How it works
From one question to a back-tested answer
A Sanenza engagement starts with one decision your team has to make. We build the population, simulate the options and show how well the model predicted outcomes you already know.
The process
Four steps, about ten weeks
Week 1
Define the question
Agree the decision, the compound, the comparator and the endpoints. Agree which past outcomes we will back-test against.
Week 2–3
Build the population
Cohorta conditions a cohort on indication, region and the subgroups you care about, and checks it against your real data.
Week 4–7
Simulate and search
Insilon runs every arm. Protova searches the design space and Vigila adds safety constraints.
Week 8–10
Validate and report
Evidra scores predictions against known outcomes, then the report states what the model can and cannot support.
Who does what
Clear responsibilities from day one
| Task | Your team | Sanenza |
|---|---|---|
| Choose the decision and endpoints | Owns | Advises |
| Provide trial and real-world data | Owns, under your agreements | Ingests in a secure tenant |
| Build and calibrate the cohort | Reviews | Owns |
| Run simulations and design search | Reviews | Owns |
| Approve any design or text | Owns | Prepares |
| Regulatory interaction | Owns | Supports with documentation |
Configuration
A question written as configuration
Every run is defined in one versioned file, so a reviewer can see exactly what was simulated.
# sanenza.yaml: a dose-selection question (example) question: dose_selection compound: SZ-0412 indication: nsclc_2l cohort: source: cohorta region: [eu, us] size_per_arm: 10000 amplify: [age_75_plus, egfr_variant_rare] arms: - dose_mg: 20 - dose_mg: 40 - dose_mg: 60 endpoints: [orr_12w, grade3_ae, exposure_auc] validation: backtest: [phase1_sz0412, analog_compound_b] review: named_approver_required
Outcomes
What you get at the end
Report
A ranked set of designs
Each design with predicted power, efficacy, safety and subgroup effects, and the reasons it ranks where it does.
Evidence
A back-test you can inspect
How well the model predicted outcomes you already know, with calibration plots and the data behind them.
Assets
A reusable cohort and model
The calibrated cohort and surrogate stay available for the next question on the same programme.
Questions
About the process
How much data do we need?
Enough to calibrate and back-test: usually phase 1 data for the compound, or outcomes from close analogues. We assess this in week 1.
Can we run it without sharing raw data?
Data stays in a single-tenant environment on confidential-computing GPUs. A deployment in your own cloud is planned for 2027.
What if the back-test is poor?
Then the report says so, and says which decisions the model should not be used for. That result is useful too.
Walk through the method before you commit
A briefing covers each step with a Sanenza scientist. If it fits, a pilot takes one decision through ten weeks and a back-test you can check.