Three families of approach. Three different questions they can answer.
The phrase "AI in drug development" hides at least three architecturally different approaches that answer different questions and have different limitations. The choice between them is not a marketing preference — it is a structural property of what the model can compute.
| Property | Family 1 Generative AI Insilico, BenevolentAI, BPGbio | Family 2 ML digital twin Unlearn.AI, Foresight, ETHOS | Family 3 Coupled-physiology ODE Raganelė, Certara, esqLABS, OSP |
|---|---|---|---|
| Trained on | Chemistry, patent, omics datasets | Patient EHR trajectories | Peer-reviewed physiology — not trained at all in the ML sense |
| Predicts | Candidate molecular structures, target activity | What kind of patient like this has historically had next | What this organ system does under this dose, this schedule, this covariate set |
| Counterfactuals? | ✗ — no parameter to perturb | ✗ — confounded by what physicians did | ✓ — change the parameter, re-integrate, read a different trajectory |
| Best for | Early-stage hit discovery, target ID | Trial control-arm augmentation, recruitment forecasting | Dose / schedule, drug-drug interaction, AE counterfactuals, repurposing, formulation |
| Regulatory familiarity | Emerging; case-by-case | EMA qualification for specific uses (recent) | 40-year EMA + FDA track record under PBPK / QSP / MIDD |
| Failure mode | Generates a plausible molecule that does not exist physically; needs wet-lab gate | Confidently confabulates a counterfactual that is not a counterfactual | Calibration drift if parameters are not anchored — diagnosable, fixable |
| Where we work | Not us | Not us | This is us |
Why the distinction matters
The three families are not in competition with each other for the same question. They are good at different questions. The error to avoid is asking one family a question only another family can answer.
Generative-AI families are at their best upstream of compound identification — they propose structures, screen libraries, prioritise targets. Once a molecule is in hand and its mechanism is roughly understood, the question shifts from "what should this molecule be?" to "what does this molecule do in a body?" That second question is a coupled-physiology question. It is what an ODE family was designed for.
ML-digital-twin families are at their best forecasting observed trajectories — recruitment timing, control-arm endpoints, enrolment-rate variance. Once the question is counterfactual — what would the AE rate have looked like at half the dose, or in patients who would have been excluded — the family runs out of mechanism to perturb. A coupled-physiology model has the mechanism by construction.
What we add on top
PBPK and QSP are mature fields. The platforms that anchor them — Simcyp, PK-Sim, MoBi — are excellent. What we add is composition speed: a proprietary computational stack, originally developed for derivatives pricing in quantitative finance, lets the whole composed body run end-to-end at sub-second wall-clock per scenario. This is what makes sweep, counterfactual, and combination-search workflows practical rather than aspirational.
The mechanism is mature. The composition speed is new. The combination is what lets the model answer the questions every pharma team already knows it wants to ask.
See it on your own published case
Pick a trial from your group's portfolio. We replicate it at our cost, send back the result, and discuss what extending to a different cohort or schedule would look like.
Engage →