63 Runs and No Convergence: When RsNLME Succeeded Where Others Couldn't

A pharmacometrics team tried to develop a population PK/PD model of drug-induced neutropenia using a published framework with 16 ordinary differential equations incorporating GCSF dynamics, feedback loops, and maturation compartments, against a dataset of approximately 55,000 records. On a conventional estimation platform they ran 63 attempts using methods including FOCEI, SAEM, and various ODE integrators: SAEM would begin optimizing and then rapidly collapse to zero with a flatlined objective function, FOCEI produced integration errors, and only the less-sophisticated FO method succeeded but was inadequate for the model's complexity. Migrating the model to RsNLME, Certara's R-based NLME platform, using the QRPEM algorithm and LSODA integrator, it converged on the first attempt, produced standard errors, and all goodness-of-fit plots looked appropriate, with a runtime of 12.5 hours -- a result that could not be replicated when exported back to the original platform.

The customer is anonymised in Certara’s own case study — not withheld by us.

Certara is a biosimulation software company (Simcyp, Phoenix) whose PK/PD and QSP models predict drug behavior in the body; the FDA itself licenses Certara's software to review sponsor submissions.

This sits in the Build stage of the AI use-case lifecycle: Named examples from the foundation-model labs and platforms pharma teams use to actually construct, train and generate the model or application.