The 2026 ASTRO Annual Meeting opens in Boston on September 26, and the Clinical Trials session begins the following afternoon with a question that has become considerably more interesting over the past year: can the predictive PAM50 signal observed in NRG-GU006/BALANCE be reproduced in an independent randomized salvage radiotherapy dataset?
Shuang Zhao of MD Anderson Cancer Center will present Evaluation of PAM50 Molecular Subtypes in Post-Prostatectomy Salvage Radiotherapy: An NRG Oncology-RTOG 0534 Analysis on Sunday, September 27 at 1:30 pm ET in the Grand Ballroom, as the opening presentation of CT 01.
Why SPPORT is the right trial for this question
NRG Oncology/RTOG 0534, better known as SPPORT, enrolled men with a persistently detectable or rising PSA after prostatectomy and randomized them to three strategies: prostate-bed radiotherapy alone, prostate-bed radiotherapy plus 4 to 6 months of short-term androgen deprivation, or pelvic nodal radiotherapy plus prostate-bed radiotherapy plus short-term androgen deprivation. At a median follow-up of 8.2 years, five-year freedom from progression was 70.9%, 81.3%, and 87.4% across the three arms. SPPORT established the benefit of adding short-term androgen deprivation to salvage radiotherapy and provided the only randomized evidence supporting further intensification with elective pelvic nodal irradiation.
For a biomarker analysis, the importance of SPPORT is not that the trial was positive. It is that treatment was randomized. Most genomic classifiers in prostate cancer have been developed and evaluated in observational cohorts, where treatment selection is confounded by clinical risk in ways no amount of matching removes. SPPORT provides something rarer: a setting in which treatment effect can be examined within a randomized comparison.
What PAM50 does in prostate cancer
PAM50 was developed in breast cancer, and its application to prostate cancer was established in a 2017 analysis led by Zhao. Applied to several thousand prostatectomy specimens, the classifier separated prostate cancers into three molecular subtypes: Luminal A, Luminal B, and Basal.
The provocative observation was not that the subtypes had different outcomes. Luminal B tumors appeared to carry the poorest prognosis, but also appeared to derive greater benefit from postoperative androgen deprivation than Luminal A or Basal tumors. That distinction is fundamental. A biomarker that identifies patients with a poor prognosis is useful for estimating risk. A biomarker that identifies patients who derive greater benefit from a particular treatment is useful for selecting treatment. The 2017 study generated the hypothesis that PAM50 might do both, but its treatment comparison was retrospective, which left the central question unresolved.
The randomized signal already exists
The field moved beyond retrospective association at ASTRO 2025 with the presentation of NRG-GU006/BALANCE.
BALANCE was a phase II, double-blinded, placebo-controlled trial that enrolled 295 men between 2018 and 2020 with a PSA of 0.1 to 1.0 ng/mL after prostatectomy and no evidence of nodal or distant disease. Patients received salvage radiotherapy with either apalutamide 240 mg daily for six months or placebo, and were stratified prospectively by PAM50 subtype as Luminal B versus non-Luminal B.
The treatment effect was concentrated almost entirely in Luminal B disease. Among the 127 Luminal B patients, five-year biochemical progression-free survival was 72% with apalutamide versus 54% with placebo, with a hazard ratio of 0.45. Among non-Luminal B patients, the corresponding figures were 70% and 71%, with a hazard ratio of 0.95. Metastasis-free survival followed the same pattern, with five-year estimates of 95% versus 82% in Luminal B and a hazard ratio of 0.27, against no separation at all in non-Luminal B.
That changed the status of PAM50 from a hypothesis-generating classifier to the first prospectively validated predictive biomarker for hormone therapy in prostate cancer.
The ASTRO 2026 SPPORT analysis is therefore not the first randomized test of the hypothesis. It is something arguably more demanding: an opportunity to determine whether the signal reproduces in an independent randomized dataset, in a trial that was not designed to look for it.
Two features make that test harder than BALANCE. First, BALANCE stratified on PAM50 prospectively and was built around the hypothesis. SPPORT was designed and conducted years before this question existed, so the analysis depends on archived tissue and post-hoc subgrouping. A signal that survives those conditions is more impressive than one found in a trial constructed to find it.
Second, and more substantively, the two trials tested different pharmacology. BALANCE used apalutamide as monotherapy, a direct androgen receptor antagonist. SPPORT used conventional LHRH-based short-term androgen deprivation. Whether the PAM50 signal reflects sensitivity to AR-pathway blockade in general, or to potent AR antagonism specifically, has not been tested by anyone. If Luminal B tumors show the same differential benefit from conventional ADT, the biological claim becomes considerably more robust. If they do not, that is arguably the more informative result.
SPPORT may also allow investigators to ask a question BALANCE could not: whether molecular subtype modifies the effect of pelvic nodal radiotherapy.
Prognostic or predictive?
This distinction is where the Sunday presentation will succeed or fail.
A prognostic biomarker tells us how a patient is likely to do. A predictive biomarker tells us whether the effect of treatment differs according to that biomarker. The two are frequently blurred because a strong prognostic marker produces striking Kaplan-Meier separation even when treatment benefit is identical across biomarker groups.
The key evidence is therefore not that Luminal B patients receiving androgen deprivation do better than Luminal B patients who do not. The relevant question is whether the magnitude of treatment benefit differs between Luminal B and non-Luminal B disease, which requires a treatment-by-biomarker interaction. For the hormonal question in SPPORT, the cleanest comparison is Arm 1 versus Arm 2, where the radiotherapy volume is otherwise similar and the principal randomized difference is short-term androgen deprivation. That is the comparison to watch.
Four things to watch
1. Tissue availability. Correlative studies from older cooperative-group trials rarely include the entire randomized population. How many SPPORT participants had evaluable prostatectomy tissue, and whether the molecular cohort resembles the original randomized population in stage, grade, PSA, and margin status, sets the ceiling on how much weight the result can carry. The smaller and more selected the cohort, the more cautious the interpretation.
2. Interaction p-value versus subgroup effect. A forest plot showing a larger treatment effect in Luminal B tumors is not equivalent to demonstrating that treatment effect differs statistically by subtype. The interaction test is the key result, and whether it was prespecified will matter.
3. Which randomized contrast is used. Arm 1 versus Arm 2 is the interpretable comparison for the hormonal question. Pooling Arms 2 and 3 against Arm 1 mixes the effects of androgen deprivation and pelvic nodal irradiation. That may increase statistical power, but it makes the biological interpretation substantially less clean.
4. Whether pelvic nodal radiotherapy is addressed. This may be the most interesting part of the analysis. SPPORT is uniquely positioned to ask whether molecular subtype modifies the benefit of elective pelvic nodal irradiation. If PAM50 predicts differential benefit from nodal treatment as well as hormonal treatment, the implications extend beyond systemic intensification and raise the possibility that molecular biology could eventually inform radiotherapy volume selection itself. That would be genuinely new territory.
What is not on the table
Even a compelling result will require restraint. This remains a retrospective correlative analysis of archived tissue from a trial that completed accrual more than a decade ago, before PSMA PET entered routine salvage staging, using conventional fractionation and with variable surgical nodal assessment. Contemporary salvage practice is not the clinical environment in which SPPORT was conducted.
There is also a practical constraint that gets lost in the enthusiasm. The prostate PAM50 classifier is currently available for research use only, through Veracyte's Decipher GRID platform. It is not an orderable clinical test in the way Decipher itself is, and no positive analysis presented on Sunday changes that.
The biomarker landscape is crowded as well. Decipher retains the broader clinical evidence base among genomic classifiers in prostate cancer and is being incorporated prospectively into treatment assignment in NRG-GU009 and GU010, although those trials concern intact-prostate rather than salvage disease. For PAM50 to become clinically decisive, biological association is not enough. It must add information beyond conventional clinicopathologic factors and existing genomic tools.
What RadOncToday thinks
The design is the story here. Prostate oncology has accumulated many biomarkers that identify patients at higher risk of recurrence, and very few that identify which treatment a patient is more likely to benefit from. BALANCE moved PAM50 into the second category. A single positive randomized study, particularly a phase II with 127 patients in the informative subgroup, is rarely the end of that story.
It is worth being direct about the awkward position the field is now in. On presenting BALANCE, Daniel Spratt argued that the magnitude of benefit in Luminal B, combined with its complete absence in non-Luminal B, made a confirmatory phase III ethically difficult to run. That may well be right. But it means that if a definitive trial is not going to happen, retrospective analyses of existing randomized datasets are not a lesser substitute for validation. They become the only remaining opportunity for the hypothesis to be wrong. Sunday's presentation carries more weight because of that, not less.
Our prediction
We expect PAM50 to retain a strong prognostic association, and the interaction with hormonal therapy to remain directionally consistent with BALANCE and the earlier retrospective data. The most likely scenario is that the biomarker cohort will be substantially smaller than the randomized SPPORT population and that the interaction analysis will be informative rather than definitive.
That would still matter. Independent replication across two randomized datasets, using two different classes of androgen-axis treatment, would considerably strengthen the biological case. What would be more surprising, and more important, is a convincing interaction between PAM50 subtype and pelvic nodal radiotherapy. If that signal emerges, the question shifts from whether molecular subtyping can guide systemic intensification to whether it might help determine the extent of radiotherapy itself. That would justify a very different conversation.
What none of it would justify is routine PAM50 testing in salvage practice on Monday morning. The assay is not clinically available outside a research platform, and even a positive SPPORT interaction remains a retrospective molecular analysis of archived tissue rather than prospective PAM50-directed treatment assignment. A good interaction slide should generate a trial, not a new order set.
We will be covering the presentation from ASTRO 2026.
This is the first article in the RadOncToday series previewing the Clinical Trials and Plenary sessions at ASTRO 2026 over the three weeks before the meeting.