is actionable today and what remains a research instrument

The criterion is narrow and it is worth applying honestly. A measurement is actionable when a different result would produce a different action that is available now. Everything else is description, however good the biology behind it.

Multi-region sequencing, single-cell and spatial profiling, radiomic texture analysis, clonal deconvolution and target-specific positron emission tomography are all research instruments in 2026. Each measures something real. None of them has a prospective trial showing that acting on the result improves an outcome.

Reproducibility is the gate that most of them have not yet passed. When 18 pathologists from 15 institutions scored 170 breast biopsies, agreement on cases scored immunohistochemistry 0 was 25%¹³. A research assay with worse observer agreement than that cannot carry a treatment decision, and the burden of showing otherwise sits with whoever proposes it.

References

1. Yates LR, Gerstung M, Knappskog S, et al. Subclonal diversification of primary breast cancer revealed by multiregion sequencing. Nat Med 2015 21:751-759. PMID 26099045
2. Mardamshina M, Karagach S, Mohan V, et al. Integrated spatial proteomic analysis of breast cancer heterogeneity unravels cancer cell phenotypic plasticity. Nat Commun 2025 16:10482. PMID 41290667
3. Navin N, Kendall J, Troge J, et al. Tumour evolution inferred by single-cell sequencing. Nature 2011 472:90-94. PMID 21399628
4. 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
5. Garcia-Murillas I, Schiavon G, Weigelt B, et al. Mutation tracking in circulating tumor DNA predicts relapse in early breast cancer. Sci Transl Med 2015 7:302ra133. PMID 26311728
6. Razavi P, Li BT, Brown DN, et al. High-intensity sequencing reveals the sources of plasma circulating cell-free DNA variants. Nat Med 2019 25:1928-1937. PMID 31768066
7. Gebhart G, Lamberts LE, Wimana Z, et al. Molecular imaging as a tool to investigate heterogeneity of advanced HER2-positive breast cancer and to predict patient outcome under trastuzumab emtansine (T-DM1): the ZEPHIR trial. Ann Oncol 2016 27:619-624. PMID 26598545
8. Parihar AS, et al. 18 F-Fluoroestradiol PET/CT for predicting benefit from endocrine therapy in patients with ER-positive breast cancer. J Nucl Med 2025. PMID 40081952
9. Roth A, Khattra J, Yap D, et al. PyClone: statistical inference of clonal population structure in cancer. Nat Methods 2014 11:396-398. PMID 24633410
10. Schiavon G, Hrebien S, Garcia-Murillas I, et al. Analysis of ESR1 mutation in circulating tumor DNA demonstrates evolution during therapy for metastatic breast cancer. Sci Transl Med 2015 7:313ra182. PMID 26560360
11. Bidard FC, Kaklamani VG, Neven P, et al. Elacestrant (oral selective estrogen receptor degrader) versus standard endocrine therapy for estrogen receptor-positive, HER2-negative advanced breast cancer: results from the randomized phase III EMERALD trial. J Clin Oncol 2022 40:3246-3256. PMID 35584336
12. Turner NC, Mayer EL, Park YH, et al. Switching to camizestrant at ESR1 mutation emergence before disease progression during first-line treatment of hormone receptor-positive advanced breast cancer (SERENA-6). Lancet Oncol 2026. PMID 42442380
13. Robbins CJ, Fernandez AI, Han G, et al. Multi-institutional assessment of pathologist scoring HER2 immunohistochemistry. Mod Pathol 2023 36:100032. PMID 36788069