Forecast the readout before the trial runs.
OVIVO reconstructs a clinical trial in silico — the full protocol, a virtual population matched to the patients you'd actually enroll, every one run through the same mechanistic chain — and rolls the result up to the trial's own endpoints. Then it commits to a call before the trial reads out, and checks it against reality when it does.
The call is made blind, before the outcome is known — so it can't be reconstructed after the fact, and it can't be faked.
Most trial forecasting is a guess about resemblance. We forecast from the biology that decides the result.
The usual approach asks how a trial statistically resembles past trials — the indication, the phase, the endpoint, the company's history — and forecasts the readout from the pattern. It can be right, and it can be right for the wrong reasons. It can't tell you why a trial will succeed or fail, and it can't tell you what to change, because it never models the thing that actually moves the endpoint: the drug acting on the disease, in these patients.
OVIVO does. We build the trial from its mechanism — exposure, engagement, pathway, disease, and safety — across a population of distinct virtual patients, and derive the endpoint from the chain. The forecast arrives with a reason attached, and a lever you can pull.
A trial, reconstructed from protocol to readout
Each station takes the one before it. A protocol becomes a population; the population is run patient by patient through the full causal chain; the individual outcomes roll up to the endpoints the trial prespecified — and the whole thing carries its uncertainty to the end.
-
1
Protocol & endpoints
Receives the trial design.
Treatment arms, dose and schedule, inclusion and exclusion criteria, the primary and secondary endpoints, follow-up duration, and the analysis plan — encoded exactly as the protocol specifies, so the simulation is judged against the trial's own definition of success.
Sources (selection)ClinicalTrials.gov · published protocols · FDA labels · EMA SmPC
-
2
Virtual population
Receives the eligibility criteria.
A cohort of distinct virtual patients whose age, sex, comorbidity distributions, organ function, and pharmacogenetics are calibrated to published baseline tables for the indication — not one average patient, but the spread of real ones a site would enroll, including the outliers that decide a trial.
Sources (selection)published baseline tables · NHANES · gnomAD · disease registries
-
3
Per-patient causal chain
Receives each virtual patient and the regimen.
Every patient is run through the full arc — pharmacogenomics, exposure, target engagement, pathway, disease progression, and safety — the same connected simulation chain that models a single therapy, now executed across the whole cohort.
Sources (selection)PharmVar · ChEMBL · Reactome · DrugBank
-
4
Endpoint aggregation
Receives every patient's trajectory.
Individual outcomes are rolled up to the trial's actual endpoints, arm by arm — response rates, change from baseline, time-to-event — using the prespecified statistical model, so the output is the quantity the trial was designed to report.
Sources (selection)prespecified analysis plans · published endpoint definitions
-
5
Confidence propagation
Receives the outcome distribution.
Uncertainty is carried from every parameter through to the endpoint, so the result is a forecast carrying an honest interval and an evidence grade — never a point estimate dressed as a certainty. When the evidence won't support a call, the platform says so and declines to make one.
-
6
The verdict, on the record
Receives the forecast.
A direction, a magnitude, and a hit-or-miss call against the primary endpoint — committed before the trial reads out and timestamped, then checked against the result when it lands. The forecast is fixed before the answer exists.
Sources (selection)ClinicalTrials.gov readouts · published results · regulatory filings
A forecast, not a postdiction
It is easy to build a model that explains a trial after it has read out — the field is full of them. The only test that means anything is the one made in advance. Every call here is committed before the outcome is known, so it can be checked, and so it can be wrong.
Two ways to call a trial
The field's standard way of forecasting a trial and ours differ at the root: one reasons from resemblance, the other from mechanism. We say it once, plainly.
Forecasting by analogy
- Forecasts the readout from how a trial statistically resembles past trials — indication, phase, endpoint, sponsor history.
- Blind to the biology: it can be right for the wrong reasons, and gives no account of why.
- Offers nothing to change — no dose, no cohort, no endpoint it can point you to and improve.
Forecasting from mechanism
- Models the drug acting on the disease, patient by patient, and derives the endpoint from the causal chain.
- Every forecast carries a reason and its provenance — you can follow the logic from genome to readout.
- Because it's mechanistic, it's actionable: change the dose, the cohort, or the endpoint and the forecast moves with it.
A model trained on trial outcomes can only echo the past. A mechanistic engine can be checked against a trial it never saw.
We don't ask the field to believe a story about mechanism. We keep a scoreboard.
Across a large and growing set of real trials, the platform makes blind, pre-readout calls — committing to a direction, a magnitude, and a verdict in advance — and we keep a continuously audited record of how they land. The record is bound to calendar time: it can't be reconstructed after the fact, and it can't be reproduced by anyone who wasn't making the calls before the outcomes were known.
We hold our highest-confidence calls to the strictest bar. A confident call on a trial that goes on to fail is treated as a failure of the whole system — surfaced, investigated, and understood before anything else moves. When we tell a partner what we think their trial will do, we can show them how often we've been right when we said the same about someone else's.
The discipline is the credential.
Decisions you'd rather make before the spend
The same reconstruction, pointed at the choices that decide a program — each answer carrying the chain of reasoning behind it, so you can defend it to a board or a regulator.
Protocol optimization
Compare doses, schedules, endpoint definitions, and durations before you commit. A trial that would have failed on an inadequate dose or a misaligned endpoint can be corrected before the first site opens.
Futility detection
If the model forecasts an effect too small to clear the primary endpoint at the planned sample size, that's worth knowing before the Phase IIb spend. Not every such trial should be stopped — but some should be redesigned.
Patient enrichment
Define the biological criteria that forecast response and sharpen enrollment toward the patients most likely to benefit — more signal for a smaller cohort, without discovering a failed enrichment strategy mid-trial.
Subgroup & stratification
Simulate differential response across genetically defined subpopulations, comorbidity strata, and demographics. See which segments drive the effect — and which carry elevated risk — before the stratification is set.
Competitive landscape
Reconstruct a competitor's trial architecture and read what its compound is likely to produce — mechanistically, not by analogy — so the landscape is understood before the readout, not after.
Rare & orphan disease
Where populations are small and natural-history data is thin, individual-level simulation generates meaningful signal that conventional trial design can't — and it doesn't need a large published trial to anchor its reasoning.
Trials we've reconstructed
A selection from a large and growing validation library — published trials rebuilt in silico and checked against their reported outcomes. Names you'll recognize, spanning oncology, cardiometabolic disease, neurodegeneration, immunology, and rare disease.
Each is a trial the platform did not train on — which is exactly what makes the comparison worth anything.
A selection · the library grows every week · validated against reported outcomes