How the work is done and checked
Written for the scientific evaluator and for QA.
Verification by construction
Each mechanism, mapped to ICH M15.
| What we do | How | M15 expectation it serves |
|---|---|---|
| Every run logged | Inputs with checksums, code version (Git commit), software version and outputs, recorded automatically | Traceability of the analysis; code and data available for review |
| Reproduction | Each reported run is re-executed from the delivered files in a clean environment and compared with pre-set tolerances | Reproducibility of reported results |
| Code review | Model code is checked against the MAP specification by a second pharmacometrician | Verification of model implementation |
| Dataset double programming | A second derivation program, written independently from the specification, is compared with the first record by record | Verification of data handling |
| Human sign-off | Each discrepancy is resolved and signed off by the lead pharmacometrician. The verification log is delivered with the report | Documented verification, reported in the MAR |
run 014 2026-03-04 10:12 nlmixr2 · FOCEi commit a41c9e2 inputs adppk_v3.csv sha256 9f2c…e81 model run014.R result OFV −2481.36 condition number 41.2 converged verify re-executed in clean env ΔOFV 0.00 ✓ dataset re-derived from SDTM 0 discrepancies ✓ decision D-07 retain allometric WT on CL signed RB
Where AI fits, and where it does not
- Our agents are custom-built for pharmacometric work and run locally on our own servers. Client data never goes to external AI services.
- All estimation and simulation runs in NONMEM, nlmixr2 or PK-Sim, in pinned environments, from code anyone can re-run.
- No estimate, simulation or statistic in a report comes from a language model.
- A person makes every scientific decision, logged with its rationale, and signs every deliverable.
Relationship to FDA's draft guidance on AI
FDA's January 2025 draft guidance on AI in regulatory decision-making proposes a credibility assessment for AI models whose output supports decisions on safety, effectiveness or quality. It is a draft.
Here, the evidence comes from conventional models in established software, not from an AI model, and independent verification is designed to catch any effect of agent assistance. Each report states where agents were used; whether a use is in scope is for the agency to judge.
Modeling standards
Defaults, adapted to each question in the analysis plan.
| Estimation | FOCE with interaction as default for continuous data; SAEM or importance sampling for complex models; Laplacian for categorical and time-to-event data |
| BLQ handling | Likelihood-based (M3) when the BLQ fraction is material; simpler methods otherwise, with a sensitivity analysis |
| Covariates | Pre-specified, physiologically motivated candidates. Full covariate model or stepwise search (forward p < 0.05, backward p < 0.001), chosen in the MAP |
| Model selection | Likelihood ratio test or information criteria, diagnostics, parameter precision, shrinkage and scientific plausibility |
| Diagnostics | Goodness-of-fit plots, VPC or prediction-corrected VPC, NPDE where informative |
| Uncertainty | Nonparametric bootstrap or sampling importance resampling (SIR), propagated into every simulation used for a decision |
The full standards document is available on request.
Software and environment
NONMEM and PsN; R with nlmixr2, rxode2 and PKNCA; PK-Sim and MoBi (Open Systems Pharmacology). The exact versions are recorded in every report.
- R package versions locked with renv; environments pinned as container images
- All code, configuration and reports under Git, with the commit recorded for each run
Submission standards
- Analysis datasets following CDISC ADaM conventions for population PK (ADPPK)
- define.xml and a reviewer guide for every dataset
- Datasets in SAS transport (XPT) format
- Reports placed for eCTD Module 5.3.3.5, with cross-references
- Code delivered as plain text files, readable without our tools
Risk-proportionate review
Independent senior review is included whenever model risk is high, and optional otherwise, at a fixed fee quoted on the call.
What we do, and what partners cover
- NCA and TFLs for study reports and CSRs
- Population PK and PK/PD
- Exposure-response, including concentration-QTc
- Mechanistic models: target-mediated disposition, disease and biomarker models
- PBPK
- Virtual clinical trials and trial simulation
- MIDD strategy and regulatory documents
- Clinical and medical input
- Biostatistics
- Regulatory affairs and publishing
Want to see it on your question?
A 15-minute call with the pharmacometrician who would do the work.