Now, the phase 1b/2 data have finally been published in The New England Journal of Medicine.

A quick recap: daraxonrasib inhibits the active RAS(ON) form. Because even RAS wild-type tumors may depend on RAS activation, this provides the rationale for investigating daraxonrasib beyond RAS-mutant populations.

Response rates: which number matters?

In most early phase data, response rates are not the primary endpoint, which is often safety measures. In RMC-6236-001, this was the case, yet the abstract emphasizes the most favorable response-rate metrics.

“In a subgroup of 26 patients with RAS G12 mutations who were treated with second-line daraxonrasib at a dose of 300 mg, an objective response to therapy was reported in 35% (95% confidence interval [CI], 17 to 56).[…] Among the 38 patients with RAS G12, G13, or Q61 mutations, 29% (95% CI, 15 to 46) had an objective response.”

However, there are many response rates in RMC-6236-001

In small phase 1/2 trials, with highly selected patients, the relevance of breaking down the results according to the line of therapy is not as obvious as in larger phase 3 trials.

The response rate in the overall treated population (24%), together with the breakdown by RAS status (with 27% in G12-mutant population), are both highly relevant, yet not explicitly reported as standalone values by the authors.

Comparison with other studies

Here is an informal contextualization with prior selected trials in a comparable setting (disclaimer, this is not a systematic review!).

In this figure, ORRs are reported for the overall study population of each trial, with no breakdown according to line of therapy.

Importantly, the characteristics of the populations included varied across studies. In NAPOLI-1, 53% of patients received NALIRI as second-line therapy in the metastatic setting, 34% as later-line treatment, and 13% had previously received systemic therapy, although not in the metastatic setting. In the Japanese study (jRCTs031210040), all patients had metastatic pancreatic cancer refractory to first-line gemcitabine-based chemotherapy, with no further detailed breakdown reported. In CodeBreaK 100, 55% of patients had stage IV disease at initial diagnosis, and all patients had metastatic disease at enrollment; 79% had received two or more prior lines of therapy. In KRYSTAL-1, 76% of patients had received two or more lines of systemic therapy before enrollment.

Bottom line: early signals of activity, but still within the historical ballpark

The much awaited phase 3 RASolute 302 trial will provide more precise estimates of response rates in a larger and more homogeneous population (second-line therapy in the metastatic setting).

Survival?

Interpreting survival data from single-arm early-phase trials should be done very cautiously. In order to receive multiple lines of therapy, patients need to be fit enough and likely have a more indolent tumour biology, unlike rapidly progressing cancers that can sometimes lead to death before any therapy can even be administered.

As a result, and this is somehow counterintuitive, we may see longer survival in later-line early-phase trials than in first-line phase 3 trials.

An example is durvalumab in urothelial carcinoma. In an early phase I/II study conducted mostly in previously treated/post-platinum patients, median OS was reported at 18.2 months. However, in the first-line phase III DANUBE trial, durvalumab monotherapy showed a shorter median OS: 14.4 months in the PD-L1–high population and 13.2 months in the ITT population.

As such, caution is warranted when drawing conclusions from PFS or OS data from early-phase trials.

Toxicity in the 300 mg cohort: more details are needed.

Below is a focus on the 300 mg cohort, the dose chosen for phase 3 trials.

There is a striking difference between the rate of grade ≥3 treatment-related adverse events (TRAEs, 34%) and the summed frequency of individual grade ≥3 TRAEs occurring in >10% of patients (13%).

Even though Table S2 reports additional safety data, it still leaves 10 patients with undocumented grade ≥3 TRAEs in the 300 mg cohort, since only events leading to dose modification and occurring in ≥2% of patients were described.

Toxicity was not dominated by a single adverse event, but rather distributed across multiple less frequent toxicities. However, the type of the remaining unreported grade ≥3 TRAEs (12% of patients) is not described.

Enthusiasm should not hamper cautious appraisal

Daraxonrasib shows unquestionable early activity in metastatic RAS-mutated pancreatic cancer. However, the data remain within a range comparable to that observed in other studies. Toxicity is a significant concern, and grade ≥3 toxicities are not fully detailed.

While we can understand the natural enthusiasm in such a poor-prognosis setting, we should at the same time try not to deviate from an objective and thorough appraisal of the data. Looking forward to the upcoming ASCO presentation! (See my previous post on what to look for in the full data presentation.)

To set the stage, in 2025, an unprecedented decision was made by the Australian Pharmaceutical Benefits Advisory Committee (PBAC) to recommend tumour-agnostic reimbursement of nivolumab with or without ipilimumab for all immunotherapy-sensitive advanced or metastatic cancers (see here). The recommendation allows treatment decisions to rely on clinicians’ “appropriate discretion” and “the best available evidence” without providing a clearer definition.

This recommendation could allow for a broad utilization of dual checkpoint inhibition (nivolumab plus ipilimumab) without any biomarker requirement.

Together with our colleague David Thomas, we explain why such a decision sets a concerning precedent. In the same issue of the journal, Vivek Subbiah and Razelle Kurzrock published a counterpoint article as part of a vigorous yet respectful academic debate.

A survival benefit of dual immunotherapy alone over anti-PD(L)1 alone has never been proven

Let’s start with the most immunosensitive tumor type: melanoma. Even in the CheckMate 067 trial, upfront “ipi-nivo” has not been formally shown to be superior to nivolumab alone in terms of overall survival. This is explained here, and summarized in the slide below.

In other tumor types, excluding melanoma, 2 meta-analyses have shown no overall survival benefit of the combination over nivolumab alone (Serritella and Shenoy, Rangwala et al.).

To be clear, dual immunotherapy may still represent an effective strategy in selected settings, for instance in patients progressing after first-line anti-PD(L)1 monotherapy in melanoma. In other tumour types, dual checkpoint inhibition has also demonstrated benefit when compared with control arms other than anti-PD(L)1 monotherapy, such as tyrosine kinase inhibitors in renal-cell carcinoma.

However, in patients with no prior exposure to immunotherapy, the evidence base to date does not support a survival benefit with dual immunotherapy alone over anti-PD(L)1 monotherapy alone in any tumour type.

The DART trial only reinforces our concerns

The DART/SWOG S1609 trial was a large phase 2 trial testing ipilimumab (lower dose of 1 mg/kg every 6 weeks) plus nivolumab across many rare cancers. Overall response rate (ORR) was the primary endpoint. The trial took 6 years to enroll 727 patients into multiple rare tumour cohorts. The signal was heterogeneous: more than half of the cohorts had no clinical activity as defined in the trial, while other rare tumour types showed meaningful activity. Across all patients, the median ORR was only 12 %.

While dual checkpoint inhibition had significant activity in some groups, the specific contribution of ipilimumab beyond nivolumab in DART remains unknown: the vast majority of patients (95%) had not received prior anti-PD-(L)1 therapy.

Physical toxicity

A key concern with dual immunotherapy is toxicity.

Immune-related adverse events are very different from those occurring with chemotherapy or oral inhibitors. They can affect any organ, and when high-grade, often requires weeks of corticosteroids. Immune-related toxicity can induce long-lasting impairment, with a potential impact on quality of life.

Again in melanoma, the 10-year follow-up data from the CheckMate 067 trial showed a 63% grade 3-4 rate in patients receiving ipi-nivo, as compared with 25% with nivolumab alone.

The dose used in CheckMate 067 was 3 mg/kg every 3 weeks for 4 cycles. A lower dose, 1 mg/kg, is known to be less toxic, however, still significantly higher as compared with anti-PD(L)1 monotherapy.

In DART, using the low dose of ipilimumab (1 mg/kg every 6 weeks), grade 3–4 treatment-related adverse events were reported in 39.2% of patients. Moreover, treatment-related deaths occurred in 1.8% of patients.

Here is a summary of the DART/ SWOG S1609 trial results:

• prospective trials are feasible in rare cancers, and this should be commended

• more than half of cohorts (54.7%) showed no clinical activity

• overall response rate = 12%

• unknown contribution beyond anti-PD(L)1 alone (with 95% of patients with no prior exposure)

• grade 3-4 toxicities = 39.2%

• treatment-related deaths = 1.8%

As such, we reach different conclusions and disagree that the DART trial supports a broad reimbursement for dual immunotherapy. Rather, we believe it reinforces our concerns.

Financial impact of the reimbursement decision

We noted that in Australia: “The projected expenditure associated with this decision is substantial, exceeding AU$1 billion, with an estimated AU$100–200 million attributable to use beyond existing indications. Funding commitments of this magnitude inevitably carry opportunity costs within a fixed health‑care budget, as resources allocated to one intervention are unavailable for alternative health programmes that may deliver greater population‑level benefit.”

One way forward: a biomarker-driven strategy together with informed data collection

In the Netherlands, the DRUP trial is an adaptive platform trial that provides off-label access to approved targeted therapies and immunotherapies for patients with advanced cancers lacking standard treatment options, while also generating evidence.

Last month, results were published reporting on 1,610 patients within DRUP, who were treated with 37 different off-label targeted or immunotherapies matched to their molecular profile. The report shows that while clinical benefit was seen in some biomarker-defined subgroups, the overall objective response rate (ORR) was only 16%.

The DRUP investigators recommended that: “To maximize patient benefit, […] off-label precision medicines should be used only within frameworks that systematically evaluate efficacy and toxicity, support biomarker refinement and enable stepwise assessment toward potential future label expansion.”

If anything, the conclusion of the DART trial, with a lower ORR than DRUP, should also be that such off-label use requires systematic outcome collection.

It is therefore important that reimbursement decisions that are not backed by clinical evidence, but are instead primarily motivated by addressing unmet need, are accompanied by structured data collection and planned re-assessments. Otherwise, as a health system, you can never know whether you are helping or harming patients. On one hand, you may expose your target population to unnecessary toxicities, and on the other hand, divert healthcare resources away from other patients. In this case, the use of a combination therapy is only justified if monotherapy is not sufficiently effective and this requires use in a trial setting and structured data generation.

Check out our full manuscript for more insights! (openly available here)

In RASolute 302, patients either received daraxonrasib, an oral RAS(ON) multi-selective, non-covalent inhibitor; or chemotherapy.

More than 90% of patients with metastatic pancreatic cancer have a RAS mutation with about 85% carrying a KRAS G12 mutation. Contrary to other KRAS inhibitors, daraxonrasib inhibits the active RAS(ON) form. Even RAS-wild-type tumors may be dependent on RAS activation, which is the rationale for investigating daraxonrasib beyond RAS-mutant populations.

The press-release says: “In the overall (intent-to-treat) study population, daraxonrasib demonstrated a median OS of 13.2 months versus 6.7 months for chemotherapy, with a hazard ratio of 0.40 (p < 0.0001). Daraxonrasib was generally well tolerated, with a manageable safety profile and with no new safety signals.”

The results will be presented in a few weeks, during ASCO2026 plenary session, and a simultaneous publication in the New England Journal of Medicine would not be surprising.

Most doctors, scientists, and academics agree that we should not practice medicine by press release. However, following the RASolute 302 press release, social media and the press have echoed the results, mostly within a one-sided framing and an abundance of superlatives, presenting daraxonrasib as a “miracle” or a “revolution.”

While doubling median survival in patients with such a poor prognosis is highly unusual, caution is warranted when interpreting any results. Here are some initial thoughts, and what I think we should be looking for during ASCO.

Medicine by press-release? Why this is problematic.

If we look at the toxicity data for daraxonrasib given as monotherapy, as in RASolute 302, data from patients with advanced pancreatic cancer treated in the phase 1/1b study were reported in a previous Revolution Medicines press-release issued in December 2024 (screenshot).

A quick read gives the false impression that there is minimal grade 3 or higher treatment-related toxicity.

This wording indicates that no individual adverse event exceeded 10% of patients. What is missing from the press-release is that, when taken together, approximately 34% of patients in the 300 mg cohort experienced grade 3 or higher treatment-related adverse events (see below).

In other words, this does not mean low toxicity, but rather distributed toxicity.

Toxicity data will be much awaited from the ASCO 2026 presentation. We can already get a glimpse of how “nasty” the drug can be. Here, I’m quoting Senator Ben Sasse, who has publicly shared his personal story after being diagnosed with advanced pancreatic cancer and enrolling in RASolute 302.

Ben Sasse : “I take it orally, but it’s a nasty drug. It causes crazy stuff like my body can’t grow skin and so I bleed all out of a whole bunch of parts of me that shouldn’t be bleeding.”

Former Senator Ben Sasse has given numerous interviews describing what he is going through as a patient with advanced pancreatic cancer. He has explicitly mentioned pain reduction and significant tumor shrinkage, and noted that “this experience has seemingly extended both quantity and quality of life” (here). At the same time, his public appearances and recounting also highlight the significant toxicities associated with the treatment.

So when we come back to the most recent press release from Revolution Medicine: “Daraxonrasib was generally well tolerated, with a manageable safety profile and with no new safety signals.” I wonder if patients, including Ben Sasse, would find that this wording is accurately reflecting his and other patients’ general experience.

Medicine by press-release? If we have no choice, let’s do it!

Let’s gather data from the press-release and from clinicaltrials.gov, and let’s see what we can make of it:

• the trial included 500 patients, unselected for specific RAS mutations

• primary endpoints = PFS and OS in RAS-G12-mutant patients

• secondary endpoints = OS and PFS in all patients

• “Trial met all primary and key secondary endpoints, including progression-free survival and overall survival”

• “In the overall (intent-to-treat) study population, daraxonrasib demonstrated a median OS of 13.2 months versus 6.7 months for chemotherapy, with a hazard ratio of 0.40 (p < 0.0001).”

• Study start (first patient): October 16, 2024; Enrollment (500 patients) completed: at least in December 2025; Total accrual duration: maximum 14 months; Press release 13 April 2026 (data-cut off likely in March 2026); Maximum follow-up duration about 17 months

The headline OS result is based on a secondary endpoint in the broadest patient population, while the primary endpoints, PFS and OS in the KRAS G12-mutant population, remain unreported

Why the OS estimates in the entire population could be overestimated.

Several factors may have contributed to an inflation of the reported results.

1. Power is usually calculated for the primary endpoint(s). As such, OS in the ITT population may be insufficiently powered. Additionally, hierarchical testing can reduce the available alpha, effectively reducing power in the ITT analysis.

2. Nested versus adjacent subgroups analyses. As the primary endpoints are in the RAS G12-mutant population, one key question for the overall population is: are non–G12 RAS-mutant patients truly deriving benefit, or is the observed effect in the “overall population” mainly driven by the G12-mutant population?

3. Early stopping rules carry a risk of inflating effect estimates, as shown in empirical work. Were such rules used in RASolute 302?

4. Efficacy–effectiveness gap? Patients in clinical trials often do not reflect real-world patients, who typically derive less benefit and may experience more toxicity. This is referred to as the efficacy–effectiveness gap. Even if the reported survival benefit is considered reliable, how it will translate into real-world practice remains an open question.

Longer follow-up matters: the NAPOLI-1 example.

Will longer follow-up data will ultimately be made available. The question is real, as the results are presented as being “primary and final”.

To illustrate the effect of longer follow-up, here is the NAPOLI-1 trial as an example, where we have two data points from the 2016 and 2019 reports. NAPOLI-1 was conducted in a setting comparable to RASolute 302. As shown in the estimations below, with longer follow-up, an early separation tends to diminish, with a less pronounced statistical gain.

This illustrates how early apparent benefit may attenuate with longer follow-up.

Key questions for ASCO and the full publication

• Statistical details: early stopping rules if specified, alpha spending for each endpoint (particularly the ITT secondary endpoints), number of interim analyses planned, etc.

• What is the median follow-up and OS maturity, what about censoring?

• The shape of Kaplan–Meier curves: early separation ? curve crossing ? any tail?

• ITT vs primary endpoints population: how much of the OS benefit in the entire ITT population may have been driven by the prespecified RAS-G12 population

• Is the objective response rate (ORR) maintained? In the phase 1/1b study, the confirmed ORR was 35% in the RAS G12 population and 29% in the overall population. Classically, ORR tends to decrease from early-phase to later-stage studies. As a reminder, sotorasib (a KRAS G12C inhibitor) had an ORR of 37.1% in the early-phase CodeBreaK 100 trial, which later decreased to 28.1% in CodeBreaK 200.

• Toxicity data!

• Quality of life data: if presented: how and when they were measured, compliance rate, timing of measurements, etc.

Long-term efficacy? Drugs facing cancer biology

In 2011, George Sledge highlighted that the genomic era could bring both promise and chaos. He distinguished “stupid” cancers, like those driven by a single driver mutation, such as chronic myeloid leukemia (CML). In these settings, inhibitors can be transformative: in CML, life expectancy became comparable to the general population after the introduction of imatinib.

Yet, most solid tumors are different. They rarely rely on a single driver, and instead harbor co-mutations and multiple clones. In advanced disease, resistance at some point is the rule rather than the exception. Although daraxonrasib does not target a single mutation but rather RAS across multiple contexts, advanced pancreatic cancer remains a biologically “smart” cancer.

To which extent a broad RAS(ON) inhibition can temporarily suppress the genomic and adaptive complexity that usually drives resistance in pancreatic cancer is unclear.

My (temporary) conclusion.

Even though I’m awaiting for the data, I see daraxonrasib as a pharmacological breakthrough, however associated with significant toxicities and inherent biological hurdles when it comes to long-term efficacy.

On social media, discussions have at times become emotional, with any critical appraisal of a trial being perceived as a lack of empathy, which is unfortunate. We should not forget that financial stakes are high. Personal stories should not replace careful interpretation of the data.

In oncology, we often say a drug should help patients live longer and/or better. In the case of daraxonrasib, even with a survival benefit, the toxicity profile and the uncertainty around long-term benefit may make the choice not as straightforward for a fraction of patients.

Whether a treatment that delays, but does not prevent, death, while adding significant toxicity, constitutes a “revolution” is a question best answered by patients, not by Key Opinion Leaders (KOLs) or companies.

In this new publication, we review if the current evidence support the use of systemic therapy in high-risk Smoldering Myeloma, or if observation should remain the standard (which is, spoiler, our conclusion).

The AQUILA trial was a global, open-label trial, that tested whether patients with high-risk Smoldering Myeloma should receive daratumumab (for up to 36 months) or active surveillance. The primary endpoint was Progression-Free Survival, Overall Survival (OS) was a secondary endpoint.

AQUILA was published in the New England Journal of Medicine, and showed an improvement in PFS in patients allocated to receive daratumumab (HR = 0.49; 95% CI, 0.36–0.67; P<0.001), with 5-year progression-free survival of 63.1% versus 40.8% with active monitoring.

The trial also report an overall survival difference favoring daratumumab (HR = 0.52; 95% CI, 0.27–0.98), and the trial led to the approval by both the FDA and the EMA, and its inclusion as a Category 1 option in the NCCN guidelines for high-risk smoldering myeloma.

The survival difference observed in AQUILA has been cited as a reason that this study should change the standard of care to treat all patients with high-risk smoldering myeloma. This is the reason why we focus today on survival data. Our overall conclusion is simple: observation with appropriate surveillance should remain the standard-of-care in high-risk smoldering myeloma.

Key Subsequent Data, once again brought to us exclusively by IQWiG!

From the seminal publication, we knew that 105 patients in the control group received at least one subsequent therapy, and that 35 patients received daratumumab or another anti-CD38 compound as part of their first-line therapy. Even though this number was low (33.3% of patients receiving subsequent systemic therapy), some could have contend that daratumumab or other anti-CD38 therapy could have been given later in the course of the disease: this piece of data was, to date, unreported.

However, as highlighted in a previous post, the german Health Technology Assessment body, IQWiG (Institut für Qualität und Wirtschaftlichkeit im Gesundheitswesen), is requiring granular subsequent therapy data to be provided by sponsors, and are making their assessments publicly available.

Based on their February 2026 report (available here), we now know that “a maximum number of 61 (58%) patients (…) received subsequent therapy with CD38 antibodies, regardless of the line of therapy.”

As a result, at least 44 control group patients 1) progressed, 2) received a subsequent therapy, and 3) were never exposed to daratumumab or other anti-CD38 therapy at any time. These patients are making at least 22.4% of control arm patients. Even though we still don’t know how many patients died without ever receiving daratumumab, this key metric provide important context. On the figure below, the absolute difference in 5-year survival is contextualized with this metric:

To my knowledge, this key metric reported in the IQWiG report is not available elsewhere, highlighting the unique value of their work and reports.

OpenEvivence failed: do not rely on AI tools for appraising the data

There is major enthusiam about new AI tools, which I partly share, in particular for coding. A recent trend concerns the value of OpenEvidence, a platform which is officially partnering with journals like the New England Journal of Medicine or JAMA.

Logically, you would expect a New England Journal of Medicine publication like AQUILA to be accurately summarized. Yet, on April 11, 2026, while reviewing the proofs of our work, I tested OpenEvidence. Here is a copy paste of my “conversation”.

“Could you please provide me the summary of the overall survival results from AQUILA ?”

OpenEvidence:

Me pushing back: “But I don’t understand, you say it was statistically significant, however in the hierarchical testing, it should only be tested after ORR and PFS2 no ? And PFS2 results are not reported, nor is the p-value for OS ?”

OpenEvidence:

This example is not trivial. Many think OpenEvidence accurately summarizes key data from seminal trials, and here it failed to do so with the survival data from a trial that could affect thousands of patients worldwide. My take: do not be dependent on AI tools. If the computer shuts down, you should still be able to talk to your patients.

Informative censoring in survival data? Likely.

We noted an imbalance in censoring rates for both PFS and OS. It is plausible that patients dropping out in excess from the control arm were healthier, more connected, seeking for active therapy outside the trial, and had a lower-risk of event as compared to patients remaining in the control arm. Those patients, had they stayed, could have been late censored (at data-cutoff) instead of being early censored. This is called “disappointement-driven informative censoring”.

In AQUILA, we generated pseudo-Individual Patient Data and, using the BREAKING-ICE App©, we conducted a sensitivity analysis to determine the minimum number of patients censored over the first 12 months in the control arm whose censoring status would need to be changed from early to late censoring for the confidence interval to cross one.

We found that changing the status of 3 patients was sufficient. Notably, even under this conservative sensitivity analysis, censoring rates at 12 months remained significantly higher in the control group (3.6% vs 0.5%).

The analysis is openly available here (select “AQUILA-OS”), and shown below:

Summary of AQUILA survival data

1. an absolute difference of 6.1% is observed at 5-year

2. this difference is not formally statistically significant

3. likely driven by suboptimal subsequent therapy in control group patients (22.4% received a subsequent therapy without ever being exposed to an anti-CD38 therapy; how many patients died without ever receiving datatumumab?)

4. likely driven by informative censoring

Observation should remain standard-of-care, and this is actually in line with NCCN guidelines!

Based on the survival data, as well as many other aspects, we conclude that “observation including appropriate surveillance should remain the standard of care for high-risk smoldering multiple myeloma.”

Interestingly, it should be noted that a category 1 NCCN guidelines recommendation does not imply a preferred option. Here is a screenshot from Version 5.2026 NCCN guidelines where panelists do not recommend systemic therapy as the preferred option.

Check-out our full paper for more insights (here). See also the recent video on this topic with here!

It is the first immunotherapy for extensive-stage small-cell lung cancer (ES-SCLC) to meet the PASKWIL criteria.

This drug is offered on its website in China for US$290 per vial, yet in Europe it costs roughly 5.5 times more. You might think that this is due to the effort of making the drug relevant for the European market. But no… there were in total only 7 European patients recruited in the trial. The effects of the drug are also rather modest: in combination with chemotherapy, it delays progression or death by about 1.4 months compared with chemotherapy alone. And the median overall survival benefit is 4.5 months (HR 0.63, 95% CI: 0.49–0.82).

Today’s question is simple: although the PASKWIL criteria provide an invaluable framework for evaluating novel indications and therapies, are HTA agencies equipped to assess them when multiple compounds from the same class are tested?

Most profitable class of drugs

It is easy to lose track of how many immunotherapies like serplulimab are currently available. Nivolumab and pembrolizumab were the first to enter the market in 2014 and have since developed the most indications. Currently, there are around 14–15 or even more me-too versions of these drugs on the market, but this is only the beginning of what is expected to be a tsunami of such agents.

In total, more than 180 companies are developing over 200 of these immunotherapies.

The reason for such investments is profitability. Companies have found ways not to compete directly in the market. Studies are often conducted slightly differently, in different indications and settings, or in combination with different agents. We are often led to believe that one drug is slightly better than another for a specific indication based on “expert opinion” or the overinterpretation of subgroups or post-hoc analyses; but in fact there is no robust evidence to justify such conclusions.

There are few within-class head-to-head trials, and even fewer showing a survival gain of one compound over the other (see here).

The landscape of immunotherapies for extensive-stage small-cell lung cancer

Serplulimab is not the only immunotherapy authorised in Europe for ES-SCLC. Two other immunotherapies, atezolizumab and durvalumab, are also authorised. However, their effect is even smaller, with median overall survival benefits of 2 months (HR 0.76, 95% CI: 0.60–0.95) and 2.4 months (HR 0.75, 95% CI: 0.63–0.91), respectively. (See Assessment of the EMA comparing the results here).

But there is also a history of negative trials. For instance, KEYNOTE-604, a Phase 3 trial of pembrolizumab plus chemotherapy, failed to meet the overall survival endpoint required for full FDA approval, and its accelerated approval was withdrawn for metastatic SCLC in March 2021. Also, nivolumab plus chemotherapy (EA5161) failed to reach statistical significance versus chemotherapy alone (Figure 1).

To our surprise, no HTA agencies appear to consider these negative trials in their assessments, and thus they may miss the overall picture. Surely, we do not want reimbursement decisions to be based on coincidence or bias. If one looks at the overall effect of immunotherapies in ES-SCLC, they will not pass the Dutch test.

Multiplicity and me-too drug development

This introduces the concept of multiplicity which has emerged with the surge of me-too drug development. It is not excluded that the same class of agents, tested again and again, will ultimately yield some positive results by chance alone. In such cases, statistical estimates can be reassessed taking into account multiplicity, using methods like Bonferroni or Stein estimation.

In a recent work from Carlisle and colleagues focusing on immune checkpoint-inhibitors trials, the authors used Stein estimation to account for the crowded trials with me-too compounds. They found that, after adjusting for the totality of trials, estimates of results, in terms of hazard radio, decreased in 35% of trials, and amoung trials with an original p-value that was inferior to 0.05, 27% of those lost significance after adjustment (their work here).

In other words, while considering the redundancy of trials in similar class of drugs, it is not excluded that some positive results will be the result of a statistical artifact, which should be accounted for with appropriate methods.

Reimbursement decisions should consider the totality of evidence

In the following scenarios, we use analogies to illustrate how reimbursement decisions are made in practice and how they are intended to be made.

• Scenario 1: Imagine you are standing in a drugstore, looking at different brands of a supplement on the shelf. You read the packaging, compare prices, and then, based on your needs and budget, decide whether to choose one. This is similar to how reimbursement decisions are often made.

• Scenario 2: Imagine that, before going to the drugstore, you search PubMed for studies on different variants of that supplement. You realise that the evidence is not straightforward: some studies are positive, while others are negative. In such cases, meta-analyses, systematic reviews, and even statistical approaches accounting for the numerous trials, can help determine whether the overall effect is clinically meaningful for you.

This is essentially what health technology assessors are expected to do. Their role is to consider all available evidence when evaluating a medicine, including negative studies involving the same class of drug for that indication. Without strong HTA, healthcare systems risk funding expensive drugs with little benefit.

Where are subsequent therapy data from NIAGARA?

In a previous video, I made a side-by-side comparison between NIAGARA and AMBASSADOR trials (video available here). Both trials were run in the peri-operative setting of bladder cancer (NIAGARA is neo-adjuvant followed by adjuvant, AMBASSADOR is purely adjuvant). Both trials demonstrated an EFS/DFS gain in the experimental arm. Trials’ design (left) and their corresponding EFS/DFS results (right) are shown below:

However, only NIAGARA demonstrated an overall survival gain. The side-by-side comparison was primarily aimed at highlighting that, in global trials like NIAGARA, a survival gain may be driven by the limited access to optimal care for patients who recur.

Such limited access is explained when some enrolling countries are simply lacking access to the best care outside the trial. Conversely, in trials like AMBASSADOR, that was exclusively run in the US, optimal access to post-protocol care, which is desirable!, may have explained the absence of survival gain.

In a paper published in the European Journal of Cancer (full paper here), we posit that such global trials are rewarded in terms of guideline recommendations and approvals.

Here is the side-by-side comparison showing how NIAGARA was “rewarded” as compared to AMBASSADOR.

The subsequent therapy data from NIAGARA are now available in the IQWiG assessment report!

Even though the discussion with the NIAGARA trialists during the first presentation in September 2024 suggested that post-protocol access to immune checkpoint inhibitors would likely be low, we were awaiting for those data until very recently.

They are still, to my knowledge, unpublished.

However, NIAGARA was recently evaluated by IQWiG. In their assessment, some key subsequent therapy data are now reported. Based on those, we can also compare them to AMBASSADOR (see table below).

A common, yet misleading, argument is to use real-world data (often involving frailer patients) to explain the low rates of post-protocol care in trials, when the underlying cause is limited global access. A more appropriate comparison is with trials such as AMBASSADOR, which are highly valuable to illustrate what can be expected as optimal subsequent therapy in a clinical trial setting.

Here, we can derive two conclusions for patients in the control group of NIAGARA:

1. confirmation of a very limited access to checkpoint inhibitors, received by only 25% of patients with an EFS event.

2. a limited access to any subsequent therapy, 38% of patients with an EFS event in NIAGARA, significantly lower than 57% of patients with a DFS event in AMBASSADOR.

Both may have biased the trial toward an overall survival gain in NIAGARA.

Why we should follow IQWiG and other HTA bodies reports.

It should be noted that other HTA bodies may also be providing key data. In the case of NIAGARA, the CDA-AMC Canadian agency also reported subsequent therapy in their report (available here, Table 8).

A specificity is that IQWiG is also systematically requiring the first subsequent treatment as well. Subsequent therapy data lacking granularity can confuse their interpretation (we have written about this together with , see here)

The sponsors, when submitting a dossier to IQWiG, must specifically provide :

“details regarding subsequent treatments administered after discontinuation of the study medication (for oncology-related issues, please also provide separate details regarding the first subsequent treatment).”

They are also requiring, systematically, the numbers of censored patients:

“ In survival analyses, the Kaplan-Meier curve should be plotted, including information on patients at risk and the proportion of censored patients over time (at multiple time points). A separate Kaplan-Meier curve should be plotted for each endpoint for which such an analysis is performed.”

IQWiG is doing an outstanding work. Their reports, which are publicly available, are usually translated into English a couple of months after publication. English reports are available on the IQWiG website, switching to english language, and searching for the relevant drugs. Definitely worth following closely!

Here is my full talk. I would like to thank the IQWiG team again for the invitation, for their warm welcome, and for their outstanding work. You should definitely follow IQWiG’s activities and read their reports, which may contain more data than what is available in publications.

+ At 41:40, get a glimpse of what can be done with the new V2 version of the BREAKING-ICE App© and its Assistant-KM companion app, with the updated sensitivity analysis based on the 5-year follow-up of the NATALEE trial.

The BREAKING-ICE App© starts from Kaplan–Meier curves reconstructed from pseudo individual-level data (pseudo-IPD) using the IPDfromKM package and data extracted from original curves. These are not real patient data: they are numerical estimates consistent with published curves and tables, used to explore informative censoring through sensitivity analyses. Upstream file preparation can be performed with Assistant-KM-BREAKING-ICE©.

A central goal of the BREAKING-ICE App© is to visually and instantaneously explore how alternative assumptions about censoring may affect curves and statistical outputs. The BREAKING-ICE App© has been published in The Lancet Oncology, June 2025, available here.

The BREAKING-ICE App© V2 is available here.
The brand-new Assistant-KM-BREAKING-ICE© is available here.

Both come with a user guide, and a video tutorial will be available soon. Please let me know what you think!

In our new paper, published in the Journal of Cancer Policy and openly accessible here, we analyze how and how quickly the National Comprehensive Cancer Network (NCCN) updates its guidelines in response to Food and Drug Administration (FDA) approvals and withdrawals over time.

The NCCN produces some of the most influential and referenced oncology guidelines worldwide. These guidelines strongly influence hematologists’ and oncologists’ clinical decision-making, and in the United States, when the NCCN recommends a drug, it regularly becomes covered by Medicare, Medicaid, and private insurers.

Given this influence, we were curious about how quickly NCCN updates its recommendations following regulatory decisions, as these updates likely affect patients, clinicians, and payors.

The Question: Do Guidelines Keep Pace with Regulatory Reality?

We examined how quickly the NCCN updates its guidelines after decisions from the U.S. Food and Drug Administration. Specifically, we looked at 83 FDA cancer drug approvals and 8 FDA withdrawals between October 2022 and July 2024. We also subdivided those FDA approvals into traditional regular approvals and accelerated approvals.

Our main findings:

The NCCN incorporated FDA approvals very quickly, with a median of 19 days post-approval. In contrast, the NCCN updated withdrawn indications significantly slower, with a median of 57 days post-approval.

The NCCN removed only 3 of 8 withdrawn indications from its guidelines. The majority (5 of 8) remained recommended in some form, and four of these were recommended with either uniform or majority NCCN expert panel consensus (Category 2A and 2B recommendations).

We found no significant differences in the time to NCCN update between regular and accelerated approval statuses, with median times of 20 days and 16 days, respectively.

Further insights into how NCCN panels make decisions

We also examined the NCCN’s recommendations for prescribing these therapies. We found that only one FDA-approved treatment (1.2%) included in the NCCN guidelines lacked uniform expert consensus (>85%) as a category 2B recommendation, which still comes with majority consensus (>50%).

This finding suggests that NCCN experts largely support the FDA’s approval decisions, in contrast to their attitude toward withdrawals as highlighted above.

As we point out in our discussion, this tendency toward universal consensus raises concerns, especially since 38% of all FDA approvals and nearly 80% of accelerated approvals come from single-arm, uncontrolled studies (see here).

Additionally, we found that in 22% of cases, the NCCN made additions or modifications to recommendations before receiving FDA approval. This does not imply that the NCCN is bypassing the FDA. Both the NCCN and the FDA respond to similar data published in the literature, and the NCCN may have updated these recommendations in anticipation of upcoming FDA approvals. However, this finding further reinforces our previous finding that the NCCN tends to incorporate new drugs quickly, even faster than approval, in contrast to delays in removals.

This study’s implications:

1. Delays in updating withdrawn indications may lead to doctors prescribing cancer drugs that lack proven survival or quality of life benefits, while exposing patients to the toxicities of those very drugs.

2. Given that the NCCN recommendations affect insurance coverage, continuing to recommend ineffective drugs may drive unnecessary spending of patients’ and taxpayers’ dollars.

3. Many cancer drugs are approved based on limited evidence, especially through the accelerated approval pathway. If guidelines rapidly adopt these therapies but hesitate to remove them when confirmatory studies show these drugs are not effective, the result may be persistent low-value care.

Ultimately, this study highlights that the NCCN has competing priorities that are very difficult to balance. The NCCN aims to equip clinicians with as many treatment options as possible that insurers cover and to provide evidence-based recommendations to aid cancer patients. These competing priorities likely contribute to our study’s findings.

Potential Solutions

1. Have two separate bodies to perform the role of the NCCN. One that provides clinical guidance and another that provides coverage decisions and cost-effectiveness analyses.

2. Alternatively, a federal standard or law could be established that compendia incorporate and align with FDA regulatory guidance, especially regarding withdrawals.

Many thanks to my mentors, and , for their guidance, support, and efforts in moving this project forward.

I’ll conclude with the take-home point: guideline makers should continue to quickly adopt new drugs to maximize patient access and inform clinicians, but they should also remove withdrawn drugs with equal efficiency when confirmatory studies call their efficacy into question, leading to FDA withdrawal.

In addition to receiving maintenance HER2 and endocrine therapies, 518 patients were randomly assigned to receive palbociclib – an anti-CDK4/6 inhibitor – or no other therapy. The primary end point was investigator-assessed progression-free survival (PFS).

The study was just published in the New England Journal of Medicine and there are several points to raised while looking at this study report.

Does the reported PFS gain seem reliable?

In PATINA, the authors report a 15 months median PFS gain, with 44.3 months median PFS in the palbociclib group and a 29.1 months median PFS in the control group (HR = 0.75; 95%CI = 0.59 - 0.96; two-sided P=0.02).

Two key elements have to be noted, that could have amplified the gain as reported.

1- Informative censoring.

The numbers of censoring patients are (sadly) unreported in the NEJM publication, however, one can estimate those by using digitized curves. In the BREAKING-ICE App, select PATINA-PFS and you see a significant early imbalance in censoring with 5.3% of patients censored over the first 6 months in the palbociclib group and 12.1% of them in the control group.

Why patients left in excess after being assigned the control group?

Some may have been disappointed by allocation to the control arm. Some may have been eligible for other trials and left PATINA, possibly encouraged by trialists.

Now here is the crux of the issue: are patients lefting PATINA in excess the same as those remaining in the control group? A reasonable assumption in that those patients may have been more connected, had more options, and may have been lower-risk patients in terms of presenting an event. If so, the remaining control patients would be enriched for higher-risk patients, potentially inflating the observed PFS difference.

The Reverse Kaplan-Meier analysis explore whether informative censoring could have occured, by flipping events and censoring events. In PATINA, both this plot and its corresponding “Cox” analysis, being “statistically significant”, show informative censoring could have occured. (Go to the Reverse-KM App© and select PATINA-PFS, see screenshot below).

2- Open-label design and investigator assessment.

We know some subjectivity can affect investigators when “calling” a progression event based on a CT scan: that’s why many trials use blinded-independent central review (BICR).

Here, the combination of an open-label trial with no BICR could have also amplify the PFS gain.

• “Early progression calls” in the control arm. In an open-label setting, investigators may be more inclined to declare progression earlier in control patients when scans are borderline or tumor growth is still below formal RECIST thresholds, because there’s a practical incentive to move quickly to the next line of therapy or another trial.

• “Late progression calls” in the experimental arm. If clinicians believe the new therapy is helping, they may be less likely to call a progression on equivocal imaging, thereby delaying the recorded progression date.

• Symptoms deterioration interpretation differing by arm. The decision to prescribe non-scheduled CAT-scan is often influenced by symptoms and clinical course. Control patients with worsening symptoms may trigger a CAT-scan sooner than a patient in the experimental arm, with the same symptoms being attributed to toxicity. In the last situation, the official progression is delayed to the next scheduled CAT-scan.

In conclusion, given the possible biases introduced by informative censoring, an open-label design, and the lack of BICR, the PFS gain as reported may have been artifically inflated.

Is PFS a direct clinical endpoint?

The answer is no. PFS is a composite time-to-endpoint enpoint: progression (largely based on CAT-scan thresholds) or death can both count as an event. (see the figure below).

One issue is that often, and this is the case in PATINA, we don’t have the“breakdown” of PFS events. In other words, we don’t know whether the event was a CAT-scan progression, or death (see this work in Nature Reviews Clinical Oncology, by Anushka Walia and , with a key figure below).

Without kwowing the breakdown of events, the relevance of a PFS gain is difficult to assess, as death and an asymptomatic progression are very different events, still counting similarly in PFS estimates.

In PATINA, we have only an indirect clue that most events were disease progressions and not deaths in the Figure S1. CONSORT Diagram, as the reason for discontinuation of treatment was death in only a few instances (2 patients in the palbociclib group, and 1 patient in the control group)

Is a PFS gain clinically meaningful?

This question is slightly different from the definition of PFS: even though PFS is not a direct clinical endpoint, we often based our treatment decisions based on progression. As a results, and even if the progression is totally asymptomatic, the event will have a clinical impact on patients simply by triggering a change in care.

“A surrogate endpoint is an endpoint that patients didn’t know mattered until a doctor told them it did” (from )

Some may find this quote provocative, but it highlights that the relevant question about any surrogate, and this applies to PFS, is whether the gain will ultimately led to better outcomes for patients, which are generally accepted as either living longer (Overall Survival gain), living better (Quality of life - QoL - gain), or both.

In PATINA, some have argued a median PFS gain of 15 months (29 to 44 months) is undoubtely clinically meaningful, as this delay the time to the next line of therapy, which is often chemotherapy. This argumentation is assuming that those patients have a QoL gain by delaying the progression event.

This misses 1) the QoL data from PATINA are not yet published, and the data presented so far show HRQoL was “maintained” 2) patient are receiving an additionnal therapy – palbociclib – which does not really fit into the category of a “well-tolerated drug”.

Is palbociclib in PATINA “well tolerated”?

As seen in the supplementary (Table S5 below), 55% of patients receiving palbociclib in PATINA had at least one dose interruption; 28% had 1 dose reduction and another 30% had 2 dose reductions.

In Table S7, we can see that 18.1% of patients experienced adverse events “at least possibly related to treatment that led to palbobiblic discontinuation”.

Not all adverse events affect patients equally: some are asymptomatic lab abnormalities, while others are clinically meaningful symptoms. In PATINA, among the 15 most frequently reported adverse events in the palbociclib group, 11 of 15 (below in yellow, excerpt from Table 2) are primarily clinical rather than laboratory abnormalities.

Conclusions — context with DESTINY-Breast09 and HER2CLIMB-05

The main efficacy results from PATINA – the PFS benefit – may have been amplified by informative censoring and the unblinded nature of both treatment allocation and progression assessment.

We lack the breakdown of PFS events which could inform treatment-decision making with patients : there is a huge difference between a CAT-scan progression and death.

While the QoL data conclude that HRQoL was maintained (presented, not published), many limitations can affect QoL analyses in general, and QoL data should not obfuscate the obvious: the added toxicity of taking an additionnal compound.

The timing of the PATINA publication is surprising, occuring more than one year after the results were presented in December 2024 in San Antonio. Recent publications of novel strategies in first-line or maintenance treatment of patients with HER2+ advanced or metastatic breast cancer may be one possible explanation.

DESTINY-Breast09 studied first-line HER2+ advanced/metastatic breast cancer (54% of patients were HR+) and showed that trastuzumab deruxtecan plus pertuzumab significantly improved BICR-assessed PFS versus taxane–trastuzumab–pertuzumab (THP) (median 40.7 vs 26.9 months; HR = 0.56) with high response rates but a higher risk of adjudicated interstitial lung disease (ILD) or pneumonitis (12.1%, including rare grade 5 events).

HER2CLIMB-05 studied first-line maintenance in HER2+ metastatic breast cancer (53% of patients were HR+) after induction therapy and showed that adding tucatinib to trastuzumab/pertuzumab significantly improved investigator-assessed PFS versus placebo (median 24.9 vs 16.3 months; HR = 0.64), with diarrhea and transaminase elevations common and 13.5% discontinuing tucatinib due to treatment emergent adverse events.

PARP-inhibitors mechanism of action

PARP is a protein involved in repairing the DNA when wrong copies are made, which tipically occurs in cancers.

PARP inhibitors are drugs that are most effective when other DNA repairing genes are already impaired in cancer. These genes are called homologous recombination repair genes (HRR), which can lead to homologous recombination repair deficiency (HRD). Those are usually grouped in two buckets:

1. BRCA1 and BRCA2 mutations (BRCAm, “m” stands for mutation) makes one group.

2. and non-BRCA mutations (non-BRCA HHRm) constitute the other group, with mutations individually less frequent: ATM, ATR, CHEK1, CHEK2, FANC, MRE11A, PALB2, RAD51C and others.

Three trials combining PARPi and ARPI

The three trials are PROpel, TALAPRO-2 and MAGNITUDE. They resulted in five different labels showing inconsistencies between the US Food and Drug Administration (FDA) and the European Medicines Agency (EMA). In our paper, we highlight issues which are mainly:

1. The flaws in the trial designs of mainly PROpel and TALAPRO-2

2. The lack of harmonization of HRD

3. Uncertain benefits of combination therapies and control arm issues

4. The differences in approvals and labels between the FDA and the EMA

Trial designs

In PROpel, 796 mCRPC patients were randomly assigned to olaparib or placebo with abiraterone (plus prednisone or prednisolone). The trial adopted an all-comer design with no stratification or prespecified alpha-controlled analysis for BRCAm/HRRm, despite the EMA’s recommendation to do so.

• the primary endpoint radiologic progression-free survival (rPFS) was met.

• no statistically significant difference in overall survival (OS) between the two groups was observed.

• both regulatory agencies suspected that the potential benefit might have been driven by the subgroup of patients with BRCAm (n = 85), in which both rPFS (HR 0.24) and OS (HR 0.30) were significantly improved.

• in contrast, in the non-BRCAm group (n=427), both endpoints were not met.

  ---

In TALAPRO-2, 805 mCRPC patients were randomly assigned to talazoparib or placebo with enzalutamide. This trial also adopted an all-comer design, albeit with a stratification for HRR mutation. A second cohort of HRRm patients only followed, and the combined HRRm population across both cohorts was analysed.

• both independent primary end points of rPFS in the all-comers cohort and in the combined HRRm population were met.

• rPFS and OS for non-BRCA HRRm were not significant.

• final OS analysis was significant in the overall population. However, better in HRRm subgroup (HR 0.55, 95% CI 0.36–0.83; p = 0.0035) but not significantly different in non-HRRm or unknown patients.

  ---

In MAGNITUDE, there were two cohorts: 423 HRRm mCRPC patients in cohort 1 and 247 non-HRMm CRPC patients in cohort 2. They were randomized to either niraparib or placebo combined with abiraterone (plus prednisone or prednisolone). Cohort 2 was stopped early because of futility in a pre-specified interim analysis. As the greatest effect was expected in BRCAm, a protocol amendment was implemented in 2020, approximately one year after the start of the trial, to modify the statistical analysis plan. The BRCAm subgroup was assigned as the primary efficacy population to be analyzed first. If statistical significance was reached in this subgroup, the entire Cohort 1 population would then be tested.

• rPFS in the BRCAm subgroup was met

• rPFS in the overall Cohort 1 was also met

• no significant effect was seen in non-BRCA HRRm patients and even a potential detrimental effect on overall survival (OS HR 1.13, 95% CI 0.77–1.64).

Lack of harmonization of HRD

The definition of HRD is not consistent across trials. For example, the PROpel trial did not consider ATR, NBN, FANCA and MRE11A genes 4 whereas TALAPRO-2 included MLH1, a mismatch repair gene, which is not part of the HRR pathway.

Ideally, researchers should be given access to both biomarkers and clinical data. Aggregating data can increase statistical power and help us better understand how different HRRm impact treatment outcomes.

Uncertain benefits of combination therapies - and suboptimal control arm

In contrast to monotherapy, where ineffective drugs are discontinued, PARP inhibitor use can continue even if the effectiveness is driven only by ARPIs. This potential over-treatment comes with increased toxicity and higher healthcare costs.

Also, in these studies, patients in the control arm who had previously received an ARPI arguably received substandard care by being assigned another ARPI (the so-called “ARPI-switch”). The EMA also raised concerns about patients with visceral disease who had not received docetaxel but had been treated with an ARPI, and therefore restricted the indication to those for whom chemotherapy is not clinically indicated.

Differences in labels between regulatory agencies

Of the three trials, only MAGNITUDE (niraparib) led to the same conclusion from both the EMA and FDA, as its design effectively identified the benefiting subgroup. The discrepancies observed in the assessment of the other two trials stem from differences in their design.

In general, if a positive trial demonstrates benefit in a biomarker-positive subgroup, FDA guidance allows approval only for that population. In contrast, in the case of both PROpel and TALAPRO-2, EMA authorized a broad label for the overall population. (see Table 1 below)

In other words, the FDA was more restrictive, with a biomarker-driven approach, than the EMA when considering labels (see also unpublished figure below).

There is a sad paradox. MAGNITUDE had a better design and informs treatment decisions better than the other trials, leading to a more restricted label, consistent in both US and Europe. In contrast, the two other companies seem to be rewarded wtih broader labels despite the design flaws and despite ignoring recommendations by the agencies.

Conclusion

Our work highlights how inconsiscies in trial designs may lead to different conclusions. An optimal trial would prospectively determine HRR status using a standardized definition, followed by stratified randomization into biomarker-defined cohorts (e.g., BRCA1/2, non-BRCA HRRm, and non-HRRm). Each cohort would be independently powered to enable definitive conclusions, and to avoid the risk of one cohort driving the benefit of a wider cohort (“adjacent subgroups” design instead of “nested subgroups” design). The control arm should reflect contemporary standard of care, allowing a fair comparison between combination therapy vs optimized sequential use.

Even though no trial is perfect, these funding principles would limit the risk that a trial is ultimately non-informative stemming from its very inception.

See our full paper openly available here.

The efficacy-effectiveness gap

An important difference must be made upfront between “efficacy”: what is seen in randomized trials, and “effectiveness”, i.e. what will happen in the real-life settings. Often, a difference will be observed between those two, which is called the "efficacy-effectiveness gap”.

What is an explanation for the difference? Simple - because patients in trials are different from those encountered outside them, and this is problematic.

Age as a key driver of this gap

Part of the mismatch is due to differences in age and health status between trial participants and the general population with the respective condition. The need for better age representation is especially important in oncology, where the risk of cancer increases as people age, yet exclusion criteria related to health conditions often prohibit older people from clinical trial participation, even though they may commonly be treated with the drug in clinical practice.

The observation that clinical trials enroll younger and healthier individuals than in the general population with a given condition is not new, and in fact, the US Food and Drug Administration (FDA) has issued guidance documents on several occasions to encourage trialists to include a study population that is more representative of the population with a condition. However, differences in age between clinical trials and the real-world may still persist, which is why we set out to assess any potential differences.

Our study - covering more than 20 years of registration trials

To answer this question, we searched for the age of trial participants and tumor types for all drug approvals going as far back as 2002. We noted multiple metrics of the age of trial participants, including median age and the percentage of people above certain age thresholds (>65 years, >75 years, etc.). To get at the real-world age for each tumor type, we used the Surveillance, Epidemiology, and End Results (SEER) data for the tumor type being tested in the registration trials for each approval.

The median age of the general population (as per SEER data) was 67 years of age, and the median age of the trial participants was 62 years of age.

We found that the age gap between registration trials and SEER data has declined over time (7 years during 2002-2012 vs 5 years during 2013-2019 vs 4 years during 2020-2024), but the age gap still persists. SEER ages were older than clinical trial ages for all tumor types except for thyroid and leukemia.

How to improve upon those findings?

One possible solution is to conduct trials specifically in older adults. In our review of trials, we found only 5 trials (of 495 non-pediatric trials) that restricted the inclusion criteria to older adults. Of those, 2 used a cut-point of 55 years of age - an age that is not typically considered as “older”.

Limiting the exclusion criteria in a trial may also be a way to get a more representative population. While comorbidities (often occurring as exclusion criteria) in the study population may complicate the interpretation of study findings, they are also more common as people age. Allowing individuals with comorbidities, unless there is a specific reason to exclude them from trial participation, would not only result in a more representative population, it would provide more realistic information on safety and benefit in the context of conditions and medications that older adults commonly have and take.

In conclusion, while there has been progress on shrinking the age gap between clinical trials and the real world population, there is still room for improvement.

Check out the full paper here!

Note for readers: the work presented in this post was prepared prior to my employment at the US Food and Drug Administration (FDA), and the views expressed here don't represent the views of the FDA.

Still, in few instances, interventions appear to have exceptional effect size. In those cases, with and , we proposed a novel trial design that both maintain a randomization component while being acceptable and ethical for patients. We published this proposal in Trials after a study showed impressive results with dostarlimab – an anti-PD1 therapy – in patients with MSI-high rectal cancer.

How to study the few potential “parachutes” in medicine?

The abstract of our work sets the stage:

“Dostarlimab (Jemperli, GlaxoSmithKline) is an anti-programmed death receptor-1 monoclonal antibody (anti-PD-1) recently tested in a non-randomized, phase II trial (NCT04165772) which included patients with mismatch repair-deficient, locally advanced rectal cancer. Among the first 12 patients treated with dostarlimab, 100% achieved a clinical complete response with no patients experiencing progression or recurrence to date. Most impressive, none required chemotherapy, radiotherapy or surgery the prevailing standard of care. In this paper, we discuss the impressive results of this trial and how they relate to cancer policy, as well as propose a novel trial methodology to assess dostarlimab.”

The study was presented during ASCO 2022 and published the same day in the New England Journal of Medicine. Below are the study schema and key results:

To gain knowledge, incorporate randomization whenever possible

A class-effect? In the case of dostarlimab, everyone wondered if such impressive results were due to the drug itself or whether similar results could have been achieved with any other anti-PD1 therapy. Randomization may answer this very question.

Parachutes themselves have evolved over time, and paragliding today is not exactly similar to jumping with parachutes 50 years ago in terms of physical trauma and risks taken.

One can pre-specify rules – like a number of events – defining the “Skydiving parachute zone” (on the left). In this zone, you don’t want to randomize patients to “no parachute”, however you can use randomization to clarify, for instance, if there is a class-effect or not. Randomization between different “parachutes” – different anti-PD1 molecules – is both reasonable and informative.

If the “Skydiving parachute zone” rules are crossed, then the trial will move to a more conventional “Equipoise area” (on the right) where randomization versus the usual standard of care is again justifiable and acceptable. We illustrated this in the key figure below.

Concluding words

Although randomization is not always appropriate or feasible in medicine, the question of incorporating some degree of randomization into trial design should always be considered. Randomization is a powerful tool for generating knowledge and should not be dismissed, even for interventions or compounds with seemingly very large effect sizes. (our full work is accessible here).

In many trials, placebo are used as add-on to a backbone therapy, trying to isolate the experimental add-on therapy. For instance, you will have trials comparing “chemo + immuno” versus “chemo + placebo”. In our work, we did not include those, we focused on trials where the control arm is “only” an inert compound.

An Ethical Framework for the Use of Placebo

The World Medical Association, in the Helsinki Declaration, describes the Ethical Principles for Medical Research Involving Human Participants. And when it comes to the control arm and the use of placebo, the declaration is very clear:

“participants (…) will not be subject to additional risks of serious or irreversible harm as a result of not receiving the best-proven intervention.”

Wisely, they add:

“Extreme care must be taken to avoid abuse of this option.”

Does progression or death constitute “irreversible harm”?

In our work, we looked at FDA approvals over almost 16 years (2009 to November 2024) and found 19 registration randomized trials in the advanced or metastatic setting using such inert control arm. One was excluded from our final analysis as it didn’t report subsequent therapy, leaving 18 trials, most using placebo-only (17 trials) while one was using “best supportive care”.

We used a simple and powerful proxy: how many patients received a subsequent active therapy after receiving an inert control arm?

We found that overall 49% of control patients got subsequent therapy!

In other words, those patients, with advanced or metastatic cancer, were randomized to an inert substance and later, upon progression, finally received an active therapy, including potentially life-prolonging compounds.

Below are the 18 trials with, in blue, the proportion of control arm patients receiving subsequent therapy (and in orange those from the experimental therapy).

The more subsequent therapy in controls, the higher the “benefit” from the novel therapy.

It is logical that if more patients received subsequent therapy, this indicates that they were more likely eligible for an active therapy upfront rather than an inert control arm. Thus, a high rate of subsequent therapy suggests that a large proportion of patients were genuinely deprived of potentially effective treatment during the randomized phase.

Consistent with this, we found a significant inverse association, with a 0.07 decrease in the hazard ratio (indicating greater magnitude of benefit) for every 10% increase in control patients receiving subsequent therapy.

Even though we pooled PFS and OS in this analysis, “pooling was justified because the core question concerns trial validity under inert controls, hazard ratios guide regulatory approval, and both OS and PFS distortions may arise from the same issue—depriving control patients of active therapy.”

A simple rule to assess the control arm - “Would you randomize your own mother?”

In our discussion, we acknowledge that “some level of post-protocol care in control-arm patients is inevitable”. Also, if a framework is provided by the Helsinki Declaration, this means the use of placebo, in some cases, can be justified.

However, in our work, the minimum rate of subsequent systemic therapy was 22 %, really broadly questioning the use of placebo and inert controls in those settings.

Whenever you consider a control arm, ask yourself: would I be comfortable assigning patients to it? “Would you randomize your own mother?” (from ). If the answer is no, it’s a strong signal to investigate rigorously whether the control is outdated, distorted, or unfair.

Check out the full paper for more insights, and thanks to my co-authors

Raj is a researcher I admire for his evidence-based approach and his precise understanding of clinical trials and disease biology, which allows him to offer key and practical insights to clinicians. Raj is also a friend, and he kindly agreed to talk about smoldering myeloma following the recent FDA approval of daratumumab in November 2025 for “adults with high-risk smoldering multiple myeloma (SMM)”, based on the AQUILA trial.

On social media, Raj summarized his “approach to patient counseling post-AQUILA” in a single slide, to which also contributed (see below). In the video, he walks us through the definition of smoldering myeloma, including high-risk disease, the seminal trials, and how he approaches this in clinical practice. For those interested in the censoring pattern I mentioned in the video (at 28:27), please go to the BREAKING-ICE App and select AQUILA-OS (with more to come on that in the future).

I hope you enjoy the video, including the great Q&A session with Sunny and Milos!

A key question, however, is how many patients derive long term benefit? Much of the research around ICIs has had short follow-up, and while results are impressive for some tumor types, long-term data are not always reported.

Indeed, the number of approvals for this drug class has increased, and the percentages of individuals who are eligible for and respond to these drugs have also increased (full access here, and a key figure below).

Several frameworks, including the American Society of Clinical Oncology (ASCO) Value Framework Tail of the Curve and the European Society for Medical Oncology Magnitude of Clinical Benefit Scale (ESMO MCBS) have been developed to capture long-term benefit. However, there are discrepancies between the frameworks, and they both have limitations. As such, conclusions about long-term data from these frameworks may be limited.

Our new work - assessing long-term survival data

To better assess long-term benefit, we reviewed all FDA-approved oncology ICIs in the metastatic setting (2011-2023) (full access here). Using trial information, we searched for the most recent publication reporting overall survival (OS) for each ICI.

The median duration of follow-up for all drugs was 26 months. Of the 88 drugs evaluated, 20 (23%) qualified for ASCO’s tail of the curve bonus and 27 (31%) did not report median OS. Moreover, 50% of studies did not report OS at 24 months, 68% did not report OS at 36 months, and 89% did not report OS at 60 months. Just under one-third of trials report OS data for patients as long as three years, and just over 10% report OS data at five years. Ideally, we would have a better idea of long-term benefit for all drugs and not just a minority of them.

One take-away from the study is that long-term OS and benefit is limited to a few tumor types, such as melanoma. Hepatobiliary and cervical tumors not only have short follow-up, but they often have small differences in the percentage of patients surviving between the intervention and control arms. Long-term benefit may be the exception rather than the rule when it comes to ICI therapies.

Another take-away is that long-term response appears to be at least partially related to tumor biology, not just the difference in OS provided by the ICI. This was demonstrated by the lack of correlation between the longest time with at least 10% of patients at risk and the absolute difference in survival.

It is important to consider that these numbers reflect clinical trial experience, so real-world benefit may be even less since clinical trials often include younger and healthier patients than patients in the real-world.

In summary, the gathering and reporting of long-term survival data on ICIs should be incentivized to provide information to providers who can assess the value of these drugs for their patients.

Let me walk you through this work and explain why studying censoring patterns across longer follow-ups can help identify where informative censoring may have occurred.

The power of digitized curves: enabling estimates of censored patients

In short, when data sharing is absent or limited, or when the number of censored patients are not reported in Kaplan–Meier curves (after each number-at-risk), one can still use a powerful approach: first digitize the curves, then run code to generate pseudo–individual patient data (pseudo-IPD). These reconstructed data produce curves that closely mirror the originals and allow you to estimate the number of censored patients, among many other aspects worth exploring.

In this new work, I used digitization and the BREAKING-ICE App© to extract the number of censored patients at each time point across different follow-ups from the same trials.

In a trial with immature data, a high rate of early censoring is expected because many patients are still on-trial and are censored at the data cut-off — this is “administrative censoring.” With longer follow-up, these patients won’t remain early censored: they will either present the event or be censored later if they’re still on-trial without an event at the next data cut-off, and this continues with each subsequent update. Conversely, if early censoring does not decline as follow-up increases, the likelihood of informative censoring becomes higher.

The NADINA trial: a reassuring evolution of the censoring pattern

NADINA was a trial testing neoadjuvant immunotherapy (in one arm) compared to adjuvant therapy only, in patients with resectable, macroscopic stage III melanoma. When NADINA was first presented and published, there were questions about the possibility of informative censoring because of the high rates of censored patients — 33% and 27% in the experimental therapy and control groups, respectively. First, and this is commendable, the authors did report the number of censored patients in their publication (see below, numbers in bracket after the number at risk).

Moreover, they addressed the possibility of informative censoring by providing a reverse Kaplan–Meier plot in their supplementary appendix (more on this type of analysis here).

Based on updated data presented at ESMO 2024 and 2025, I was able to plot the censoring patterns with longer follow-ups. It clearly shows that the high early numbers decrease with longer follow-up, which is reassuring regarding informative censoring.

The NATALEE trial: a different pattern

NATALEE was a trial investigating the use of 3 years of adjuvant ribociclib in patients with early breast cancer. We and others have raised concerns about the possibility of informative censoring (also detailed here).

The censoring pattern is now evolving very differently with longer follow-ups. There is an area where high and early censoring remains stable, allowing us to identify an “informative censoring area” — here up to the first 18 months — where informative censoring likely occurred.

Of course, this does not mean it couldn’t occur later; however, the identification of an “informative censoring area” in NATALEE reasonably justifies conducting sensitivity analyses within the period where informative censoring most likely occured. This relevant window extends beyond 6 months—a period chosen in a recent publication—and restricting analyses to such an early interval may miss a substantial part where informative censoring could have occurred.

Significant early censoring? Wait for longer follow-up.

The main takeaway point is that when facing significant early censoring, longer follow-up may help distinguish administrative censoring from informative censoring. Regarding informative censoring, the more mature data in NADINA are reassuring, but this is not the case in NATALEE. Another way to address this issue—even with earlier follow-up—would be real data-sharing, which is often limited or absent.

“(…) when assessing trial outcomes, health technology assessment bodies or regulatory agencies may request longer follow-up to better determine the extent to which an early censoring pattern is likely to be informative.”

You can read the full paper here — it’s open access!

I invited him to discuss a recent paper he co-authored with Gilbert Welch and led by Vishal Patel. The study, published in JAMA Internal Medicine, is titled “The Rise in Early-Onset Cancer in the US Population—More Apparent Than Real.”

In our conversation, Ade walks us through the background of the research question, the key findings, and the open questions that remain. What follows is a live discussion between us, joined by core members of the VKPrasadLab and other guests.

A gem from this interview is this final quote from Ade:

“Whenever somebody tells you that the incidence of something is going up a lot, ask about the outcome of interest, like mortality, especially as it relates to cancer”

I hope you enjoy this rare and thoughtful exchange on a crucial topic: how to understand and contextualize incidence numbers within the broader population.

(free link to the paper here)

In such cases, a classic situation is when the disease remains stable for a long period of time before progressing. In those situations, it remains difficult to know whether this apparently positive outcome is occurring simply because the patients were highly selected—with spontaneously slow-growing tumours and indolent biology—or whether the outcome is actually related to the therapy under examination.

One way to distinguish between these possibilities is to compare, for the same patients, their outcomes before and after the therapy. This is the core idea behind the Iceberg Plot, which was designed and published in The European Journal of Cancer soon after I joined the VKPrasad Lab.

The swimmer plot - no prior therapy data

In Swimmer plots, as the one below (from this publication), you have no data on prior therapy. Each patient has a “lane”, and the lane is beginning with the novel therapy.

Comparing new therapy data to prior therapy.

We reanalyzed data from a JAMA Oncology report of a cohort study involving 1138 patients with advanced solid tumors. Next-generation sequencing (NGS) was successfully performed in 1015 patients (89.2%), identifying potentially actionable alterations in 817 (80.5%). Among these, 132 patients (16.2%) received sequencing-directed therapy (SDT), of whom 49 (37.1%) experienced clinical benefit and 26 (19.7%) achieved responses lasting 12 months or longer.

Notably, the study reported prior therapy data for 125 of the 132 patients treated with SDT, enabling the construction of the Iceberg Plot. In this plot, the new therapy (here, SDT) appears above the ocean line, while the time prior to it represents the “invisible” part of the iceberg.

This visualization immediately contextualizes the results, offering a clearer understanding of treatment responses in relation to patients’ prior therapies.

Below are two figures summarizing the steps required to transform a classical Swimmer Plot (shown on the left in the first figure) into the final Iceberg Plot (second figure).

Von Hoff criteria and growth kinetics.

The central question is not only whether the tumor grows, but whether it grows faster or slower under a given therapy — in other words, this concerns tumor growth kinetics.

In this re-analysis, we also assessed how many patients met the Von Hoff criteria, a metric proposed by Daniel D. Von Hoff and colleagues. According to this criterion, a patient is considered to derive clinical benefit if the progression-free survival (PFS) on the new therapy is at least 1.3 times longer than the PFS on the prior therapy. In the analyzed cohort, the proportion of patients meeting this criterion was remarkably low, with only 2 of 97 patients (represented by the red points in the figure below) satisfying it.

Key point

When interpreting outcomes such as PFS or other composite measures that include stable disease (like the disease control rate – DCR) from uncontrolled data, one should always consider whether the observed outcomes reflect a rather indolent tumour biology or are related to the therapy itself.

In such cases, assessing growth kinetics—for instance, by comparing outcomes with the prior line of therapy—can be informative.

Instead of Swimmer Plots, Iceberg Plots should be more widely used in oncology trial reports. (full paper here)