Insilon · Simulate

Run a drug across a whole population, not a handful of scenarios

Insilon couples PK-PD and physiology surrogates with biomolecular response models, so you can sweep dose, schedule and subgroup in hours instead of weeks.

Example simulationExample

SZ-0412 · 40 mg QD · 10,000 patients

Median AUC
1,240 ng·h/mL
90% interval
610–2,380
Predicted ORR, 12 weeks
31% (27–35%)
Grade 3+ AE
8% (6–10%)
Seed
4417
Reproducible from run recordLogged
Example. Values are illustrative.

Capabilities

What Insilon does

Surrogates

Physics-ML PK-PD

Neural surrogates of physiology and PK-PD systems trained against mechanistic models with Modulus.

Molecular

Biomolecular response

Target, ligand and pathway effects from BioNeMo models feed each patient's response.

Sweeps

Dose, schedule, subgroup

Run the full grid of candidate arms across the same cohort in one job.

Uncertainty

Calibrated intervals

Every efficacy and safety output carries an interval, not a single number.

Reproducible

Deterministic runs

The same inputs, versions and seed give the same result, logged to Evidra.

Scale

Batched on GPU

Tens of thousands of coupled systems per arm, batched with TensorRT and Triton.

How it works

Inside Insilon

  1. Load

    Read the cohort and compound model.

  2. Expose

    Simulate PK across every patient.

  3. Respond

    Simulate PD, efficacy and safety.

  4. Summarise

    Return distributions with intervals.

Specification

Specification

Items marked planned are in design with pilot partners.

ItemDetail
ModelsPK-PD and physiology surrogates; BioNeMo response; outcome models
Scale (target)104 to 106 patients per run
OutputsExposure, efficacy, safety distributions with intervals
ReproducibilityVersioned models, seeds and inputs per run
CoverageSelected therapeutic areas first; wider mechanistic coverage planned
StatusIn development

Questions

About Insilon

How do surrogates compare with our mechanistic models?

Each surrogate is trained and checked against a reference mechanistic model, and its error against that model is reported.

Can we bring our own PK-PD model?

That is the plan: import a NONMEM or Monolix model as the reference and train a surrogate against it. This is planned with pilot partners.

How fast is a run?

The design target is under an hour for a 10,000-patient arm. Actual times depend on model depth and will be measured in pilots.

See Insilon in a briefing

We walk your team through Insilon on a public example, then discuss where a pilot could test it against outcomes you already know.