Part VII · Detection, diagnosis, and biomarker assessment · Chapter 39

Prognostic and predictive assays and somatic profiling

The assays agree about populations and disagree about patients, and the second fact is the one that reaches the clinic.

1 · Multigene prognostic and predictive assays

Multigene assays exist to answer one question. Can chemotherapy be safely omitted in a patient whose clinical features would otherwise indicate it.

Three prospective randomised trials answered versions of that question, and their designs differ in a way that determines how each result transfers.

TAILORx enrolled 10,273 women with hormone receptor-positive, HER2-negative, node-negative disease, of whom 9,719 were eligible with follow-up. The 6,711 with a midrange 21-gene recurrence score of 11 to 25, which was 69% of the eligible population, were randomised to endocrine therapy alone or to chemoendocrine therapy¹. Endocrine therapy alone was non-inferior, with a hazard ratio of 1.08 and a 95% confidence interval of 0.94 to 1.24. At 9 years, invasive disease-free survival was 83.3% against 84.3%, and overall survival was 93.9% against 93.8%. Chemotherapy benefit varied with the combination of score and age, with a P value of 0.004, and some benefit appeared in women aged 50 or younger with a score of 16 to 25.

MINDACT asked a different question. Among 6,693 women, 1,550 had high clinical risk and low genomic risk by the 70-gene signature, which is 23.2% of the enrolled population². Among those who received no chemotherapy, 5-year survival without distant metastasis was 94.7%, with a 95% confidence interval of 92.5% to 96.2%. The absolute difference against those who received chemotherapy was 1.5 percentage points, in favour of chemotherapy.

RxPONDER extended the 21-gene assay to node-positive disease. Among 5,083 randomised women with one to three positive nodes and a score of 25 or lower, the chemotherapy effect differed by menopausal status, with a P value of 0.008³. In postmenopausal women, 5-year invasive disease-free survival was 91.9% against 91.3%, with a hazard ratio of 1.02 and no benefit. In premenopausal women it was 89.0% against 93.9%, with a hazard ratio of 0.60 and a 95% confidence interval of 0.43 to 0.83.

TAILORx, MINDACT and RxPONDER are therefore three different experiments, and the clinical application is developed in Genomic assays and the decision to omit chemotherapy.

2 · How the assays differ and where they disagree

The assays were built from different genes, trained on different cohorts, against different endpoints, with cut-points chosen for different purposes. They should not be expected to agree, and they do not. Their validations were also run separately. The PAM50 risk of recurrence score added more prognostic information than the 21-gene score in node-negative disease in TransATAC⁴, and predicted ten-year distant recurrence in a Danish national cohort⁵. EndoPredict added information independent of clinicopathological factors⁶. A within-patient comparison of six signatures in 774 postmenopausal women from the ATAC trial ranked them for distant recurrence over ten years⁷. The PAM50 risk of recurrence score, the Breast Cancer Index and the EndoPredict clinical score gave the most prognostic information, each significantly more than the 21-gene recurrence score. In the 183 women with one to three involved nodes, every test gave substantially less.

OPTIMA prelim ran five tests on the same 313 tumours, using the actual diagnostic assays rather than in silico approximations. The proportion called low risk, or low to intermediate risk, was 82.1% by the 21-gene recurrence score, 65.5% by Prosigna, 72.0% by IHC4, 61.4% by MammaPrint and 61.6% by NexCourse Breast⁸.

The per-patient agreement is the number that matters. Only 119 tumours, which is 39.4%, were classified uniformly as either low to intermediate risk or high risk by all five tests. The remaining 183, which is 60.6%, were assigned to different risk categories by different tests⁸.

Subtype calling behaved the same way. Three subtype assays each assigned between 59.5% and 62.4% of tumours to luminal A. Only 121 tumours, which is 40.1%, were called luminal A by all three, and 123, which is 40.7%, were discordantly subtyped⁸.

That combination is the whole lesson. The tests produce nearly the same population-level distributions and disagree about six tumours in ten. A comparison of average prognostic performance therefore cannot resolve which test a given patient should have had, and no such comparison has been done as a randomised trial.

Adding clinical variables improves prognostic information without resolving the disagreement. Integrating the 21-gene score with grade, tumour size and age produced more prognostic information for distant recurrence than either the score or the clinical factors alone⁹.

3 · Somatic sequencing panels, scope, and actionability tiers

A sequencing panel is a fixed set of genomic regions read at a stated depth. Its scope determines what it cannot find, and scope is a design decision rather than a biological fact. A panel that does not cover a promoter will never report a promoter mutation, and the report will not say so.

Actionability is the second axis and it is graded by evidence rather than by biology. Applying the ESMO Scale for Clinical Actionability of Molecular Targets to breast cancer produced roughly 40 recurrent driver alterations ranked into tiers¹⁰. ERBB2 amplification, germline BRCA1 and BRCA2 mutations and PIK3CA mutations were tier IA, supported by large randomised trials. NTRK fusions and microsatellite instability were tier IC. ESR1 mutations and PTEN loss were tier IIA, and ERBB2 and AKT1 mutations tier IIB. Somatic BRCA1 and BRCA2 mutations, MDM2 amplification and ERBB3 mutations were tier III. Seventeen genes were tier IV on preclinical evidence alone.

The tier X category is the most instructive and the least quoted. FGFR1 and CCND1 were ranked tier X because previous studies showed a lack of actionability¹⁰. FGFR1 amplification is a real driver in breast cancer and it associates with endocrine resistance¹¹. Being a driver and being a useful target are separate claims, and the tier system exists to keep them apart.

Tiers are also testable, which is what makes them evidence rather than opinion. In plasmaMATCH, neratinib produced confirmed responses in 5 of 20 patients with HER2-mutant disease and capivasertib with fulvestrant in 4 of 18 with AKT1-mutant oestrogen receptor-positive disease. Extended-dose fulvestrant in ESR1-mutant disease produced 6 responses in 74 patients and missed its target¹². Three tier II alterations, three different answers.

The somatic landscape itself is developed in The somatic genomic landscape, and plasma genotyping in Mutation tracking for ESR1 and PIK3CA to guide treatment change.

4 · Immune biomarkers and their assay discordance

PD-L1 is the worked example of an eligibility criterion that is defined by an assay rather than by a biological threshold, and the data are unusually clear.

Samples from 614 patients in IMpassion130 were scored centrally with three antibodies. The proportion called positive at an immune cell cut-off of 1% or more was 46.4% with SP142, with a 95% confidence interval of 42.5% to 50.4%, 74.9% with SP263 and 73.1% with 22C3¹³. Using the 22C3 combined positive score of 1 or more, 80.9% were positive.

The assays are not measuring the same population. Analytical concordance with SP142 was 69.2% for SP263 and 68.7% for 22C3. Almost every SP142-positive case was also positive on the other assays, while 29.6% of SP263-positive and 29.0% of 22C3-positive cases were SP142-negative¹³.

Clinical benefit followed the narrower assay. Activity of atezolizumab with nab-paclitaxel in the SP263-positive and 22C3-positive populations was driven by the cases that were also SP142-positive, rather than by the single-positive cases¹³. Attempts to harmonise cut-points, using SP263 at 4% or 22C3 combined positive score at 10, reached only about 75% concordance with SP142 at 1%.

The structure of this problem is identical to HER2-low in HER2-low and ultralow reframed as a heterogeneity problem and to the analytic validity framework in Receptor assessment and the measurement problem. A boundary chosen to define trial eligibility becomes a boundary that governs access to a drug, and its reproducibility was never the property it was selected for. Nearly thirty percentage points of a population changes eligibility depending on which antibody the laboratory happened to buy.

Tumour-infiltrating lymphocytes escape the antibody problem and inherit an observer problem instead. They are scored on a haematoxylin and eosin slide against published conventions¹⁴, and they carry prognostic weight in triple-negative and HER2-positive disease^15,16. Their biology belongs to Breast cancer immunology and their use in selection to Predictive biomarkers and patient selection and First-line chemoimmunotherapy and the PD-L1 threshold.

5 · DNA repair and homologous recombination deficiency assays

Homologous recombination deficiency names three different measurements, and running them together is the commonest error in this area.

The first is a germline pathogenic variant in BRCA1 or BRCA2, which is a heritable property of the patient and is tested on blood. The second is a somatic alteration or loss in the tumour, including BRCA1 promoter hypermethylation, which silences the gene without mutating it and produces a tumour molecularly similar to a BRCA1-mutant one¹⁷. The third is a genomic scar, meaning a pattern of large-scale rearrangement or a mutational signature that records past repair failure.

Scar-based assays work. A homologous recombination deficiency score predicted response to platinum-containing neoadjuvant chemotherapy in triple-negative disease¹⁸. A classifier built from mutational signatures identified BRCA1 and BRCA2 deficiency including cases without a detected mutation¹⁹. The genomic scarring biology is in DNA repair deficiency and genomic scarring and DNA repair, replication stress, and genomic instability.

The temporal problem is the one that decides clinical use. A scar is a record of the past. It establishes that repair was deficient when the scar formed. It does not establish that repair is deficient now. A reversion mutation that restores the reading frame of BRCA1 or BRCA2 restores function and abolishes sensitivity, while leaving every scar in place. The scar-based assay will still read as deficient.

Guidelines reflect that uncertainty. ASCO found insufficient data to recommend routine testing for homologous recombination deficiency to guide metastatic breast cancer therapy²⁰. Germline testing, by contrast, has a clear action attached to it in both the adjuvant and metastatic settings, with OlympiA and TBCRC 048 defining the evidence and Germline and somatic DNA repair alterations the practice.

6 · Reporting standards and molecular tumor boards

A molecular report should state what was tested, at what depth, what was not covered, and how strong the evidence is for each alteration reported. Most of the interpretive difficulty in practice comes from the absence of the third item.

Variants of uncertain significance are produced by panel design, not by disease. A larger panel returns more of them, and the number on a report is therefore a property of the assay ordered. None of them supports an action, and a report that lists them without saying so invites the clinician to find meaning in noise.

The molecular tumour board exists to convert a variant list into a decision, and its commonest correct output is that nothing should change. That is a useful output and it is rarely recorded as one.

Three functions justify the meeting. It applies an evidence framework consistently, so that a tier IIB alteration is not treated as a tier IA one because it was the only thing found¹⁰. It knows what trials are open, which is where most tier II and tier III alterations actually lead Biomarker-selected, basket, umbrella, and platform trials. And it can distinguish a somatic finding from an incidental germline one, which carries consequences for the patient's family and requires a different pathway.

The diagnostic version of the same reconciliation discipline is in Multidisciplinary diagnostic integration, and the computational infrastructure in Data science and precision oncology.

7 · Reproducibility, clinical utility, and the evidence bar for adoption

The framework is set in Analytical validity, clinical validity, and clinical utility defined and is applied rather than restated here. Analytic validity asks whether the assay measures what it claims. Clinical validity asks whether the measurement associates with an outcome. Clinical utility asks whether acting on the measurement improves that outcome.

Almost every assay in this chapter has the first two. The bar that separates adoption from research use is the third, and it requires a design that randomises the action rather than observing the association.

TAILORx and MINDACT cleared it, and the way they cleared it is the point. Both randomised the decision that the assay was meant to inform, rather than measuring how well the assay predicted outcome^1,2. RxPONDER then showed that the same assay's predictive claim can fail in a different population³. Utility is a claim about an assay in a population in a decision, and moving any of the three can break it.

Reproducibility is the gate before utility, and it is where several assays stop. When 18 pathologists from 15 institutions scored 170 breast biopsies, agreement on cases called immunohistochemistry 0 was 25%²¹. PD-L1 antibodies assign eligibility to populations differing by nearly thirty percentage points¹³. Five multigene assays disagree about six tumours in ten⁸. An assay whose result depends materially on who ran it cannot carry a treatment decision, whatever its published hazard ratio.

The same standard applied to research instruments in What is actionable today and what remains a research instrument applies here without modification. The burden of showing utility sits with whoever proposes the test.

8 · Access to molecular testing and who is tested

An assay that exists is not an assay that is used, and the distribution of testing does not follow the distribution of disease.

Testing requires an unbroken chain. Tissue must be adequate and handled correctly Tissue acquisition, specimen adequacy, and handling. A laboratory must run the assay or a specimen must be shipped. A clinician must order it and know what the result means. Someone must pay. An available drug or an open trial must exist at the end. A break at any link produces the same outcome, which is that the patient is not tested, and the reasons are rarely visible in the medical record.

Two documented findings show how the problem reaches the assay itself rather than only its distribution. The randomised trial that provides the strongest evidence for artificial intelligence-supported screening did not collect race and ethnicity data, so its results cannot be examined by subgroup²². Among the 21,631 WISDOM participants who received a polygenic risk score, differences between major racial and ethnic groups were significant at P below 0.001, and the score had to be adapted for use in diverse populations²³. The size of those differences is not reported. A tool derived in one population carries the composition of that population inside it.

The consequences are developed where the data live, in Access to pathology, molecular testing, and imaging for access to pathology and molecular testing, in Clinical trial representation and the mechanisms of exclusion for trial representation, and in Disparities across the breast cancer continuum for the general decomposition of an outcome gap.

References

All 23 references

1. Sparano JA, Gray RJ, Makower DF, et al. Adjuvant chemotherapy guided by a 21-gene expression assay in breast cancer. N Engl J Med 2018 379:111-121. PMID 29860917
2. Cardoso F, van 't Veer LJ, Bogaerts J, et al. 70-gene signature as an aid to treatment decisions in early-stage breast cancer. N Engl J Med 2016 375:717-729. PMID 27557300
3. Kalinsky K, Barlow WE, Gralow JR, et al. 21-gene assay to inform chemotherapy benefit in node-positive breast cancer. N Engl J Med 2021 385:2336-2347. PMID 34914339
4. Dowsett M, Sestak I, Lopez-Knowles E, et al. Comparison of PAM50 risk of recurrence score with oncotype DX and IHC4 for predicting risk of distant recurrence after endocrine therapy. J Clin Oncol 2013 31:2783-90. PMID 23816962
5. Lænkholm AV, Jensen MB, Eriksen JO, et al. PAM50 Risk of Recurrence Score Predicts 10-Year Distant Recurrence in a Comprehensive Danish Cohort of Postmenopausal Women Allocated to 5 Years of Endocrine Therapy for Hormone Receptor-Positive Early Breast Cancer. J Clin Oncol 2018 36:735-740. PMID 29369732
6. Filipits M, Rudas M, Jakesz R, et al. A new molecular predictor of distant recurrence in ER-positive, HER2-negative breast cancer adds independent information to conventional clinical risk factors. Clin Cancer Res 2011 17:6012-20. PMID 21807638
7. Sestak I, Buus R, Cuzick J, et al. Comparison of the Performance of 6 Prognostic Signatures for Estrogen Receptor-Positive Breast Cancer: A Secondary Analysis of a Randomized Clinical Trial. JAMA Oncol 2018 4:545-553. PMID 29450494
8. Bartlett JM, Bayani J, Marshall A, et al. Comparing breast cancer multiparameter tests in the OPTIMA Prelim trial: no test is more equal than the others. J Natl Cancer Inst 2016 108:djw050. PMID 27130929
9. Sparano JA, Crager MR, Tang G, Gray RJ, Stemmer SM, Shak S. Development and validation of a tool integrating the 21-gene recurrence score and clinical-pathological features to individualize prognosis and prediction of chemotherapy benefit in early breast cancer. J Clin Oncol 2021 39:557-564. PMID 33306425
10. Condorelli R, Mosele F, Verret B, et al. Genomic alterations in breast cancer: level of evidence for actionability according to ESMO Scale for Clinical Actionability of molecular Targets. Ann Oncol 2019 30:365-373. PMID 30715161
11. Turner N, Pearson A, Sharpe R, et al. FGFR1 amplification drives endocrine therapy resistance and is a therapeutic target in breast cancer. Cancer Res 2010 70:2085-2094. PMID 20179196
12. Turner NC, Kingston B, Kilburn LS, et al. Circulating tumour DNA analysis to direct therapy in advanced breast cancer (plasmaMATCH): a multicentre, multicohort, phase 2a, platform trial. Lancet Oncol 2020 21:1296-1308. PMID 32919527
13. Rugo HS, Loi S, Adams S, et al. PD-L1 immunohistochemistry assay comparison in atezolizumab plus nab-paclitaxel-treated advanced triple-negative breast cancer. J Natl Cancer Inst 2021 113:1733-1743. PMID 34097070
14. Salgado R, Denkert C, Demaria S, et al. The evaluation of tumor-infiltrating lymphocytes (TILs) in breast cancer: recommendations by an International TILs Working Group 2014. Ann Oncol 2015 26:259-271. PMID 25214542
15. Denkert C, von Minckwitz G, Darb-Esfahani S, et al. Tumour-infiltrating lymphocytes and prognosis in different subtypes of breast cancer: a pooled analysis of 3771 patients treated with neoadjuvant therapy. Lancet Oncol 2018 19:40-50. PMID 29233559
16. Loi S, Drubay D, Adams S, et al. Tumor-infiltrating lymphocytes and prognosis: a pooled individual patient analysis of early-stage triple-negative breast cancers. J Clin Oncol 2019 37:559-569. PMID 30650045
17. Glodzik D, Bosch A, Hartman J, et al. Comprehensive molecular comparison of BRCA1 hypermethylated and BRCA1 mutated triple negative breast cancers. Nat Commun 2020 11:3747. PMID 32719340
18. Telli ML, Timms KM, Reid J, et al. Homologous recombination deficiency (HRD) score predicts response to platinum-containing neoadjuvant chemotherapy in patients with triple-negative breast cancer. Clin Cancer Res 2016 22:3764-3773. PMID 26957554
19. Davies H, Glodzik D, Morganella S, et al. HRDetect is a predictor of BRCA1 and BRCA2 deficiency based on mutational signatures. Nat Med 2017 23:517-525. PMID 28288110
20. Henry NL, Somerfield MR, Dayao Z, et al. Biomarkers for systemic therapy in metastatic breast cancer: ASCO guideline update. J Clin Oncol 2022 40:3205-3221. PMID 35759724
21. Robbins CJ, Fernandez AI, Han G, et al. Multi-institutional assessment of pathologist scoring HER2 immunohistochemistry. Mod Pathol 2023 36:100032. PMID 36788069
22. Lang K, Josefsson V, Larsson AM, et al. Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI). Lancet Oncol 2023 24:936-944. PMID 37541274
23. Fergus KB, Heise RS, Madlensky L, et al. Integrating breast cancer polygenic risk scores at scale in the WISDOM Study: a national randomized personalized screening trial. Genome Med 2025 17:97. PMID 40877879