What KEYNOTE-522 settled, and why every attempt to take a drug back out still reads out at the operation.
1 · Scope, and why pathological complete response carries most weight in this subtype
Early disease · TNBC · neoadjuvant
Choosing what to add to preoperative chemotherapy, and what to take out
What can a preoperative trial settle about a new drug, and what has to wait for a randomised trial after surgery?
Rationale
The preoperative window is the fastest read in this disease. The tumour is still in place. The endpoint arrives at the operation. Paired biopsies make the mechanism visible. So it is the natural place to try a conjugate, a targeted addition or a physical primer, well before anyone commits to a trial with survival as its endpoint.
Experiment
An adaptive platform runs candidate agents against a shared chemotherapy control and graduates them on a Bayesian probability of success in a future trial. A conjugate was given alone in a single arm, with treatment after those cycles left to the treating physician. A JAK inhibitor was added to preoperative chemotherapy in inflammatory triple-negative disease, with a pharmacodynamic conversion as the primary endpoint. Preoperative radiation dose was randomised as an immune primer, with immunological endpoints. The randomised conjugate trials in this disease all sit after surgery, in patients with residual disease.
Finding
Conjugate monotherapy cleared a minority of untreated tumours. That is enough to say the class is active before surgery. It is not enough to say it belongs there. The JAK inhibitor lowered its pathway marker in the biopsies taken after the short run-in. Sustained dosing alongside the taxane raised the same signalling rather than lowering it. Response in that rare population stayed low.
Limitation
A graduation probability is a prediction about a trial not yet done rather than a result. The conjugate trial is single-arm and response-guided. What patients received after those cycles was at their physician's discretion. Its event-free survival therefore describes a pathway rather than a drug. The inflammatory trial enrolled very few patients. Its response figure rests on a handful of events. Its own primary endpoint was a pharmacodynamic conversion proportion, and the published report never states that proportion. Its run-in sits inside the randomisation rather than before it, so its two comparisons overlap. The radiation trial's endpoints are immunological. Its dose comparisons are exploratory and do not rank doses for efficacy. None of the randomised conjugate trials asks a preoperative question. There is no randomised basis for moving one in front of surgery.
Next question
Does the conjugate stay behind surgery, where the residual-disease trials are already randomising it, or is a preoperative comparison against the standard regimen worth mounting first?
Allocating further treatment on what surgery finds
Does what surgery finds after neoadjuvant treatment identify who needs more treatment?
Rationale
Giving the same chemotherapy before surgery rather than after it does not change survival. The reason to give it first is that surgery then reports whether the drugs worked. In triple-negative disease that report separates the population more sharply than anything measurable before treatment. A complete response is common in this subtype. Residual invasive disease relapses within a few years of surgery rather than over the decades that hormone receptor-positive disease can take.
Experiment
The foundational trials randomised the timing of chemotherapy, and then the addition of a taxane to the preoperative course, with survival as the endpoint and pathological response recorded alongside. A platform trial read response against outcome across its molecular subtypes on a shared chemotherapy control. Randomised trials that added carboplatin to preoperative chemotherapy reported the response rate first and their survival endpoints years later. The registration trial for chemo-immunotherapy carried both endpoints inside one randomisation.
Finding
Response sorts the population steeply. Patients who clear the breast and the axilla relapse rarely. Residual disease accounts for most of the events. That sorting is prognostic within an arm. It has repeatedly failed to behave as a surrogate between arms. Carboplatin raised the response rate over the same chemotherapy without it, and the later outcome analyses of those same trials found no gain. Where a trial moved both endpoints together, that is agreement rather than evidence that the first predicts the second.
Limitation
The timing trials enrolled all comers before subtype was routinely recorded. The response gradient they established is being read into a subtype they did not define. What counts as a complete response is a convention rather than a fact. The trials differ on whether residual in-situ disease disqualifies it. Residual cancer burden grades what that dichotomy discards. What the pathologist reports also depends on how the specimen and the post-chemotherapy axilla were sampled. Residual disease marks resistance to the regimen that was actually given. It says nothing about drugs the tumour never met.
Next question
If residual disease marks the population carrying the events, does giving those patients a further course of a different chemotherapy lower recurrence?
Pathological complete response matters more here than anywhere else in this book, and the reason is quantitative. A pooled analysis of 12 trials and 11,955 patients measured what clearance of the breast and nodes is worth in this subtype. The event-free survival hazard ratio was 0.24, with a 95% confidence interval of 0.18 to 0.33. The overall survival hazard ratio was 0.16, with an interval of 0.11 to 0.251. No other readily available variable in this disease separates outcome that sharply.
That gradient was established before subtype was routinely recorded. NSABP B-18 and NSABP B-27 randomised when chemotherapy was given and then what was added to the preoperative course, carried survival as the endpoint, and recorded response alongside it. Both enrolled all comers. The sorting this subtype is now managed by was measured in a population that did not distinguish it.
The same analysis is why the endpoint cannot be used to judge a drug. At trial level, the increase in response frequency between treatment groups explained almost none of the variation in event-free survival, with a coefficient of determination of 0.03. For overall survival it was 0.241. The distinction between a patient-level prognostic marker and a trial-level surrogate is worked out in Endpoints, surrogates, and the overall survival question, and it governs every table below.
Two trials in this chapter show both halves of that at once. Carboplatin raised pathological complete response in CALGB 40603 and produced no long-term gain2. Nab-paclitaxel with carboplatin raised it over nab-paclitaxel with gemcitabine in WSG-ADAPT-TN, and the invasive disease-free survival hazard ratio for the arm producing the higher response rate was 1.043. Inside that same trial, five-year invasive disease-free survival was 90.6% in patients who cleared the breast and the axilla and 62.8% in those who did not. One dataset, two directions. The gradient within an arm is steep and the shift between arms carries almost nothing.
GeparNuevo shows the failure inside a single trial rather than across a pooled dataset. Its response endpoint was not met and its survival curves separated regardless, and the authors' own analysis puts that separation outside the response effect4. Whatever moved those curves reached patients whose tumours were not eradicated, which is exactly the population a response-mediated mechanism cannot reach.
What counts as a complete response is also a convention, and the trials here do not share one. GeparSixto scored it as no residual invasive or in-situ carcinoma in breast and nodes. WSG-ADAPT-TN and INFORM allowed residual in-situ disease. CALGB 40603 reported the breast alone and the breast with the axilla as two separate rates. Eradication from the breast alone carries a weaker association with outcome than eradication from both1, and what the pathologist can report depends on how the specimen was sampled and how the treated axilla was handled.
Residual disease is not one prognosis either. It is a spectrum, which is why residual cancer burden and the continuous measures in Residual cancer burden and continuous response measures carry more than a binary result. Stromal lymphocyte density adds a further axis. It is prognostic in this subtype, is scored to a published convention5,6, and is readable on the residual disease specimen as well as on the diagnostic core, which the long-term GeparNuevo report does7. Those subgroups are small, so the direction is the finding rather than the size.
So read pathological complete response as an instruction about what to do next rather than as a verdict on what was just given. That is the use that survives the surrogacy failure, and the use on which Adjuvant and post-neoadjuvant therapy depends.
That is also the limit of what a preoperative study can establish alone. The trials in this chapter that stop at the operation agree on one thing and one thing only. They show that an agent is active against untreated triple-negative disease, and they show it within months of the first dose.
They differ in almost everything else, and the difference is design rather than drug. I-SPY 2 runs candidate agents against a shared chemotherapy control and graduates them on a Bayesian probability of success in a future phase III trial8. That probability is a prediction about a trial nobody has run. NeoSTAR gave a conjugate alone for four cycles in a single arm, and what followed was left to the treating physician9. Its event-free survival therefore describes a treatment pathway rather than a drug. TBCRC 039 added a JAK inhibitor to preoperative chemotherapy in triple-negative inflammatory disease, with a pharmacodynamic conversion as its primary endpoint10. P-RAD / TBCRC 053 randomises preoperative radiation dose as an immune primer and reads out on immunological endpoints. Four different currencies, and none of them is recurrence.
The inflammatory trial earns its space because its pharmacodynamic result inverts. Phosphorylated STAT3 fell in the biopsies taken after a short run-in of the inhibitor alone. Sustained dosing alongside the taxane raised the same pathway rather than lowering it10. The run-in sat inside the randomisation rather than in front of it, so the two comparisons overlap. The trial also enrolled very few patients, and its response figure rests on a handful of events.
None of this licenses moving a drug in front of surgery. The randomised conjugate trials in this disease all sit behind it. ASCENT-05/OptimICE-RD11, TROPION-Breast0312 and SASCIA / GBG 102 randomise a conjugate in patients left with residual disease after neoadjuvant chemotherapy. A single-arm response rate in an untreated tumour and a randomised outcome in residual disease are not two readings of one experiment. There is no randomised basis for a preoperative conjugate, because nobody has built the comparison.
So the division of labour in this chapter is fixed. A preoperative trial retires a candidate quickly and makes a mechanism visible in paired biopsies. It cannot license adding a drug to a curative regimen, and it cannot license taking one out.
Each section carries its own table and its own question from the [Chapter Stories](stories.html) map. A trial recurs where it bears on more than one of them. The stories carry no figures at all, so a story and a table can never disagree about a number. Read the population column of every table before the result column, because these trials were built on different populations.
In practice
Report response as breast and nodes together. Eradication from the breast alone carries a weaker association with outcome than eradication from both1.
Use response to assign the next treatment, not to evaluate the last one. Assigning on observed response is what Adjuvant and post-neoadjuvant therapy rests on, and it is the patient-level use that survives the surrogacy failure.
When a regimen raises pathological complete response and has no mature survival data, say that the survival question is open. The trial-level correlation between the two is close to nothing1.
Ask for stromal lymphocyte scoring on the residual disease specimen where it is available. It is prognostic and it costs a slide7,6.
2 · Anthracycline and taxane backbones
Choosing what to add to preoperative chemotherapy, and what to take out, continued
Does the taxane formulation change what the preoperative window shows?
Rationale
Nab-paclitaxel removes the solvent that limits how much paclitaxel can be given and changes how the drug reaches the tumour. A higher delivered dose without the solvent was expected to raise the response rate, and the preoperative setting measures response directly.
Experiment
Randomisation of nab-paclitaxel against paclitaxel, each for four cycles before an anthracycline regimen, in HER2-negative early or locally advanced disease, with pathological complete response as the endpoint.
Finding
The pathological complete response rate was not improved.
Limitation
The trial enrolled HER2-negative disease rather than this subtype alone, so the result is an average over two populations whose response rates differ by a factor of several. A negative result on response does not establish that the two formulations are equivalent on toxicity, which is the axis on which the choice is usually made.
Next question
If the formulation does not change response, does it change who can complete the regimen?
ETNA
The chemotherapy backbone in this disease was inherited rather than optimised for it. Every trial in this chapter adds something to a sequential taxane and anthracycline regimen established across all subtypes. None of the component drugs was selected for triple-negative biology.
Giving the same chemotherapy before rather than after surgery does not change survival. In the individual patient data meta-analysis of ten trials, fifteen-year distant recurrence was 38.2% with neoadjuvant treatment against 38.0%, with a rate ratio of 1.02. Local recurrence was higher, at 21.4% against 15.9%, with a rate ratio of 1.3713. Those trials predate modern surgical planning and response-adapted local therapy. The point that survives is that the timing buys information and conservative surgery rather than a survival advantage in itself.
The taxane question was asked twice with the same drugs and answered twice differently. GeparSepto / GBG 69 found that nab-paclitaxel raised pathological complete response over solvent-based paclitaxel across all subtypes14. ETNA tested the same substitution in HER2-negative disease and found no significant difference15. Neither trial isolated the formulation. The first gave nab-paclitaxel at 150 mg/m2 until an amendment lowered it to 125, against a comparator at 80 mg/m214. The second gave 125 mg/m2 throughout against a comparator at 90 mg/m2 on a different cycle length15. A dose difference and a schedule difference are inside what is reported as a formulation comparison.
Dose density belongs to the same inheritance, and its evidence is in Adjuvant chemotherapy regimens and dose density. Several trials here use a dose-dense anthracycline phase and others do not, which is one more reason cross-trial response rates are not comparable.
The live backbone question is whether the anthracycline is needed at all.
Regimens that remove the anthracycline from the preoperative phase or defer it until after surgery.
Only the last row randomises that comparison against an outcome.
atezolizumab added to carboplatin and nab-paclitaxel vs carboplatin and nab-paclitaxel, both followed by adjuvant anthracycline
event-free survival. The pathological complete response tabulated here is a secondary endpoint
pathological complete response 48.6% with atezolizumab added to carboplatin and nab-paclitaxel against 44.4% without, odds ratio 1.18 (0.74-1.89), P=0.48
Patients aged 18 or older with stage I to III triple-negative breast cancer.
Weekly paclitaxel plus carboplatin followed by dose-dense doxorubicin and cyclophosphamide versus carboplatin plus docetaxel
Pathologic complete response in breast and axilla
Pathologic complete response was 54% with weekly paclitaxel plus carboplatin followed by doxorubicin and cyclophosphamide and 54% with carboplatin plus docetaxel. Residual cancer burden 0 or I was 67% in both arms.
Patients with triple-negative breast cancer, oestrogen and progesterone receptor under 1% and HER2-negative, cT1c to cT4c, cN0 or node-positive.
Nab-paclitaxel 125 mg/m2 with gemcitabine 1000 mg/m2 on days 1 and 8 every 3 weeks versus nab-paclitaxel 125 mg/m2 with carboplatin AUC 2 on days 1 and 8 every 3 weeks, each for 12 weeks.
Pathological complete response (ypT0/is ypN0) by therapy arm; the trial was also powered for early response, defined as a Ki-67 decrease of more than 30% or fewer than 500 invasive tumour cells in the three-week serial biopsy.
Pathological complete response was 28.7% (95% CI 0.22 to 0.36) with nab-paclitaxel plus gemcitabine and 45.9% (95% CI 0.38 to 0.54) with nab-paclitaxel plus carboplatin, 95% CI for the difference 6.2% to 27.9%, P = .002. Pathological complete response was lower in early non-responders than in early responders, 19.5% versus 44.4%, P < .001, and was 50.0% in patients whose early response was unclassifiable.
SCARLET3 · ongoing
TNBC stage I-III
carboplatin, docetaxel and pembrolizumab vs the anthracycline-containing KEYNOTE-522 regimen
event-free survival
not yet reported
The rows that have reported converge on the response question. A carboplatin and taxane doublet clears the breast and nodes about as often as the same drugs followed by an anthracycline. Only NeoSTOP randomised that comparison directly, and its two arms landed on the same response rate18, with the same residual cancer burden distribution. What separated them was haematological toxicity and cost rather than efficacy. Grade 3 or 4 neutropenia occurred in 60% of the anthracycline arm against 8% without it. That trial randomised 100 patients and read out at surgery.
The rest of the table diverges by what each trial chose to randomise. WSG-ADAPT-TN gave no anthracycline up front in either arm and randomised the partner drug instead19. Treatment after those twelve weeks depended on how the tumour had responded, so its survival comparison is confounded by what came next rather than by what was allocated3. NeoPACT added pembrolizumab to the short regimen in a single arm16. NeoTRIP kept the anthracycline and moved it behind surgery, then randomised the checkpoint inhibitor on what was left in front17. SCARLET is the only row that randomises the short regimen against the standard one with event-free survival as its endpoint, and it has not reported.
So the decision rests on a trial that has not read out. Three of the five rows were powered for what is found at the operation, on the endpoint this chapter has already shown moves without survival moving with it. The two powered for recurrence have not reported it.
The table also has no column for what the omission is meant to buy. Cardiac failure and treatment-related leukaemia arrive years after the last dose, in a population young enough that the relevant follow-up runs into decades. None of these trials was built to count those events, and the anthracycline-free regimens carry their own costs in neuropathy and marrow suppression. The argument is taken up in The other immunotherapy trials and why they read out differently, because in practice it cannot be separated from the immunotherapy question.
3 · Platinum and the germline BRCA question
Choosing what to add to preoperative chemotherapy, and what to take out, continued
Does adding carboplatin to preoperative chemotherapy do more than raise pathological complete response?
Rationale
Triple-negative tumours are enriched for defective homologous recombination. Platinum works by making the interstrand crosslinks that this repair pathway exists to resolve. So a drug the rest of breast cancer has little use for should earn its place here. The preoperative setting is where that prediction could be read quickly.
Experiment
Randomised addition of carboplatin to a taxane and anthracycline backbone, in trials that differed in the backbone and in what else was randomised alongside it. The endpoint in each was pathological complete response. Later trials put the same addition against event-free or disease-free survival instead, giving the platinum before surgery, after it, or either way.
Finding
Carboplatin raised pathological complete response over the same backbone without it, in trial after trial. The trials that used survival rather than response as their endpoint found fewer events in the carboplatin arm. One of the response trials was followed for years afterwards and found no long-term advantage from the carboplatin it had added. One trial that set carboplatin against a different partner drug on the same taxane backbone saw its response advantage gone by the time survival was counted.
Limitation
These trials vary the platinum and the backbone together. The variable under test is not the platinum alone. They differ in which anthracycline was given, in whether a second experimental agent was randomised alongside in a factorial, and in how much of the planned chemotherapy survived the added toxicity. In the trial where the response advantage disappeared, treatment after the opening block of chemotherapy depended on how the tumour had responded. Its survival comparison is therefore confounded by what came next. One of the survival trials violated the proportional hazards assumption its summary estimate is computed under. That estimate does not describe the whole of its follow-up. Pathological complete response is also prognostic within an arm rather than a demonstrated surrogate across arms. The older trial that doubled it by adding a preoperative taxane, and changed no survival curve, is the standing demonstration of the difference.
Next question
If the benefit comes from damage the tumour cannot repair, can the tumours that cannot repair it be identified before treatment starts?
Does a homologous recombination deficiency assay, or a germline BRCA mutation, identify who needs the platinum?
Rationale
If platinum works because a crosslink goes unrepaired, tumours that can repair one should not need it. A test for the defect would let the drug be allocated rather than given to everyone. That matters here because platinum toxicity is what limits delivery of the rest of the regimen.
Experiment
Two randomised trials built around the marker rather than around the drug. One set single-agent platinum against a single-agent taxane, and asked whether a deficiency score above a fixed cut-point predicted response to the platinum specifically. The other set single-agent platinum against an anthracycline regimen in germline carriers. A third trial selected patients by the deficiency assay and put a PARP inhibitor in the platinum's place. A single-arm window study gave a PARP inhibitor to patients selected for none of this.
Finding
The deficiency score did not separate response to platinum from response to a taxane. In germline carriers, platinum produced no more complete responses than the anthracycline regimen those patients would have had anyway. Selecting on the assay did not make the platinum-free arm clear the response threshold set for it. Unselected tumours given a PARP inhibitor responded often enough to make the marker look less like a gate than a description.
Limitation
The biomarker trial was sized to test a predictive claim rather than to compare two drugs. Its arm-to-arm difference is not an efficacy result. A large share of its patients crossed to another preoperative regimen before surgery. The assay could not be scored on all of the rest. Entry to the germline trial required a mutation rather than triple-negative disease. Hormone receptor-positive patients therefore sit inside its answer. The assay-selected trial tested one arm against a fixed threshold rather than against its comparator. Its arms can be described and not compared. No trial here kept an assay-negative arm on the same treatment, which is the comparison that would show what the selection bought.
Next question
If the assay cannot allocate the platinum, can it allocate a PARP inhibitor? And can that drug replace the chemotherapy rather than join it?
TBCRC 030INFORMGeparOLAPETREMAC
Can a PARP inhibitor stand in for the chemotherapy before surgery rather than be added to it?
Rationale
A PARP inhibitor kills cells that cannot repair a double-strand break. That is the same vulnerability the platinum is aimed at. If the oral drug does that work, the case for the cytotoxic backbone weakens. Its toxicity is of a different kind, which is the whole appeal in a disease being treated for cure.
Experiment
The designs run from addition to replacement. One trial randomised a PARP inhibitor on top of a platinum and taxane backbone, with a third arm carrying neither. One substituted the PARP inhibitor for the platinum in assay-selected patients. Two gave a PARP inhibitor as the only preoperative treatment, one for a full course in germline carriers, one as a short window before the chemotherapy in unselected disease. A further single-arm study pairs a PARP inhibitor with a checkpoint inhibitor and no chemotherapy at all.
Finding
Added to a platinum-containing backbone, the PARP inhibitor added nothing to it. Substituted for the platinum in selected patients, it did not clear the response threshold the trial had set itself, although it landed near the platinum arm. Given alone for a full course in germline carriers, it cleared tumours at a rate close to what combination chemotherapy achieves, and still fell short of its own prespecified bar. A short window of it, in patients unselected for a repair defect, produced objective responses in many of the tumours it was given to.
Limitation
The addition trial used a PARP inhibitor that traps the enzyme on DNA weakly. A null result there is a result about one molecule rather than about the class. The monotherapy trials are single-arm. Their response rates have no concurrent comparator, so they get read against chemotherapy rates from other trials in other populations. One of them reports its response on an evaluable denominator rather than on everyone enrolled. A window of a few weeks followed by planned chemotherapy says nothing about whether that chemotherapy could have been left out. The chemotherapy-free combination has no PARP-only arm. Nothing in it separates the checkpoint inhibitor's contribution from the PARP inhibitor's.
Next question
If chemotherapy is to come out, the other candidate for carrying the treatment is the immune response. Does adding a checkpoint inhibitor before surgery lower recurrence?
BrighTNessGeparOLANEOTALAPETREMACTBCRC 056
Platinum before surgery.
Four of these rows randomised the drug into the backbone. Two substituted it, one in germline carriers and one testing a repair assay as a predictor of response.
Germline BRCA1/2 mutation carriers with cT1-3 tumours of at least 1.5 cm, cN0-3, HER2-negative, stage I to III breast cancer; 70% had triple-negative disease and the remainder ER-positive HER2-negative disease.
Preoperative cisplatin 75 mg/m2 every 3 weeks for four doses versus doxorubicin 60 mg/m2 with cyclophosphamide 600 mg/m2 every 2 to 3 weeks for four doses, each followed by surgery.
Pathological complete response (ypT0/is, N0) confirmed by central review, the trial asking whether pathological complete response was at least 20% higher with cisplatin than with doxorubicin-cyclophosphamide.
Pathological complete response was 18% with cisplatin and 26% with doxorubicin-cyclophosphamide, risk ratio 0.70 (90% CI 0.39 to 1.2). Residual cancer burden 0 or 1 was 33% with cisplatin and 46% with doxorubicin-cyclophosphamide, risk ratio 0.73 (90% CI 0.50 to 1.1).
Patients with germline BRCA1/2 wild-type or unknown stage I to III triple-negative breast cancer.
Twelve weeks of preoperative single-agent cisplatin versus single-agent paclitaxel, with crossover to an alternative provider-selected chemotherapy offered for inadequate clinical response.
Association of the homologous recombination deficiency score, positive at 33 or above, with pathological response defined as residual cancer burden 0 or 1 after single-agent cisplatin or paclitaxel.
Residual cancer burden 0 or 1 was 26.4% with cisplatin and 22.3% with paclitaxel. No significant association was observed between homologous recombination deficiency score and response to cisplatin, odds ratio for residual cancer burden 0 or 1 if homologous recombination deficiency positive 2.22 (95% CI 0.39 to 23.68), or to paclitaxel, odds ratio 0.90 (95% CI 0.19 to 4.95), and there was no evidence of an interaction.
stage II-III triple-negative and HER2-positive disease; 315 triple-negative
carboplatin added to weekly paclitaxel and non-pegylated liposomal doxorubicin vs the same backbone without carboplatin
pathological complete response, ypT0 ypN0
pCR 53.2% with carboplatin vs 36.9% in triple-negative disease, P=0.005. No gain in HER2-positive disease, interaction P=0.015
Adding carboplatin to a neoadjuvant backbone raises pathological complete response. The rows above that randomised it into a backbone and read out at the operation agree on that, across different anthracyclines and different schedules. Two questions follow. Whether the response gain becomes a survival gain, and whether any test identifies the patients who need the drug.
The divergence between those trials is in the backbone rather than in the platinum. GeparSixto added carboplatin to weekly paclitaxel and a liposomal anthracycline given concurrently, with bevacizumab in every triple-negative patient25. CALGB 40603 added it to weekly paclitaxel followed by dose-dense doxorubicin and cyclophosphamide, in a factorial that also randomised bevacizumab24. BrighTNess added it to weekly paclitaxel followed by the same anthracycline pair, in a three-arm design whose other question was a PARP inhibitor23. Each varies the platinum and something else at once. In GeparSixto and CALGB 40603 the added toxicity also cost treatment delivery, which is a second variable moving with the drug under test. In BrighTNess it did not substantially affect delivery of the rest of the regimen.
The survival evidence is mixed, and the mixture is informative rather than contradictory. Long-term follow-up of CALGB 40603 found no outcome gain from the carboplatin it had added2. BrighTNess pointed the other way at four and a half years, on event-free survival between its carboplatin and paclitaxel arm and paclitaxel alone26. That is a randomised carboplatin comparison. It is not the veliparib question the trial was designed around, and overall survival did not differ between its arms. GeparSixto reported a disease-free survival advantage in its triple-negative stratum27. WSG-ADAPT-TN set carboplatin against gemcitabine on the same anthracycline-free taxane, and its response advantage was gone by the time survival was counted3. None of these trials was powered for the survival question it is being asked.
PEARLY is what the decision rests on. It randomised carboplatin into an anthracycline and taxane regimen, powered on event-free survival rather than on response, and met that endpoint20. Two features qualify it. Patients could be treated before or after surgery, so its population is not the preoperative one this chapter otherwise describes. And its overall survival favoured carboplatin without reaching significance, which is the endpoint a curative addition ultimately has to move. CITRINE asked the question entirely after surgery and was positive on disease-free survival, with its proportional hazards assumption violated28. A single hazard ratio computed under an assumption the data reject does not describe the whole of the follow-up.
The demonstration that keeps all of this honest is older than any of them. In NSABP B-27 a preoperative taxane doubled pathological complete response and changed neither disease-free nor overall survival29. That is the gap between the two uses of the endpoint, shown once, with the drug held constant.
The part that surprises people is that the platinum benefit is not confined to germline carriers. In the homologous recombination analysis of GeparSixto, carboplatin raised response sharply in deficient tumours and modestly in non-deficient ones. The interaction test was not significant, and disease-free survival improved in both groups27. Deficiency predicted response to the regimen as a whole, and not benefit from carboplatin specifically.
The apparent contradiction with the metastatic evidence dissolves once the designs are compared. In advanced disease, TNT tested carboplatin against docetaxel in unselected disease and found no difference in response overall. Among germline BRCA1 and BRCA2 carriers, carboplatin produced double the response rate of docetaxel, at 68% against 33%, with a biomarker-by-treatment interaction p value of 0.0130. No such benefit appeared in patients selected by BRCA1 methylation, by low BRCA1 messenger RNA, or by a high score on a commercial homologous recombination deficiency assay. A germline mutation predicted platinum sensitivity there and the assays built to capture the same phenotype did not. A substitution design can expose differential sensitivity. The neoadjuvant trials added carboplatin to a full anthracycline and taxane backbone. Every patient still received the drugs a BRCA-associated tumour is already sensitive to, so the platinum increment is measured on top of them. An addition design cannot reveal the same contrast, whatever the germline result says.
The two selection trials in the table were built the other way round and both failed. TBCRC 030 set single-agent platinum against a single-agent taxane and asked whether a deficiency score above a fixed cut-point predicted response to the platinum specifically. It did not22. That trial was sized to test a predictive claim rather than to compare two drugs. A large share of its patients crossed to another preoperative regimen before surgery. The assay could not be scored on all of the rest. INFORM gave germline carriers single-agent platinum against an anthracycline regimen, and the platinum produced no more complete responses than the chemotherapy those patients would have received anyway21. Entry there required a mutation rather than triple-negative disease, so hormone receptor-positive patients sit inside its answer. No trial has selected patients by the assay and kept an assay-negative arm on the same treatment. That is the comparison that would show what the selection bought, and nobody has run it.
If the assay cannot allocate the platinum, the next question is whether it can allocate a PARP inhibitor, and whether that drug can replace the chemotherapy rather than join it. The designs run from addition to substitution to monotherapy, and they get less comparable as they go. BrighTNess added veliparib to a carboplatin and taxane backbone and it added nothing23. Veliparib traps the enzyme on DNA weakly, so that null belongs to one molecule rather than to the class. GeparOLA put olaparib where the carboplatin would have gone, in patients selected by a deficiency score or a BRCA mutation. It tested its own arm against a fixed threshold rather than against its comparator, and it did not clear the threshold31,32. NEOTALA gave talazoparib alone for a full course in germline carriers and fell short of its prespecified bar, on an evaluable denominator rather than on everyone enrolled33. PETREMAC gave olaparib as a short window before planned chemotherapy in tumours unselected for a repair defect, and many of them responded34. A window followed by planned chemotherapy says nothing about whether that chemotherapy could have been left out. TBCRC 056 pairs a PARP inhibitor with a checkpoint inhibitor and no chemotherapy at all. It carries no PARP-only arm, so it will not separate the contributions of the two drugs.
The practical position that survives all of this is short. Germline testing changes management here and somatic repair assays do not.
Caution
Carboplatin is already inside the regimen that most stage II and III patients now receive35. No trial has removed it from a chemoimmunotherapy backbone, so its contribution in that context is unmeasured. The question is not whether to add platinum. It is whether the platinum that is already there is doing anything, and that trial does not exist.
4 · Checkpoint blockade and what KEYNOTE-522 settled
Choosing what to add to preoperative chemotherapy, and what to take out, continued
Does adding a checkpoint inhibitor to preoperative chemotherapy lower recurrence, and does it have to be given before surgery?
Rationale
Triple-negative tumours carry more tumour-infiltrating lymphocytes and more neoantigen than the other subtypes. Chemotherapy releases antigen as it kills. A checkpoint inhibitor given while the primary tumour and its draining nodes are still in place should prime a wider response than one given to a patient who has neither left.
Experiment
Several randomised additions of a checkpoint inhibitor to preoperative chemotherapy. They differ in the drug, in whether the backbone carried a platinum, and in whether the checkpoint inhibitor continued after surgery. Set against them, two trials gave a checkpoint inhibitor only after surgery, with no preoperative phase at all. A further trial is randomising the omission of the post-surgical half in patients whose tumours cleared.
Finding
Adding a checkpoint inhibitor raised pathological complete response over the same chemotherapy with placebo in most of these trials. In the trial that also continued it after surgery, the event-free survival advantage held. A large trial of a different checkpoint inhibitor, given across both halves of the pathway, missed its primary endpoint. Both trials that waited until after surgery were negative on their primary endpoints.
Limitation
The practice-defining trial varies more than the checkpoint inhibitor. Its backbone carries a platinum. Its treatment continues after surgery. Neither of those was randomised separately. The trial cannot say which part is load-bearing. The trial that ran no post-surgical phase reported a survival separation its own response endpoint did not support, on few events. Checkpoint ligand expression did not select here the way it does in metastatic disease. The co-primary analysis in the marker-positive population did not cross its significance boundary. The assay and its cut-point differ between these trials in any case. The adjuvant-only trials differ in drug and in population as well as in timing. The case for the preoperative window is therefore drawn between trials rather than inside one. Whether the post-surgical half is needed at all is only now being randomised. That trial has not reported.
Next question
If the parts of the package were never randomised against each other, the first one to take out is the cytotoxic backbone. How much chemotherapy still has to sit underneath the checkpoint inhibitor?
high-risk stage II-III HER2-negative disease; 29 of 69 in the experimental arm triple-negative
pembrolizumab added to weekly paclitaxel then doxorubicin and cyclophosphamide vs the same chemotherapy, adaptively randomised against a shared control
pathological complete response, with Bayesian graduation
estimated pCR 60% with pembrolizumab vs 22% in triple-negative disease, 44% vs 17% across HER2-negative disease
stage II-III TNBC, enrolled without a PD-L1 criterion
pembrolizumab with neoadjuvant chemotherapy and then adjuvant pembrolizumab vs placebo with the same chemotherapy and then adjuvant placebo
pathological complete response and event-free survival
pCR 64.8% vs 51.2%, difference 13.6 points (5.4-21.8). 3-year EFS 84.5% vs 76.8%, HR 0.63 (0.48-0.82)
5-year OS 86.6% vs 81.7%, P=0.002, at median follow-up 75.1 months
SCARLET3 · ongoing
TNBC stage I-III
carboplatin, docetaxel and pembrolizumab vs the anthracycline-containing KEYNOTE-522 regimen
event-free survival
not yet reported
not yet reported
KEYNOTE-522 settled one thing completely and left several others open. Adding pembrolizumab to neoadjuvant chemotherapy and continuing it after surgery improves overall survival in stage II and III triple-negative disease, having first improved event-free survival37,38. The trial carries a response endpoint and a time-to-event endpoint inside one randomisation, and its control arm is the chemotherapy the clinic would otherwise give. No other row in this chapter has all three properties. An overall survival benefit in a curable population is rare, which is why the regimen is standard rather than optional. It is also the only row in that table with a randomised overall survival comparison.
Four things it did not settle are worth stating precisely, because each is regularly treated as though it had been.
The other three rows are variations on the regimen rather than tests of it. I-SPY2 pembrolizumab arm graduated pembrolizumab on a Bayesian estimate of success in a future phase III trial36. The prediction proved right, which is a point in favour of the platform design rather than evidence about the drug. A graduation is a probability statement about a trial nobody has run Biomarker-selected, basket, umbrella, and platform trials. NeoPACT and SCARLET keep the antibody and change what sits beneath it. Both put the backbone question of Anthracycline and taxane backbones under the practice-defining drug. Only SCARLET is randomised against an outcome, and it has not reported.
Three of the four rows give carboplatin alongside the taxane, and none of them randomised it. The contribution of either drug to the other therefore stays unmeasured. The fourth row, I-SPY2 pembrolizumab arm, gave pembrolizumab on a taxane and an anthracycline with no platinum at all, in HER2-negative disease reported by subtype36. A response rate from a single arm or from a platform subgroup is not a statement about the antibody alone.
5 · The other immunotherapy trials and why they read out differently
Choosing what to add to preoperative chemotherapy, and what to take out, continued
How much chemotherapy still has to sit underneath the checkpoint inhibitor?
Rationale
The cytotoxic part of the regimen carries the cardiac risk, the leukaemia risk and most of the acute toxicity. If the platinum and the immune agent are doing the work, the anthracycline is a cost with nothing on the other side of it. Taken further, the same argument asks whether any chemotherapy is needed in a tumour that already has an immune infiltrate.
Experiment
One randomised comparison set an anthracycline-free platinum and taxane regimen against the same drugs followed by an anthracycline, with response as the endpoint. A single-arm study gave that short regimen with a checkpoint inhibitor added. One trial kept the anthracycline but moved it after surgery, and randomised the checkpoint inhibitor on what was left in front. A randomised trial of the short regimen against the standard one, with event-free survival as its endpoint, is running. At the far end, a chemotherapy-free trial gave checkpoint blockade alone to tumours selected by a lymphocyte cut-point.
Finding
Anthracycline-free regimens reach response rates close to those of the anthracycline-containing ones. Where the anthracycline was deferred until after surgery, the checkpoint inhibitor's response advantage did not appear. Checkpoint blockade with no chemotherapy at all cleared some lymphocyte-rich tumours completely. It did so at the price of frequent immune endocrinopathy in patients who were otherwise well.
Limitation
All of these read out at the operation. The one trial randomising the two regimens against an outcome has not reported. The anthracycline-free comparison is small and was powered for response, which is the endpoint already shown in this cell to move without survival moving with it. The version that added the checkpoint inhibitor to that short regimen is single-arm. Its response rate has no concurrent standard-regimen control to be set against. The trial that deferred the anthracycline has published only a secondary response endpoint. It cannot be used to argue that the anthracycline was the missing ingredient. The chemotherapy-free trial ran window cohorts whose endpoint is a laboratory measure of immune activation. Only its selected surgical cohort carried a response endpoint. A lymphocyte score read on one slide against a percentage cut-point is a scoring convention rather than a property of the tumour. That trial selected on it.
Next question
If response at the operation cannot license taking a drug out, what can a preoperative trial settle at all, and what has to wait for a randomised outcome trial?
NeoSTOPNeoPACTNeoTRIPSCARLETBELLINI
Checkpoint inhibitor trials in early triple-negative disease, all but one randomised, including the two that gave the drug only after surgery.
atezolizumab added to carboplatin and nab-paclitaxel vs carboplatin and nab-paclitaxel, both followed by adjuvant anthracycline
event-free survival. The pathological complete response tabulated here is a secondary endpoint
pathological complete response 48.6% with atezolizumab added to carboplatin and nab-paclitaxel against 44.4% without, odds ratio 1.18 (0.74-1.89), P=0.48
high-risk stage II-III HER2-negative disease; 29 of 69 in the experimental arm triple-negative
pembrolizumab added to weekly paclitaxel then doxorubicin and cyclophosphamide vs the same chemotherapy, adaptively randomised against a shared control
pathological complete response, with Bayesian graduation
estimated pCR 60% with pembrolizumab vs 22% in triple-negative disease, 44% vs 17% across HER2-negative disease
durvalumab added to nab-paclitaxel then dose-dense epirubicin and cyclophosphamide vs placebo, with no durvalumab after surgery
pathological complete response
pCR 53.4% vs 44.2%, P=0.287. Primary endpoint not met
SCARLET3 · ongoing
TNBC stage I-III
carboplatin, docetaxel and pembrolizumab vs the anthracycline-containing KEYNOTE-522 regimen
event-free survival
not yet reported
One pattern in that table is consistent. Every trial that started the checkpoint inhibitor before surgery and has reported a time-to-event comparison puts its hazard ratio below one, whether or not the endpoint was met. The two that gave the drug only after surgery, ALEXANDRA/IMpassion03041 and A-BRAVE40, both missed their primary endpoints. The case for treating while the primary tumour and its draining nodes are still in place is therefore drawn between trials rather than inside one. Nobody randomised the timing.
Beyond that, reading the spread as a statement about checkpoint blockade is the standard error. The trials differ on five axes at once, and no axis was isolated in any of them.
The agent differs. Pembrolizumab binds PD-1, while atezolizumab, durvalumab and avelumab bind PD-L1 and leave the PD-L2 interaction intact.
The chemotherapy backbone differs. KEYNOTE-522 and NSABP B-59/GeparDouze used a platinum and an anthracycline39, while IMpassion031 used an anthracycline without a platinum. NeoTRIP used a platinum with the anthracycline deferred until after surgery. NeoPACT and SCARLET drop the anthracycline altogether.
The design differs. I-SPY2 pembrolizumab arm was not a triple-negative trial at all. It adaptively randomised HER2-negative disease against a shared control and reported this subtype as a subgroup36. NeoPACT has no control arm16. SCARLET randomises the backbone rather than the antibody.
The endpoint each trial was powered for differs. IMpassion031 and GeparNuevo were powered for pathological complete response, while NSABP B-59/GeparDouze and NeoTRIP were powered for event-free survival, which is a much harder endpoint to move.
The enrolled population differs. NeoTRIP took high-risk and locally advanced disease, and in a pooled reconstruction of five of these trials its four-year recurrence rate was the highest44.
The column this table does not carry is overall survival, and three rows change character when it is added. GeparNuevo missed on response and separated on survival, and the separation held at long follow-up43,4,7. IMpassion031 hit its response endpoint while its survival comparisons stayed descriptive42,45. A-BRAVE missed both co-primary disease-free survival endpoints while its descriptive overall survival estimate favoured avelumab40. Descriptive means the trial bought no alpha for it. None of these three is a survival result in the sense KEYNOTE-522 produced one. The dissociation between response and survival runs through all of them. That is the trial-level surrogacy failure of Endpoints, surrogates, and the overall survival question appearing inside a single drug class.
The anthracycline-free question sits in this section rather than in Anthracycline and taxane backbones, because every trial that asks it also varies the checkpoint inhibitor. NeoPACT reported a response rate in a single arm with no randomised comparator16. NeoTRIP deferred the anthracycline until after surgery, and the checkpoint inhibitor's response advantage did not appear at all17. That is one trial, one endpoint and one backbone. The published response result is a secondary endpoint, so it cannot be read as showing the anthracycline was the missing ingredient. SCARLET is the randomised test against the KEYNOTE-522 regimen. Until it reports, removing the anthracycline remains a hypothesis rather than an option.
BELLINI takes the same argument to its end and gives no chemotherapy at all. Checkpoint blockade alone cleared some tumours selected for a high stromal lymphocyte density46. Two things bound what that can be used for. Its window cohorts read out on a laboratory measure of immune activation rather than on response, and only its selected surgical cohort carried a response endpoint. A stromal lymphocyte percentage read on one slide against a fixed cut-point is a scoring convention rather than a property of the tumour Tumor-infiltrating lymphocytes, biology and standardized scoring. That trial selected on the convention.
Caution
Comparing pathological complete response rates across these trials is not informative. The populations, the backbones and the response definitions were not held constant. A lower rate in one trial than another is at least as likely to reflect who was enrolled as what was given.
6 · Biomarkers for immunotherapy benefit, and why PD-L1 fails here
PD-L1 selects in metastatic triple-negative disease and does not select in early disease. Both statements are correct, and the biology does not have to reverse between the two settings for that to be so. The metastatic side of the argument is in First-line chemoimmunotherapy and the PD-L1 threshold and the trials that carry it are tabulated there rather than here. Six differences separate the settings, each cumulative with the others rather than competing.
The first is that PD-L1 positive is not the output of one assay. In advanced disease, samples scored by three antibodies gave immune cell positivity of 46.4% with SP142, 74.9% with SP263 and 73.1% with 22C3, and 80.9% were positive by 22C3 combined positive score. Analytical concordance with SP142 was about 69%, and no inter-assay equivalence was found47. KEYNOTE-522 used the combined positive score, while IMpassion031 and NeoTRIP used the SP142 immune cell score. Those instruments count different cells and divide by different denominators Predictive biomarkers and patient selection. Part of what is read as a setting effect is an assay effect Immunotherapy.
The second is that the tissue being scored is not the same tissue. Early disease is scored on a treatment-naive core of an intact primary. Advanced disease is often scored on an archival primary or on a metastasis. Immune cell positivity was 63.7% in primary tumours and 42.2% in unmatched metastatic lesions in one series scored with the SP142 assay48. Lymphocyte counts, PD-L1 and a coordinated set of interferon-regulated and antigen presentation genes are all lower in metastases49.
The third follows from the second, because the dynamic range of the marker differs. The early population is shifted towards the positive end of the scale, so its negative group is both smaller and less extreme. A fixed cut-point applied to a compressed distribution separates less, whatever the underlying biology does.
The fourth is the combination partner the checkpoint inhibitor is added to. In the early setting it joins a curative multi-agent regimen containing a platinum. Cytotoxic therapy changes the infiltrate and PD-L1 expression itself50, and DNA damage activates innate sensing51. A strong combination partner compresses the difference between subgroups defined by a marker, because it supplies to the negative group some of what the marker was selecting for.
The fifth is that the two settings ask the marker to predict different things. Here it is asked to predict eradication at one operation, and then recurrence over a few years in patients treated for cure. In advanced disease it is asked to predict progression and death in disease that will not be cured. A marker can track one and fail the other without being wrong about either Endpoints, surrogates, and the overall survival question.
The sixth is that the statistical claim is weaker than it sounds. KEYNOTE-522 was not designed to test a PD-L1 interaction, and its PD-L1 negative subgroup was small35. Benefit regardless of PD-L1 means that no interaction was demonstrated. It does not mean that no interaction exists.
Within the early setting the marker has not been inert either. The PD-L1 positive co-primary population of IMpassion031 did not cross its significance boundary42. In NeoTRIP, PD-L1 was the strongest factor associated with response in multivariable analysis, with an odds ratio of 2.08, in a trial where the drug itself did not improve response17. A marker associated with response and not with the treatment effect is prognostic rather than predictive, a distinction defined in Analytical validity, clinical validity, and clinical utility defined.
Three other candidates are worth tracking and none is ready.
Tumour-infiltrating lymphocytes are prognostic and reproducibly scored under the convention in Tumor-infiltrating lymphocytes, biology and standardized scoring6,52. The first randomised signal that they might also be predictive of checkpoint benefit comes from A-BRAVE, where avelumab improved outcome only above a baseline stromal score of 30%53. That is a biomarker analysis of a trial that missed its primary endpoint, and the score is a percentage read on one slide rather than a property of the tumour. Residual cancer burden was not predictive in the same analysis.
Tumour-specific class II major histocompatibility expression was associated with response to atezolizumab and not to chemotherapy alone in NeoTRIP54. It requires quantitative imaging mass cytometry that no routine laboratory runs.
Tumour mutational burden carries tumour-agnostic evidence from studies that had no breast cohort55. The one breast cohort since reported was in metastatic disease56. High burden is also uncommon in this disease57, so the eligible population is small before the question of whether the inference transfers is reached.
One exploratory signal is worth naming without overreading. The subgroup analyses of NSABP B-59/GeparDouze pointed to benefit in patients presenting with clinical lymph node involvement39. That is a subgroup of a trial that missed its primary endpoint, and nodal status is a stage variable rather than an assay. It cannot serve as the marker this section is looking for.
Caution
Do not carry the metastatic selection rule backwards across surgery. A patient with a combined positive score below 10 is not a patient for whom neoadjuvant pembrolizumab is unproven, because the trial that established the regimen did not select on that score at all35. The reverse carry is also wrong, and First-line chemoimmunotherapy and the PD-L1 threshold states it from the other side.
Interplay
The reason a percentage of PD-L1 staining fails to predict is the reason given in Part VI for most biomarker failure in this disease. What a checkpoint inhibitor releases is a relationship between cell types, not a property of one cell. A score computed on a block reports abundance and discards arrangement, exactly as a HER2 percentage does in Spatial patterns, clustered, mosaic, and scattered, and the spatial phenotypes in Inflamed, excluded, and desert phenotypes separate patients that no abundance measure can. There is a second point specific to the neoadjuvant setting. Treatment is being given to an intact primary and its draining nodes, which is where priming happens Systemic immunity, lymph node priming, and immune memory. A marker measured on a core taken before any of that describes the starting state of a system whose behaviour depends on what the first two cycles do to it.
7 · De-escalation after pathological complete response and the immune toxicity cost
After a pathological complete response the residual risk is low and it is front-loaded. In a pooled reconstruction of five neoadjuvant chemoimmunotherapy trials, most recurrences occurred within 24 months of randomisation, in patients with and without a complete response44. A de-escalation decision taken after surgery is therefore a decision about the next two years rather than the next ten.
The only randomised test of omitting the adjuvant checkpoint phase after a complete response.
Trial
Population
Experimental vs control
OptimICE-pCR3 · ongoing
TNBC with pathologic complete response after chemoimmunotherapy
omission of adjuvant pembrolizumab vs adjuvant pembrolizumab, after pathological complete response
That single row is the entire randomised record for this question, and it has not reported. It compares omission against continuation in patients who already cleared, so the verdict it can return is non-inferiority and nothing else. It will not say what the adjuvant phase is worth in a patient with residual disease, which is a different population treated under Adjuvant and post-neoadjuvant therapy.
The cost side of the trade is specific to this setting. More than half the patients in the control arm of KEYNOTE-522 reached a complete response without pembrolizumab35. A permanent endocrinopathy in one of those patients is a different event from the same endocrinopathy in incurable disease. The structural features that make these toxicities unlike chemotherapy toxicity are in Immune-related toxicity and its management and Immune-related toxicity in a curative-intent population.
Two arguments support omission and neither of them answers the question. GeparNuevo separated on overall survival with no adjuvant checkpoint phase at all4. That was a secondary endpoint in a trial that missed its primary one. And a complete response identifies a group whose absolute risk leaves little room for an adjuvant drug to work.
The obstacle is arithmetic rather than conceptual. Three-year event-free survival after a complete response on this regimen is high, so a non-inferiority trial needs either a very large sample or a margin wide enough to permit real harm. That is the structure of the trastuzumab duration trials in The trastuzumab adjuvant trials and the duration question, where near-identical estimates produced opposite verdicts because the margins differed. Designing de-escalation trials and the non-inferiority margin problem sets out what to demand of the answer when it arrives.
In practice
Do not stop adjuvant pembrolizumab early on the grounds that the tumour responded. That is the question OptimICE-pCR is asking, and it has not answered it.
Stopping it is defensible when a significant immune-related event occurs after a complete response. The remaining absolute benefit is small and unproven. The toxicity in front of you is neither.
stage I-III TNBC with residual invasive disease after neoadjuvant therapy
datopotamab deruxtecan with durvalumab, or datopotamab deruxtecan alone, vs investigator's choice of capecitabine, pembrolizumab or both
invasive disease-free survival
Six trials in this cell have not reported a primary endpoint, and only two of them will change what is given before the operation.
SCARLET asks whether the anthracycline can be removed from a pembrolizumab and platinum regimen. A non-inferior result would take an anthracycline, and with it the cardiac and leukaemic risks, out of a young population treated with curative intent. It is the only randomised outcome trial of that question anywhere in this chapter.
NeoTRIP has yet to report the event-free survival endpoint it was powered for17. Its response result, published as a secondary endpoint, was negative. A positive survival result would add a second example of the dissociation that GeparNuevo produced, and would strengthen the case that response is the wrong readout for this drug class.
The other three sit after surgery, and they are not one experiment repeated. ASCENT-05/OptimICE-RD gives pembrolizumab in both arms and asks what a conjugate adds to it, against a control that may or may not include capecitabine11. TROPION-Breast03 randomises a different conjugate with or without an anti-PD-L1 antibody, against a control the investigator chooses from capecitabine, pembrolizumab or both12. A control arm selected by the investigator is honest about practice and imprecise about the drug. SASCIA / GBG 102 asks the same question across HER2-negative disease rather than triple-negative disease alone, so hormone receptor-positive patients sit inside its answer. All three belong to Adjuvant and post-neoadjuvant therapy, and they appear here because a preoperative regimen that raises response decides how many patients reach them.
Four questions in this chapter have no trial running at all.
Nobody is removing carboplatin from the chemoimmunotherapy backbone, so its contribution there stays unmeasured.
No trial randomises a checkpoint inhibitor in stage I disease, where the regimen is increasingly used by extrapolation. SCARLET enrols stage I patients and gives pembrolizumab in both arms.
And no trial sets one checkpoint inhibitor against another. Every comparison a clinician makes between KEYNOTE-522 and the atezolizumab and durvalumab regimens is a cross-trial comparison, with all the confounding set out in The other immunotherapy trials and why they read out differently.
References
All 58 references
Cortazar P, Zhang L, Untch M, et al. Pathological complete response and long-term clinical benefit in breast cancer: the CTNeoBC pooled analysis. Lancet 2014 384:164-172. PMID 24529560
Shepherd JH, Ballman K, Polley MC, et al. CALGB 40603 (Alliance) - long-term outcomes and genomic correlates of response and survival after neoadjuvant chemotherapy with or without carboplatin and bevacizumab in triple-negative breast cancer. J Clin Oncol 2022 40:1323-1334. PMID 35044810
Gluz O, Nitz U, Kolberg-Liedtke C, et al. De-escalated Neoadjuvant Chemotherapy in Early Triple-Negative Breast Cancer (TNBC): Impact of Molecular Markers and Final Survival Analysis of the WSG-ADAPT-TN Trial. Clin Cancer Res 2022 28:4995-5003. PMID 35797219
Loibl S, Schneeweiss A, Huober J, et al. Neoadjuvant durvalumab improves survival in early triple-negative breast cancer independent of pathological complete response. Ann Oncol 2022 33:1149-1158. PMID 35961599
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
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
Loibl S, Untch M, Huober J, et al. Durvalumab in combination with neoadjuvant chemotherapy in early triple-negative breast cancer: long-term analysis from the GeparNuevo trial. J Clin Oncol 2026 44:1580-1586. PMID 42008768
I-SPY2 Trial Consortium, Yee D, DeMichele AM, et al. Association of event-free and distant recurrence-free survival with individual-level pathologic complete response in neoadjuvant treatment of stages 2 and 3 breast cancer: three-year follow-up analysis for the I-SPY2 adaptively randomized clinical trial. JAMA Oncol 2020 6:1355-1362. PMID 32701140
Spring LM, Tolaney SM, Fell G, et al. Response-guided neoadjuvant sacituzumab govitecan for localized triple-negative breast cancer: results from the NeoSTAR trial. Ann Oncol 2023 35:293-301. PMID 38092228
Lynce F, Stevens LE, Li Z, et al. TBCRC 039: a phase II study of preoperative ruxolitinib with or without paclitaxel for triple-negative inflammatory breast cancer. Breast Cancer Res 2024 26:20. PMID 38297352
Tolaney SM, DeMichele A, Takano T, et al. OptimICE-RD - sacituzumab govitecan plus pembrolizumab vs pembrolizumab (with or without capecitabine) for residual triple-negative breast cancer. Future Oncol 2024 20:2343-2355. PMID 38922307
Bardia A, Pusztai L, Albain K, et al. TROPION-Breast03 - a randomized phase III global trial of datopotamab deruxtecan with or without durvalumab in patients with triple-negative breast cancer and residual invasive disease at surgical resection after neoadjuvant therapy. Ther Adv Med Oncol 2024 16:17588359241248336. PMID 38686016
Early Breast Cancer Trialists' Collaborative Group (EBCTCG). Long-term outcomes for neoadjuvant versus adjuvant chemotherapy in early breast cancer: meta-analysis of individual patient data from ten randomised trials. Lancet Oncol 2018 19:27-39. PMID 29242041
Untch M, et al. Nab-paclitaxel versus solvent-based paclitaxel in neoadjuvant chemotherapy for early breast cancer (GeparSepto-GBG 69): a randomised, phase 3 trial. Lancet Oncol 2016 17:345-356. PMID 26869049
Gianni L, et al. Comparing Neoadjuvant Nab-paclitaxel vs Paclitaxel Both Followed by Anthracycline Regimens in Women With ERBB2/HER2-Negative Breast Cancer-The Evaluating Treatment With Neoadjuvant Abraxane (ETNA) Trial: A Randomized Phase 3 Clinical Trial. JAMA Oncol 2018 4:302-308. PMID 29327055
Sharma P, Stecklein SR, Yoder R, et al. Clinical and biomarker findings of neoadjuvant pembrolizumab and carboplatin plus docetaxel in triple-negative breast cancer: NeoPACT phase 2 clinical trial. JAMA Oncol 2024 10:227-235. PMID 37991778
Gianni L, Huang CS, Egle D, et al. Pathologic complete response (pCR) to neoadjuvant treatment with or without atezolizumab in triple-negative, early high-risk and locally advanced breast cancer - NeoTRIP Michelangelo randomized study. Ann Oncol 2022 33:534-543. PMID 35182721
Sharma P, et al. Randomized Phase II Trial of Anthracycline-free and Anthracycline-containing Neoadjuvant Carboplatin Chemotherapy Regimens in Stage I-III Triple-negative Breast Cancer (NeoSTOP). Clin Cancer Res 2020 27:975-982. PMID 33208340
Gluz O, Nitz U, Liedtke C, et al. Comparison of Neoadjuvant Nab-Paclitaxel+Carboplatin vs Nab-Paclitaxel+Gemcitabine in Triple-Negative Breast Cancer: Randomized WSG-ADAPT-TN Trial Results. J Natl Cancer Inst 2018 110:628-637. PMID 29228315
Kim GM, Jung KH, Jeung HC, et al. A randomized phase III trial of anthracyclines followed by taxane versus taxane plus carboplatin as (neo)adjuvant therapy in patients with triple-negative breast cancer - KCSG BR 15-1 PEARLY trial. Ann Oncol 2026. PMID 42218963
Tung N, Arun B, Hacker MR, et al. TBCRC 031: Randomized Phase II Study of Neoadjuvant Cisplatin Versus Doxorubicin-Cyclophosphamide in Germline BRCA Carriers With HER2-Negative Breast Cancer (the INFORM trial). J Clin Oncol 2020 38:1539-1548. PMID 32097092
Mayer EL, Abramson V, Jankowitz R, et al. TBCRC 030: a phase II study of preoperative cisplatin versus paclitaxel in triple-negative breast cancer: evaluating the homologous recombination deficiency (HRD) biomarker. Ann Oncol 2020 31:1518-1525. PMID 32798689
Loibl S, O'Shaughnessy J, Untch M, et al. Addition of the PARP inhibitor veliparib plus carboplatin or carboplatin alone to standard neoadjuvant chemotherapy in triple-negative breast cancer (BrighTNess) - a randomised, phase 3 trial. Lancet Oncol 2018 19:497-509. PMID 29501363
Sikov WM, Berry DA, Perou CM, et al. Impact of the addition of carboplatin and/or bevacizumab to neoadjuvant once-per-week paclitaxel followed by dose-dense doxorubicin and cyclophosphamide on pathologic complete response rates in stage II to III triple-negative breast cancer - CALGB 40603 (Alliance). J Clin Oncol 2015 33:13-21. PMID 25092775
von Minckwitz G, Schneeweiss A, Loibl S, et al. Neoadjuvant carboplatin in patients with triple-negative and HER2-positive early breast cancer (GeparSixto; GBG 66) - a randomised phase 2 trial. Lancet Oncol 2014 15:747-756. PMID 24794243
Geyer CE, Sikov WM, Huober J, et al. Long-term efficacy and safety of addition of carboplatin with or without veliparib to standard neoadjuvant chemotherapy in triple-negative breast cancer - 4-year follow-up data from BrighTNess, a randomized phase III trial. Ann Oncol 2022 33:384-394. PMID 35093516
Loibl S, Weber KE, Timms KM, et al. Survival analysis of carboplatin added to an anthracycline/taxane-based neoadjuvant chemotherapy and HRD score as predictor of response - final results from GeparSixto. Ann Oncol 2018 29:2341-2347. PMID 30335131
Liu Y, Gong Y, Zhu XZ, et al. Effect of adjuvant carboplatin intensified chemotherapy versus standard chemotherapy on survival in women with high risk, early stage, triple negative breast cancer (CITRINE) - randomised, open label phase 3 trial. BMJ 2025 391:e085457. PMID 41436191
Bear HD, et al. Sequential preoperative or postoperative docetaxel added to preoperative doxorubicin plus cyclophosphamide for operable breast cancer:National Surgical Adjuvant Breast and Bowel Project Protocol B-27. J Clin Oncol 2006 24:2019-27. PMID 16606972
Tutt A, Tovey H, Cheang MCU, et al. Carboplatin in BRCA1/2-mutated and triple-negative breast cancer BRCAness subgroups: the TNT trial. Nat Med 2018 24:628-637. PMID 29713086
Fasching PA, Link T, Hauke J, et al. Neoadjuvant paclitaxel/olaparib in comparison to paclitaxel/carboplatinum in patients with HER2-negative breast cancer and homologous recombination deficiency (GeparOLA study). Ann Oncol 2020 32:49-57. PMID 33098995
Fasching PA, Schmatloch S, Hauke J, et al. Neoadjuvant Paclitaxel/Olaparib in Comparison to Paclitaxel/Carboplatin in Patients with HER2-Negative Breast Cancer and HRD-Long-term Survival of the GeparOLA Study. Clin Cancer Res 2025 31:1596-1604. PMID 39998850
Litton JK, Beck JT, Jones JM, et al. Neoadjuvant Talazoparib in Patients With Germline BRCA1/2 Mutation-Positive, Early-Stage Triple-Negative Breast Cancer: Results of a Phase II Study. Oncologist 2023 28:845-855. PMID 37318349
Eikesdal HP, Yndestad S, Elzawahry A, et al. Olaparib monotherapy as primary treatment in unselected triple negative breast cancer. Ann Oncol 2020 32:240-249. PMID 33242536
Schmid P, Cortes J, Pusztai L, et al. Pembrolizumab for early triple-negative breast cancer. N Engl J Med 2020 382:810-821. PMID 32101663
Nanda R, Liu MC, Yau C, et al. Effect of pembrolizumab plus neoadjuvant chemotherapy on pathologic complete response in women with early-stage breast cancer - an analysis of the ongoing phase 2 adaptively randomized I-SPY2 trial. JAMA Oncol 2020 6:676-684. PMID 32053137
Schmid P, Cortes J, Dent R, et al. Event-free survival with pembrolizumab in early triple-negative breast cancer. N Engl J Med 2022 386:556-567. PMID 35139274
Schmid P, Cortes J, Dent R, et al. Overall survival with pembrolizumab in early-stage triple-negative breast cancer. N Engl J Med 2024 391:1981-1991. PMID 39282906
Loibl S, Tang G, Nekljudova V, et al. Atezolizumab in early triple-negative breast cancer: the randomized phase 3 NSABP B-59/GeparDouze trial. Nat Med 2026. PMID 42680950
Conte PF, Dieci MV, Bisagni G, et al. A-BRAVE trial - a phase III randomized trial with anti-PD-L1 avelumab in high-risk triple-negative early breast cancer patients. Ann Oncol 2025 36:1492-1502. PMID 40885530
Ignatiadis M, Bailey A, McArthur H, et al. Adjuvant atezolizumab for early triple-negative breast cancer: the ALEXANDRA/IMpassion030 randomized clinical trial. JAMA 2025 333:1150-1160. PMID 39883436
Mittendorf EA, Zhang H, Barrios CH, et al. Neoadjuvant atezolizumab in combination with sequential nab-paclitaxel and anthracycline-based chemotherapy versus placebo and chemotherapy in patients with early-stage triple-negative breast cancer (IMpassion031): a randomised, double-blind, phase 3 trial. Lancet 2020 396:1090-1100. PMID 32966830
Loibl S, Untch M, Burchardi N, et al. A randomised phase II study investigating durvalumab in addition to an anthracycline taxane-based neoadjuvant therapy in early triple-negative breast cancer: clinical results and biomarker analysis of GeparNuevo study. Ann Oncol 2019 30:1279-1288. PMID 31095287
Licata L, Mariani M, Viale G, et al. Timing of recurrence after neoadjuvant chemo-immunotherapy in patients with early-stage triple-negative breast cancer. Clin Cancer Res 2025 31:3916-3921. PMID 40689866
Mittendorf EA, Assaf ZJ, Harbeck N, et al. Peri-operative atezolizumab in early-stage triple-negative breast cancer - final results and ctDNA analyses from the randomized phase 3 IMpassion031 trial. Nat Med 2025 31:2397-2404. PMID 40467898
Nederlof I, Isaeva OI, de Graaf M, et al. Neoadjuvant nivolumab or nivolumab plus ipilimumab in early-stage triple-negative breast cancer: a phase 2 adaptive trial. Nat Med 2024 30:3223-3235. PMID 39284953
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
Rozenblit M, Huang R, Danziger N, et al. Comparison of PD-L1 protein expression between primary tumors and metastatic lesions in triple negative breast cancers. J Immunother Cancer 2020 8. PMID 33239417
Szekely B, Bossuyt V, Li X, et al. Immunological differences between primary and metastatic breast cancer. Ann Oncol 2018 29:2232-2239. PMID 30203045
Pelekanou V, Carvajal-Hausdorf DE, Altan M, et al. Effect of neoadjuvant chemotherapy on tumor-infiltrating lymphocytes and PD-L1 expression in breast cancer and its clinical significance. Breast Cancer Res 2017 19:91. PMID 28784153
Parkes EE, Walker SM, Taggart LE, et al. Activation of STING-dependent innate immune signaling by S-phase-specific DNA damage in breast cancer. J Natl Cancer Inst 2017 109:djw199. PMID 27707838
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
Dieci MV, Gasparini E, Nicole L, et al. Tumor infiltrating lymphocytes as a predictor of adjuvant avelumab efficacy in patients with high-risk triple negative breast cancer. Clin Cancer Res 2026. PMID 42467210
Balko JM, Licata L, Wang XQ, et al. Tumor-specific major histocompatibility-II expression predicts pathological complete response to atezolizumab combined to chemotherapy in triple-negative breast cancer. NPJ Breast Cancer 2025 11:103. PMID 41022779
Marabelle A, Fakih M, Lopez J, et al. Association of tumour mutational burden with outcomes in patients with advanced solid tumours treated with pembrolizumab: prospective biomarker analysis of the multicohort, open-label, phase 2 KEYNOTE-158 study. Lancet Oncol 2020 21:1353-1365. PMID 32919526
Alva AS, Mangat PK, Garrett-Mayer E, et al. Pembrolizumab in Patients With Metastatic Breast Cancer With High Tumor Mutational Burden: Results From the Targeted Agent and Profiling Utilization Registry (TAPUR) Study. J Clin Oncol 2021 39:2443-2451. PMID 33844595
Barroso-Sousa R, Jain E, Cohen O, et al. Prevalence and mutational determinants of high tumor mutation burden in breast cancer. Ann Oncol 2020 31:387-394. PMID 32067680
Mamounas EP, Bandos H, White JR, et al. Omitting regional nodal irradiation after response to neoadjuvant chemotherapy. N Engl J Med 2025 392:2113-2124. PMID 40466065