The signal
before the trial.

■ 01 · The early signal

The science moves years before the trials do.

Take every pairing of a disease or target area with another concept that eventually entered a clinical trial together, and look at what the scientific record was doing beforehand. Activity in papers and patents starts climbing about six years before the first trial, more than tripling over that window. The official trial registry shows nothing until the trial already exists, and by then everyone can see it.

Based on 72,174 candidate pairings that entered trials, counting only years the record actually covers. The takeaway: the early warning lives in the scientific record, and Intangia reads it there. Waiting for the registry means finding out with everyone else.

■ 02 · Worked examples

Real drug targets, flagged years early.

What an alert looks like in practice. Replaying history in PD-1 immunotherapy, one of the most competitive fields in oncology, Intangia's twenty-partner shortlist made 82 first alerts, and 59 of them (72%) entered a trial with PD-1 within four years, against an area background of about 2%. In CTLA-4 the same replay produced 19 entrants from 53 alerts (36%). Below are nine of those alerts, all recognisable targets or indications.

How to read this: each row is one partner from the replayed shortlists. The filled dot is the year Intangia's ranking first put it in the area's top twenty. The open dot is the year its first trial with the focal target was registered. The bar between them is the head start. Every alert carried its evidence: how much paper and patent activity the pairing had accumulated, and how much clinical precedent the partner already had elsewhere.

Depth is a dial, not a limit: at twenty partners these two shortlists reach 4.8% and 8.6% of everything that eventually entered trials with the focal, and at a thousand they reach 50.5% and 75.7%, with the hit rate falling from 72% to 28% and from 36% to 8%. CD47 is the picture in one row: alerted as a CTLA-4 partner in 2019 on a patent-heavy record, 107 co-occurring patents against 5 papers, four years before its first registered combination trial. These two areas are worked demonstrations of the same system measured across all 100 areas below. They were selected post hoc as familiar examples, and they are favourable ones: PD-1 ranks 3rd and CTLA-4 26th of the 97 evaluable areas on this metric. The panel-wide figures, not these two, are the performance estimate.

■ 03 · The breadth

It works across 100 disease and target areas, not just one.

A method that only works in one field is a curiosity. This one was tested across oncology, immunology, metabolic disease, neuroscience, cardiovascular and rare disease.

How to read this: each dot is one disease or target area. The score answers a simple question: given one pairing that later entered trials and one that didn't, how often does Intangia rank the real one higher? 0.5 is a coin flip. 1.0 is perfect. Every area scores about 0.8 or better.

In 100 of the 100 areas, Intangia also beat the best simple alternative: just counting how much has been published about each pairing. These scores come from the full thirteen-signal specification; the shortlist economics below come from the simpler five-signal ranker Intangia runs in production, which scores 0.915 here and does better at the top of the list.

Every area was scored by a model that had never seen it. For each of the 100, the system learns from the other 99 and is then asked about that one cold. Nothing is fitted and marked on the same area, so these are not scores the method could have memorised.

These twelve were named as recognised targets when the panel was chosen, before any of them was scored — not selected afterwards for looking good. PD-1 is last, and is left in for that reason: it is one of the most crowded fields in oncology, and crowded areas are the hardest to rank.

■ 04 · The payoff

A 170-to-1 search becomes 1 in 4.

This is what triage buys: it tells your scientists and BD team where to spend their attention.

That is a 38-fold improvement over each area's own background rate at the shortlist, and 19-fold across the whole ranking; a matched random shortlist of the same size finds an entrant in fewer than 1 pick in 100. It is not a story carried by a few lucky fields: 94 of the 97 evaluable areas beat their matched random shortlist after false-discovery control, and a panel-level permutation puts the whole result at p ≤ 0.0005. A shortlist buys concentration rather than coverage, and the two have to be read together: twenty names reach 5.7% of everything that eventually entered trials in an area, a hundred names reach 15.9%. Every flag arrives with its receipts: which papers or patents accelerated, when, and by how much, plus how much clinical precedent the partner carries elsewhere. Shortlisted entrants arrive a median of 2 yrs before their first trial.