Introduction

The question

Cancer research produces results every week, and almost none of them is the end of cancer. A trial reports that tumours shrank in more patients on one arm than the other. A blood test finds signals of many cancers at once. A model trained on biopsies predicts who will respond to a treatment. A vaccine made for each patient wakes immune cells in half of those who receive it, one trial reports. Each of these is, to someone, a step forward. Each is also a claim that has to be weighed before anyone acts on it: before a programme spends more money, before a regulator lets a product onto the market, before a field changes what it studies next.

This paper does not ask whether any of these results is real progress. It asks the question Meridian asks of any decision: who must decide, by when, and what would make them decide again? Put in full, the question is this. When should a cancer research programme or a regulator treat a result as progress worth acting on, who decides, by when, and what would make them revisit the decision? The question has a subject, an object, a time and a condition. A result can be acted on by someone who had no duty to check it, measured by a yardstick that does not track what patients care about, accepted with no date by which anyone must look at it again, or left in place after the evidence that would overturn it has arrived.

The paper reads ten public texts on this question: four United States government documents, from the 21st Century Cures Act of 2016 to pages of the National Cancer Institute and the Food and Drug Administration, and six openly licensed scholarly articles. It reads them as a set of answers, each partial, to the same question. It reports where each text places the decision and where it leaves the decision open. It takes no side on any drug, test or programme, and nothing in it is medical, clinical, regulatory or investment advice.

The debate

Four lines of argument meet in these texts. They disagree less about the facts than about where the difficulty lies.

The first line holds that the difficulty lies in the yardstick. A cancer drug is meant to help people live longer or better. Those outcomes take years to measure, so trials often measure something faster that is supposed to predict them: the share of tumours that shrink, or the time until a tumour grows past a threshold. These faster measures are called surrogate endpoints. Robert Kemp and Vinay Prasad, Surrogate endpoints in oncology: when are they acceptable for regulatory and clinical decisions, and are they currently overused? (2017), put the problem in the first lines of their abstract:

They do not argue that surrogates should never be used. They argue for limits: a surrogate should support approval when studies have shown that it reliably predicts what matters, or when a disease is rare, dire and has few treatment options, and in both cases only once the studies that will measure survival are under way. Two further works point the same way by their abstracts. Bishal Gyawali and others, Assessment of the Clinical Benefit of Cancer Drugs Receiving Accelerated Approval (2019), report that confirmatory trials for a fifth of the cancer indications granted the FDA's faster route through 2017 showed an improvement in overall survival. Alyson Haslam and others, Overall survival for oncology drugs approved for genomic indications (2022), report that fewer than half of the randomised studies of survival for such drugs showed an improvement, while nearly all studies of the faster measure were positive. On this line, the question "is this progress?" is mostly a question about what was measured.

The second line holds that the difficulty lies in reach. A result can be real and still touch few people. Alyson Haslam and Vinay Prasad, Estimation of the Percentage of US Patients With Cancer Who Are Eligible for and Respond to Checkpoint Inhibitor Immunotherapy Drugs (2019), estimated, as an upper bound, how many patients with cancer in the United States were eligible for a class of drugs that had drawn wide interest, and how many might respond. They wrote their purpose plainly:

Allan Hackshaw and others, Estimating the population health impact of a multi-cancer early detection genomic blood test to complement existing screening in the US and UK (2021), take the same population view from the opposite side: they estimate how many more cancers a single blood test might find across a whole country, and with what mix of true and false positives. This study reads both papers as holding that progress is a count of people reached, not a property of a molecule or a test, though the two papers reach different moods from that premise.

The third line holds that the difficulty lies in the odds. Chi Heem Wong and others, Estimation of clinical trial success rates and related parameters (2019), traced development paths through trial databases and estimated the chance that a programme which enters the clinic ends in approval; by their estimates most programmes do not, and cancer programmes succeed less often than those of any other therapeutic group they studied. They begin from a claim about decisions rather than about cancer:

On this line, a single result is weak evidence of progress unless it is read against the base rate of results like it.

The fourth line is the programme's. It does not weigh single results; it sets goals and asks the field to meet them. The National Cancer Institute's page on the Cancer Moonshot states the goal in its first sentence:

Elizabeth M. Jaffee and others, Future cancer research priorities in the USA: a Lancet Oncology Commission (2017), by their abstract, expand the recommendations of the panel convened for the Moonshot into further priorities, among them regulation and financing. The National Cancer Plan, on its landing page, frames the goal as shared work:

The goal and the hope here are the Institute's, in its words. This study reads the programme line as answering the question of progress with a schedule and a list of aims, and as leaving the weighing of particular results to others.

Two further articles stand outside the debate and serve as its test cases. Stephen-John Sammut and others, Multi-omic machine learning predictor of breast cancer therapy response (2021), built a model from biopsies and tested it on patients it had not seen. Luis A. Rojas and others, Personalized RNA neoantigen vaccines stimulate T cells in pancreatic cancer (2023), report an early trial of a vaccine made for each patient. Both say, in their own words, what would have to happen before their results could be acted on.

The gap and the contribution

The four lines argue about what is true of endpoints, populations, odds and goals. None of them asks, in the same terms, who holds the decision to treat a result as progress, what time limit binds that decision, and what is written in advance as the reason to reopen it. Those are the questions of a decision record rather than of a scientific review, and they can be put to a statute and a web page as well as to an article.

The paper makes three narrow contributions. First, it shows that the statute answers the question of who decides precisely for money and not at all for results: the Director of the National Institutes of Health spends and reports, committees of Congress receive the reports, and the test written into the law is whether projects advance a strategic plan. The weighing of a cancer result for patients belongs to a regulator to which the statute gives no part in weighing results, and the articles describe that regulator's room for judgment as wide. Second, it shows that the deadlines in the texts fall on programmes and reports, not on results: the statute counts fiscal years and sets an expiry date, the Moonshot set itself five years, and the articles say that the evidence which would settle a result takes years that no text schedules. Third, it collects the conditions for revisiting a result that the texts write down, and finds that they are many and specific, from confirmatory trials to untested populations to unmeasured deaths, while the duty to act on them is seldom assigned.

Chapter 1 says how the texts were chosen and read, and then takes the three questions in turn. An appendix gives further passages of every text, word for word. The conclusion returns to the debate, states the contributions again, and names the limits of reading ten texts from one country.