Loading
12,029 dated facts read out of 19,371 records and sorted into 174 years, from 1746 to 2033. Each year has its own page listing what happened in it; this page is the shape of all of them together, and what that shape shows.
One column a year from 1900 on, as tall as the number of dated facts the corpus holds for it. Every column is a link to that year. The shape is as much the shape of what has been read into OnCo as of what happened, which is the first thing the findings below try to measure.
Before 1900 the record is a scatter of landmarks rather than a year-by-year reading: 40 years carry 53 entries between them, the earliest 1746.
Eleven questions asked of the dates once they are in one shape. Every figure carries the number of records it was computed over. A question answered over fewer than 30 cases, or contradicted by the shape of the data, is marked as one the corpus cannot support, and the working is shown anyway so the thinness is visible.
It is lengthening, not shortening, and part of that is arithmetic. The median is 20 years across 140 targets whose first approval came after the protein was described, and it rises steadily with the decade of approval: n=34, median 14 years (12 years to 17 years) for approvals in the 2000s, n=55, median 22 years (19 years to 26 years) in the 2010s and n=35, median 29 years (26 years to 32 years) in the 2020s. On 24 more targets the drug came first, by as much as 51 years: methotrexate, mercaptopurine, the steroid receptors and the hormone therapies were in use for decades before anyone had the gene they act on.
Denominator. 164 of the 1,674 targets carry both a sourced first-description year and a product with an approval; 1,654 targets carry the year at all.
What it does not show. The year is the earliest paper UniProt cites for the protein's sequence, so it dates the molecule being in hand, not the biology being understood or anyone thinking of it as a target. The widening is also partly forced: an interval cannot be longer than the time since the sequence was published, and 64 per cent of these sequences are from the 1990s, so recent approvals have more room to be slow than old ones ever had.
The median is 5 years, and the middle half runs from 4 years to 7 years. Phase makes almost no difference to it: n=260, median 5 years (3 years to 7 years) for phase 2 and n=767, median 5 years (4 years to 8 years) for phase 3. What does move is when the trial opened, and that movement is mostly an artefact: trials that opened in the 1990s took 16 years, those that opened in the 2010s 5 years and those that opened since 2020 3 years, because a trial that opened in 2022 can only appear here if it has already finished.
Denominator. 1,142 of the 5,969 trials carry both a registry start date and a reported year; 5,821 carry a start date at all.
What it does not show. This measures trials that reported, which is the whole selection problem in one sentence: a trial still running, abandoned or never published has no reported year and is absent, and the more recent the start, the more of them are missing. The start date is the registry's, the reported year is a year, so every figure is a difference of years and not of days. Every start date counted here is one the registry marks as actual or leaves unmarked; none is still a plan.
The median is 0 years, and 81 per cent of these readouts have an approval in the same year or the next one. A registrational readout and the decision that follows it are two steps of one process rather than two separate events, and in a corpus that records years rather than days they usually land on the same one.
Denominator. 708 of the 1,277 trials with a reported year have a product approved in or after that year.
What it does not show. This is a measure of which trials the corpus reads. OnCo holds landmark and registrational trials, so the trials with an approval behind them are over-represented and the trials that reported and led nowhere are under-represented. The figure is the interval for trials that did lead somewhere, not the chance that a readout leads anywhere.
It has not moved in the period the corpus covers. The median gap between the latest trial behind a guideline version and the version's own date is 1 year across all 64 versions, 1 year for versions dated 2019 or earlier (9 versions) and 1 year for 2020 onwards (55 versions). The earlier group is too small on its own to carry a trend.
Denominator. 64 of the 69 dated NCCN and ESMO versions in the corpus name at least one trial with a reported year; 5 name none.
What it does not show. The guideline history is curated by anchoring each version to the dated event behind it, usually an approval or a publication. That rule pulls the two dates together by construction, so this measures how the corpus records guideline changes at least as much as how fast guidelines move.
The median is 1 year. 40 of the 172 products carry the same year on both sides and 20 carry an earlier European year than an American one, so the two systems are closer together than the usual telling suggests.
Denominator. 172 of the 1,088 products carry both a United States and a European Union approval year.
What it does not show. Approval years here are the year of a product's first approval in each region, not of the matching indication. A product approved in America for one cancer and in Europe for another counts as a pair, which shortens the gap.
The median is 3 years, over a range of 1 year to 22 years, and the middle half runs from 2 years to 7 years. It is the widest spread of any interval on this page: the England row is a recommendation for use in the health service rather than a licence, and it follows a licence by a year for some products and by two decades for others.
Denominator. 35 of the 1,088 products carry both an England (NICE) row and an approval year elsewhere.
What it does not show. NICE rows are recorded on products a deep dive has been written for, above all the cancers with a United Kingdom pathway page, so the set is not a sample of NICE's work.
138 of the 173 resolved indications converted to traditional approval and 35 were withdrawn, a confirmation rate of 80 per cent, with 65 more still open. Confirmation takes a median 3.1 years and withdrawal a median 3.8 years, so a failure takes longer to admit than a success takes to prove. The open indications have been open a median 2.9 years already, and that figure only grows.
Denominator. 238 accelerated-approval indications on 139 products, dated 1995 to 2026, generated from all four of the accelerated-approval tables the FDA publishes, of which 20 are not cancer treatments (supportive care, dosing and formulation); one further indication on the agency's list names a product OnCo does not hold.
What it does not show. This is the American pathway and only the American pathway: the FDA publishes the four lists, no other regulator does, and the conditional approvals of the EMA and the others are not in this count. The rate is taken over resolved indications alone, because counting the open ones as failures would make it fall every time the agency granted a new one; an open indication's clock is measured to the day the table was read, so it is a floor and not a duration.
The median is 11 years, with the middle half between 4 years and 17 years.
Denominator. 80 of the 378 companies with a founding year have a product with an approval dated at or after it.
What it does not show. A product is credited to every company the record links, so an acquired product counts for the acquirer too, which shortens the interval for large companies. The founding years we hold are also weighted towards companies with something to show, so the companies that never approved anything are largely absent.
Everything before 2000 is 11 per cent of the 12,029 dated entries, against 21 per cent of the 2,604 landmarks that editors chose and wrote by hand. Landmarks are the one group picked for mattering rather than fetched, and they are 6 times more likely to sit before 2000 than a paper is. The thin decades are our reading, not a quiet field.
Denominator. 12,029 dated entries across 174 year records, of which 2,604 are hand-written landmarks and 4,840 are papers.
What it does not show. This compares two groups inside the corpus against each other. It shows that the fetched literature is recent and the hand-written history is not; it cannot say how much of the field before 2000 is missing, because there is nothing here to measure that against.
Of 12,029 dated entries, 11,067 are known only to the year, 215 to the month, 21 to the quarter and 726 to the day. 62 per cent of the day-precise entries fall in 2020 or later, because the day comes from the regulatory feeds and the readout calendar, which are the parts of the corpus that are refreshed automatically.
Denominator. 12,029 dated entries across the 174 year records.
What it does not show. Precision here is the shape of the date the source gives, not a judgement of how reliable it is. A landmark known only to the year may be far better sourced than a calendar entry known to the day.
The earliest dated entry is 1746 and the latest is 2033. Between 1900 and the last dated year every year has a record, 12 of them empty; before 1900 the corpus holds 53 entries across 40 scattered years, which is a handful of landmarks rather than a year-by-year reading. 7 years after 2026 carry 68 dated entries, all of them dates a source states for something still to come.
Denominator. 174 year records covering 288 calendar years.
What it does not show. The forward edge is only as long as the last registry completion date and roadmap watch item anyone has written down, so it shortens as those pass rather than as the field changes.
Each year links to its own page, which lists everything the corpus dates to it, grouped by what kind of thing it was. A year with nothing in it keeps its page and says so: leaving it out would read as a claim that nothing happened, when what is true is that nothing has been read into the corpus for it.
13 groups, each one a field the corpus already carried. Nothing here is a new claim: a year is a view of the other records, generated from them, so it changes when they do.
The chart stacks them into four bands: approvals, regulatory decisions, guidelines and law; trials reported; papers; landmarks, technologies, companies and expected dates. Counts are of entries, not of records: a product with an approval in three regions contributes three, and a paper listed on three people's pages contributes one.
The same dated fields, read forwards: drag the year and see what had happened by then. For everything the corpus holds in a single year, with its sources, open that year's own page: the busiest, 2020, with 828 entries; the last the corpus dates anything to, 2033; or all 174 years as a table.
Best effort: for each cancer, the most recent dated history event at or before 2026 that references products or technologies. Not a guideline; see each cancer page for the current standard of care.