Part III · Pathology and classification · Chapter 10

Clinical and molecular classification

Every taxonomy here was built to answer one question, and most of the trouble comes from asking it another.

1 · Histology, stage, grade, and clinical phenotype

Stage answers a question about extent. Grade answers a question about appearance. Neither states which pathway a tumour depends on, and both remain useful for exactly that reason.

The Nottingham system scores three morphological features separately. Tubule formation, nuclear pleomorphism and a mitotic count performed over a defined field area each receive a score, and the three scores are summed1. Grade was assessed in 1831 patients in the founding series and separated survival by a wide margin. The design point is that the criteria were made semiquantitative in order to be reproducible. That is the same motive behind every threshold in Receptor assessment and the measurement problem.

Mitotic count is the component most sensitive to handling. It depends on the field area used, on section thickness, and on the delay before fixation Preanalytics, fixation, and the invisible sources of error.

The eighth edition of the American Joint Committee on Cancer staging manual folded grade and receptor status into a prognostic stage alongside anatomy. In a validation cohort of 3327 patients at one institution, prognostic stage moved 29.5% of patients up and 28.1% down relative to anatomic stage2. The concordance index for disease-specific survival improved from 0.737 to 0.836. A registry cohort of 54727 patients gave the same direction of result.

Note what that shows and what it does not. Adding biology to an anatomic system improves prognostic discrimination. It does not turn the combined system into a biological taxonomy. Stage remains a description of extent with prognostic modifiers attached to it.

Morphology also retains predictive content that molecular profiling does not supply, and that argument belongs to Histopathology and the spectrum of breast neoplasia.

2 · Hormone receptor and HER2 based clinical categories

The working clinical categories are hormone receptor positive and HER2 negative, HER2 positive, and triple negative. They are not a taxonomy of breast cancer. They are a map from two assays onto three treatment pathways, and they were built in that direction.

Each boundary came from a treatment question. Hormone receptor positivity sits at 1% of tumour nuclei because the guideline preferred over-inclusion to denying endocrine therapy3,4. HER2 positivity sits at immunohistochemistry 3+ or in situ hybridisation amplification because those were the criteria under which trastuzumab was shown to work5. Triple negative is defined by the failure of the other two tests, which makes it the only category named for an absence.

The categories carry real information despite their origin. Patterns of distant recurrence differ systematically between them, including the site of first metastasis and the time course6. That is clinical validity. It is not the same claim as the categories marking natural divisions in the disease.

Two consequences follow for how the categories should be read.

A patient does not move between categories because their biology changed. They move because a measurement changed, or because a threshold moved. The HER2-low category in HER2-low and HER2-ultralow, entity or assay artifact is the cleanest example, since every patient in it was previously called HER2 negative.

A category boundary imported into a trial imports its reproducibility as well. That constraint is developed in Clinical trials and evidence interpretation.

3 · Intrinsic molecular subtypes and how they were derived

The derivation matters more than the result, because the derivation explains what the subtypes can and cannot be asked to do.

Perou and colleagues profiled 65 surgical specimens from 42 individuals on complementary DNA microarrays representing 8102 genes7. Twenty tumours were sampled twice, before and after a 16-week course of doxorubicin, and two were paired with a node metastasis. The decisive methodological step used those pairs. Genes were selected for low variation between two samples of the same tumour and high variation between different tumours. That set was called intrinsic, and clustering on it produced the subtypes.

So the intrinsic list was built to be stable within a tumour. A gene that moved under treatment was excluded by construction. The subtypes are therefore a description of between-tumour variation among features that stay put. That is a narrower object than a classification of breast cancer.

Sorlie and colleagues extended the analysis and split the luminal group into subgroups with different survival, with basal-like carrying the worst outcome in a uniformly treated locally advanced cohort8.

The Cancer Genome Atlas then integrated copy number, methylation, exome sequence, messenger RNA, microRNA and protein data on primary breast cancers9. Four main classes emerged when data from five platforms were combined, and each showed significant molecular heterogeneity within it. Only three genes were mutated at above 10% incidence across all breast cancers, namely TP53, PIK3CA and GATA3. Convergence of that kind across independent measurement platforms is the strongest evidence available that the classes track something real.

Caution

Nearest-centroid assignment gives every sample a subtype. A tumour lying almost equidistant from two centroids receives one label and no warning. A published subtype is the nearest of a fixed set of prototypes, not a measurement of how near it was. That distinction is invisible in a report and it drives much of the discordance described in Discordance between clinical categories and molecular class.

4 · HER2-enriched subtype as a predictor of anti-HER2 response

This is the clearest case in the book where intrinsic subtype adds information that receptor status does not already carry, and the demonstration is worth setting out in full.

A single-group phase II trial gave 18 weeks of lapatinib and trastuzumab, without chemotherapy, to 151 women with previously untreated HER2-positive stage I to IIIA disease10. Hormone receptor positive patients also received endocrine therapy. Intrinsic subtype was called by PAM50 on a baseline biopsy. At baseline 101 tumours (67%) were HER2-enriched, 22 (15%) were luminal A, 16 (11%) were luminal B, nine (6%) were basal-like and three (2%) were normal-like.

Pathological complete response in the breast occurred in 46 of 151 patients overall. Among HER2-enriched tumours it occurred in 41 of 101, which is 41%. Among non-HER2-enriched tumours it occurred in five of 50, which is 10%. The odds ratio was 6.2, with a 95% confidence interval of 2.3 to 16.810.

What makes this informative is that the comparison is internal. Every patient in the trial was HER2 positive by the routine assay. Subtype separated them anyway, so the information was not already contained in the receptor result.

A pooled analysis across five trials points the same way and sharpens it. It analysed 305 patients with early HER2-positive disease, combining the HER2-enriched call with ERBB2 messenger RNA level. Tumours that were both HER2-enriched and ERBB2-high reached a pathological complete response rate of 44.5% after lapatinib and trastuzumab, against 11.6% for everyone else11. The adjusted odds ratio was 6.05.

Three limits belong alongside the result.

The endpoint was pathological complete response to a chemotherapy-free regimen. That is the setting most sensitive to HER2 dependence, and it is not the setting most patients are treated in.

Subtype was called on research-grade RNA in a trial laboratory. Nothing here establishes that a routine subtype call would perform the same way.

A 41% response rate still leaves most of the enriched group without a complete response, and 10% is not zero. The subtype shifts a probability. It does not identify a population that can safely be treated differently.

5 · PAM50, BluePrint, and subtype calling in practice

PAM50 is a 50-gene nearest-centroid predictor. It was developed on 189 prototype samples, using microarray and quantitative reverse transcription polymerase chain reaction data. It was then tested for prognosis in 761 patients who received no systemic therapy, and for prediction of pathological complete response in 13312.

Applied to archival material it works. In 786 ER-positive tamoxifen-treated cases with a median follow-up of 11.7 years, the assay ran on formalin-fixed paraffin-embedded blocks and gave more prognostic information than clinical assays for hormone receptors or Ki-6713. Ten percent of clinically ER-positive cases were assigned to a non-luminal subtype.

BluePrint is a separate 80-gene profile, developed on 200 specimens and confirmed in four independent validation cohorts totalling 784 patients14. It sorts tumours into basal, luminal and HER2 types, and it was also tested against pathological complete response in 133 neoadjuvant patients.

The practical problem is not any single assay. It is that different single-sample predictors disagree about the same tumour. Three published predictors were applied to one in-house cohort of 53 tumours and three public cohorts totalling 77915. Pairwise agreement across the whole classification ran from a kappa of 0.238 to 0.740. Only basal-like showed almost perfect agreement, above 0.812. The proportion of tumours called basal-like was consistent whichever predictor was used. The proportions called luminal A, luminal B, HER2 and normal-like varied substantially.

That result should govern how any published subtype is read. The predictors identified subtypes with broadly similar survival curves, and they did not assign the same patients to the same subtypes.

Caution

A subtype in a paper is a property of the tumour and the classifier together. Two papers reporting on luminal B disease may not be reporting on the same patients. Basal-like is the exception. It is the only intrinsic class that can be compared across publications without first checking which predictor was used.

6 · Integrative clusters and copy number driven taxonomy

Expression-based subtypes cluster on transcription. Integrative clusters cluster on the joint structure of copy number and expression, which is a different axis and yields a different partition.

The METABRIC analysis paired DNA and RNA profiles from 997 discovery and 995 validation primary tumours, each with long-term follow-up16. Inherited variants and acquired somatic copy number aberrations were associated with expression for roughly 40% of genes. Unsupervised analysis of the paired profiles produced subgroups that reproduced in the validation set. They included a high-risk ER-positive subgroup driven by cis-acting amplification at 11q13 and 11q14, and a favourable subgroup with almost no copy number aberration at all.

The clinically interesting result came later, from the same cohort with longer follow-up. A statistical model applied to 3240 patients, 1980 of them with molecular data, mapped where and when relapse occurs across three classification systems17. Four late-recurring integrative subtypes were identified. Together they comprise about 26% of tumours that are ER positive and HER2 negative, and each carries a mean risk of recurrence between 47% and 62% up to 20 years after diagnosis. The same analysis separated a group of triple-negative cancers in which recurrence after five years is rare from a group that remains at risk.

That is a different kind of claim from a subtype label. It is a statement about timing, and timing is what an endocrine-treated patient at year eight actually needs to know.

None of this is in routine use, and the reason is clean rather than political. No trial has randomised a management decision on an integrative cluster. Why that is the deciding question is set out in Analytical validity, clinical validity, and clinical utility defined.

7 · Transcriptomic, proteomic, and integrated classifications

Every measurement layer produces its own partition of the same tumours. The disagreements between layers carry information, and they are not noise.

Quantitative mass spectrometry was applied to 105 genomically annotated breast cancers, of which 77 gave high-quality proteomic and phosphoproteomic data18. Global protein data confirmed basal and luminal clusters. It also produced a stromal-enriched group that expression clustering had not separated in the same way. Analysis of the phosphoproteome identified a cluster driven by G-protein-coupled receptor signalling that was not readily visible at the messenger RNA level.

Two lessons come out of that, and they should be kept apart.

The first is biological. Protein abundance and phosphorylation sit downstream of transcript abundance, and the steps between are not uniform across tumours. A pathway can be active without its transcripts being elevated.

The second is methodological. A stromal-enriched proteomic cluster is partly a statement about what else was in the specimen. The integrated Cancer Genome Atlas analysis found the same thing independently, identifying two protein-expression-defined subgroups that the authors attributed to stromal and microenvironmental elements9. That is the composition problem developed in Single-cell and spatial data as a challenge to bulk-defined subtypes.

Convergence across platforms is the strongest available argument that a class is not an artefact of one measurement. Divergence between platforms is not evidence that one of them is wrong.

8 · Discordance between clinical categories and molecular class

Clinical categories and molecular classes disagree at a rate that surprises people. The disagreement is systematic rather than random, which makes it interpretable.

Clinically HER2-positive disease is not uniformly HER2-enriched. In a trial that enrolled only centrally confirmed HER2-positive patients, 67% were HER2-enriched by PAM50 and 38 of 151, which is 25%, were luminal A or luminal B10.

Clinically ER-positive disease includes non-luminal calls. Ten percent of ER-positive tamoxifen-treated cases were assigned a non-luminal subtype13. The 1% to 10% ER band is considerably worse, and the numbers sit in ER and PR testing, thresholds, and the ER-low category.

Subtype is not fixed across time or site either. In 123 paired primary and metastatic samples, the rate of subtype conversion was 0% for basal-like tumours, 23.1% for HER2-enriched, 30.0% for luminal B and 55.3% for luminal A19. Luminal A tumours most often converted to luminal B.

The important move is to stop treating this as error. A clinical category answers whether a drug class is licensed for this patient. A molecular class answers which transcriptional programme the bulk of the tumour is running. Those are different questions, and there is no reason for the answers to coincide.

Discordance does matter when one system is used as a proxy for the other. An immunohistochemical surrogate for luminal B disease, using Ki-67 and HER2 alongside hormone receptors, was derived against gene expression profiling in 357 tumours and then applied to 404620. The surrogate is used because profiling is expensive, and it inherits the Ki-67 reproducibility documented in Ki-67, proliferation indices, and interobserver reproducibility. Consensus definitions built on such surrogates have been in general use for over a decade21.

9 · Single-cell and spatial data as a challenge to bulk-defined subtypes

Bulk expression is an average over every cell in the specimen, weighted by how many of each there are. A subtype called from bulk data is therefore partly a statement about composition.

A single-cell and spatially resolved atlas of human breast cancers made that explicit22. The authors built a single-cell method of intrinsic subtype classification and used it to reveal recurrent heterogeneity among neoplastic cells within individual tumours. Deconvolution of large bulk cohorts using single-cell signatures then stratified them into nine ecotypes, defined by cellular composition and associated with different clinical outcomes.

An ecotype is not a competing subtype. It is a different object. A subtype names the transcriptional programme of the dominant malignant population. An ecotype names what else is present and in what proportion.

Spatial measurement adds a third axis, which is arrangement. Imaging mass cytometry across large breast cancer cohorts resolved single-cell phenotypes in place and related them to genomic features and to outcome23,24,25. Two tumours with the same cellular composition can have different structure, and structure carries prognostic weight of its own.

Single-cell genomics makes the same point about the malignant compartment alone. Subclonal structure within one tumour was demonstrated by single-cell sequencing well before the transcriptional work26. Topographic single-cell sequencing then showed multiple clones invading together rather than sequentially27.

None of this shows that bulk subtypes are wrong. It shows that one label per tumour cannot carry composition, arrangement and subclonal structure at the same time. Much of the classifier disagreement in PAM50, BluePrint, and subtype calling in practice comes from asking a label to do exactly that.

Interplay

Classification and heterogeneity are the same problem seen from two directions. A taxonomy assigns one label per tumour, and heterogeneity is the statement that one label is not enough. Every argument in this chapter about discordance between classifiers, between platforms or between primary and metastasis is a heterogeneity argument arriving through the door of measurement rather than through the door of biology.

See Integrative biological interplay

10 · Classification as a changing framework rather than a fact

A taxonomy is a tool, and the first question about any tool is what it was made for.

Three questions separate a useful classification from a fashionable one. What was it built to predict. On what measurement. Against what comparator.

The intrinsic subtypes answer the first question differently at different points in their history, and that is the heart of the critique. They were derived to describe between-tumour variation in a gene set selected for stability within a tumour7. They were then used for prognosis8,12. They were then used to predict response10,11. Each of those is a separate claim requiring separate evidence. The evidence is strongest for the first use and thinnest for the third.

The second question is where most published disagreement lives. A class that different classifiers assign differently cannot be compared across studies15. A class derived from bulk expression cannot separate tumour from stroma without help18,22.

The third question decides whether a taxonomy enters practice. A classification changes care only when a trial has randomised a decision on it. That has happened for the multigene prognostic assays and it has not happened for intrinsic subtype, for integrative clusters, or for any single-cell derived scheme. The framework behind that statement is in Analytical validity, clinical validity, and clinical utility defined.

What survives all three questions is worth naming, because it is not nothing. Basal-like reproduces across classifiers and across measurement platforms. HER2-enriched carries predictive information beyond receptor status within HER2-positive disease. Copy-number-based clusters predict the timing of relapse rather than only its likelihood.

Caution

The instinct on meeting a new classification is to ask whether it is better than the old one. That question usually has no answer, because the two were built to predict different things. Ask instead what each was trained on, what it was validated against, and which decision it is being offered for. A taxonomy that cannot answer the third question is a research result rather than a clinical tool.

Classification in this disease has changed roughly once a decade and it will change again. The durable part is not the list of classes. It is the discipline of asking what a given class was built to do.

References

All 27 references
  1. Elston CW, Ellis IO. Pathological prognostic factors in breast cancer. I. The value of histological grade in breast cancer: experience from a large study with long-term follow-up. Histopathology 1991 19:403-410. PMID 1757079
  2. Weiss A, Chavez-MacGregor M, Lichtensztajn DY, et al. Validation study of the American Joint Committee on Cancer eighth edition prognostic stage compared with the anatomic stage in breast cancer. JAMA Oncol 2018 4:203-209. PMID 29222540
  3. Hammond MEH, Hayes DF, Dowsett M, et al. ASCO/College of American Pathologists guideline recommendations for immunohistochemical testing of estrogen and progesterone receptors in breast cancer. J Clin Oncol 2010 28:2784-2795. PMID 20404251
  4. Allison KH, Hammond MEH, Dowsett M, et al. Estrogen and progesterone receptor testing in breast cancer: ASCO/CAP guideline update. J Clin Oncol 2020 38:1346-1366. PMID 31928404
  5. Wolff AC, Hammond MEH, Allison KH, et al. Human epidermal growth factor receptor 2 testing in breast cancer: ASCO/College of American Pathologists clinical practice guideline focused update. J Clin Oncol 2018 36:2105-2122. PMID 29846122
  6. Kennecke H, Yerushalmi R, Woods R, et al. Metastatic behavior of breast cancer subtypes. J Clin Oncol 2010 28:3271-3277. PMID 20498394
  7. Perou CM, Sorlie T, Eisen MB, et al. Molecular portraits of human breast tumours. Nature 2000 406:747-752. PMID 10963602
  8. Sorlie T, Perou CM, Tibshirani R, et al. Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc Natl Acad Sci U S A 2001 98:10869-10874. PMID 11553815
  9. Cancer Genome Atlas Network. Comprehensive molecular portraits of human breast tumours. Nature 2012 490:61-70. PMID 23000897
  10. Llombart-Cussac A, Cortes J, Pare L, et al. HER2-enriched subtype as a predictor of pathological complete response following trastuzumab and lapatinib without chemotherapy in early-stage HER2-positive breast cancer (PAMELA). Lancet Oncol 2017 18:545-554. PMID 28238593
  11. Prat A, et al. HER2-Enriched subtype and ERBB2 expression in HER2-positive breast cancer treated with dual HER2 blockade. J Natl Cancer Inst 2020. PMID 31037288
  12. Parker JS, Mullins M, Cheang MCU, et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. J Clin Oncol 2009 27:1160-1167. PMID 19204204
  13. Nielsen TO, Parker JS, Leung S, et al. A comparison of PAM50 intrinsic subtyping with immunohistochemistry and clinical prognostic factors in tamoxifen-treated estrogen receptor-positive breast cancer. Clin Cancer Res 2010 16:5222-5232. PMID 20837693
  14. Krijgsman O, Roepman P, Zwart W, et al. A diagnostic gene profile for molecular subtyping of breast cancer associated with treatment response. Breast Cancer Res Treat 2012 133:37-47. PMID 21814749
  15. Weigelt B, Mackay A, A'hern R, et al. Breast cancer molecular profiling with single sample predictors: a retrospective analysis. Lancet Oncol 2010 11:339-349. PMID 20181526
  16. Curtis C, Shah SP, Chin SF, et al. The genomic and transcriptomic architecture of 2,000 breast tumours reveals novel subgroups. Nature 2012 486:346-352. PMID 22522925
  17. Rueda OM, Sammut SJ, Seoane JA, et al. Dynamics of breast-cancer relapse reveal late-recurring ER-positive genomic subgroups. Nature 2019 567:399-404. PMID 30867590
  18. Mertins P, Mani DR, Ruggles KV, et al. Proteogenomics connects somatic mutations to signalling in breast cancer. Nature 2016 534:55-62. PMID 27251275
  19. Cejalvo JM, Martinez de Duenas E, Galvan P, et al. Intrinsic subtypes and gene expression profiles in primary and metastatic breast cancer. Cancer Res 2017 77:2213-2221. PMID 28249905
  20. Cheang MC, Chia SK, Voduc D, et al. Ki67 index, HER2 status, and prognosis of patients with luminal B breast cancer. J Natl Cancer Inst 2009 101:736-750. PMID 19436038
  21. Goldhirsch A, Winer EP, Coates AS, et al. Personalizing the treatment of women with early breast cancer: highlights of the St Gallen International Expert Consensus 2013. Ann Oncol 2013 24:2206-2223. PMID 23917950
  22. 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
  23. Jackson HW, Fischer JR, Zanotelli VRT, et al. The single-cell pathology landscape of breast cancer. Nature 2020 578:615-620. PMID 31959985
  24. 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
  25. 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
  26. Navin N, Kendall J, Troge J, et al. Tumour evolution inferred by single-cell sequencing. Nature 2011 472:90-94. PMID 21399628
  27. 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