We predict the endpoint of a trial as it is designed, and show the biology that gets it there. For the team running the trial, the team watching a competitor run one, and anyone weighing a deal on the result.
By the time a Phase 3 reports, the choices that decided it were locked in at design, when phase-to-phase success still sits near 20%. Four in five Phase 3 combinations fail, and the levers that could have changed the answer closed long before anyone saw a curve.
Thin signal, large commitment. A Phase 3 gets green-lit on early data that rarely shows how a combination behaves at scale.
The failure mode is set at design. Get the line, the population or the dose step-up wrong and no amount of execution recovers it.
Resemblance is a poor guide. Two trials of the same drugs in neighbouring diseases read out in opposite directions, and standard analysis cannot see why.
Built for Clinical Development · Biometrics · Market Intelligence · BD & Licensing · Search & Evaluation · Investment DD
Clinical Development · Biometrics
Predict the endpoint for the design you are about to commit to, then find the version of it that clears on fewer patients and in less time. Includes the combinations and sequences with no precedent to interpolate from.
Market Intelligence · Portfolio Strategy
Which competitor read-outs will change practice and which are noise, what a rival's combination does to your franchise, and the benchmark your own assets have to beat on the day they launch rather than today.
BD&L · Search & Evaluation · Corporate Development
A verdict on the asset itself: the setting where it wins, what it needs beside it to get there, and whether it still beats standard of care by the time it reports. Diligence that does not rest on the seller's framing.
Venture · Crossover · Hedge funds
The same read applied to a ticker rather than a term sheet. Query the model programmatically over MCP to run it across a watchlist or a whole fund's pipeline, and refresh as catalysts move. Scientific diligence, not investment advice.
Every prediction decomposes into the biology and the trial design that drive it, so a committee can take the reasoning apart rather than trust the number.
Adding LAG-3 blockade to the PD-1 backbone in first-line melanoma moves the biology only marginally beyond the active comparator. Recognition-dependent kill is already close to saturated in most responders, so the combination sits near today's standard of care.
The model therefore expects a real but small PFS gain, smaller than the trial is powered to detect against pembrolizumab. The resistant fraction persists and the curves stay too close to clear significance.
A PASS or FAIL and a confidence for each, timestamped on @bigpicturebio before any of the trials reported. Twelve landed. Three of them are worth walking through.
ATTRACTION-6
1L gastric · nivo + ipi + chemo
Our most confident call of the meeting, and a FAIL. The stock read was a win, because the combination had already beaten chemotherapy in gastric cancer. But that benefit is carried almost entirely by the PD-L1-high minority. Dilute it across an all-comers population, add the toxicity of a second checkpoint, and the curves converge. They did: 15.7 months against 15.8.
Called correctlyHARMONi-6
1L squamous NSCLC · ivonescimab
A PD-1 and VEGF bispecific winning in Chinese trials, with the obvious worry being whether it transfers. Much of that is answerable from single-cell data: the tumour state the drug engages is present in both Western and East-Asian lung tumours, and if anything the East-Asian tumours are slightly more immune-suppressed. Interim OS came in at HR 0.66.
Called correctlyRASolute-302
2L metastatic pancreatic · daraxonrasib
A pan-RAS inhibitor in second-line pancreatic cancer, a setting that has defeated almost everything brought to it. The call turned on the drug closing the pathway broadly enough that the usual escape through a neighbouring RAS isoform is not available, in a tumour whose dependence on that pathway is close to total. Median OS came in at 13.2 months against 6.7, an HR of 0.40.
Called correctlyWe post the misses in the same place as the calls. Each one becomes a question the model asks on every subsequent trial, and a correction only counts if it improves calls on trials it was not derived from. A fuller write-up is coming.
A trial you are designing, a competitor catalyst, or an asset you are weighing. The model runs in seconds, so we can usually turn an initial review around within hours rather than the weeks a diligence process normally takes.
What comes back is a predicted endpoint, the biology behind it, and the one thing we would measure next. If it is useful we will talk about the rest.