jinko-task-from-nonmem

Jinkō Task: From NONMEM

Convert a NONMEM run into a working Jinkō trial.

PREREQUISITE: This skill needs an initialized jinko-sdk connection and an SDK satisfying its metadata.requires_sdk range. Run the jinko-sdk-setup skill (../jinko-sdk-setup/SKILL.md) and proceed only once its check passes. If that skill is not found, install it from novainsilico/jinko-skills. This skill additionally needs the converter's extra, which the SDK does not install by default: ``bash pip install "jinko-sdk[nonmem2jinko]" The converter is the nonmem2jinko package. It ships inside jinko-sdk and is imported separately. The extra adds scipy, which the numerical check integrates the reference solve with. Without it every script still runs, and --check solve` raises a message naming the extra.

Core Rules

• Run the scripts. Never hand-translate $PK; only the scripts' output is verified.
• Always check the numbers, and report the measured agreement. Never claim a

conversion succeeded without one, and never publish a report whose verdict has no result in it.

• Say which reference the numbers came from. --reference scipy isolates the

platform; rxode2 covers the reading too; nonmem compares against a $TABLE the original run wrote and needs no licence, only that file. They are not interchangeable. The rxode2 numerical reference currently refuses models needing covariate overrides rather than comparing unequal inputs.

• Blocking error means stop. Do not pass --allow-issues without telling the

user what is lost.

• Read the report before applying. Units and names are inferred, and some

inferences will be wrong for a given data set.

• Wrong time unit changes the kinetics silently. Jinkō accepts arbitrary time

units but solves and returns results in seconds. Confirm amount, volume and time against the data set.

• Compare on the platform's own time grid. Time is a requestable series;

request it. A recorded event inserts its own points, so a grid rebuilt from tMin/tStep pairs values with the wrong times.

Scripts

Project writes are dry-run by default and gated by --apply. Local artifact outputs are explicit path arguments and preserve existing files where their scripts expose --overwrite.

Orchestration

1. Ask for .ext, .phi and the data set. Each improves the conversion; the

control stream alone is enough to start.

1. Convert dry-run. Resolve what the report raises: --time-unit,

--amount-unit, --volume-unit, --unit SYMBOL=unit, --rename NONMEM=jinkoId.

1. Apply, then run.
2. Check numerically using the same --time-unit as conversion. A skipped solve

is a failure, not evidence. Report the number and its reference.

1. Render and offer an equivalence report whenever anyone other than the person

running the conversion will rely on the model.

Every write step takes --folder NAME --create-folder, which keeps one run's items together, and --json-out FILE, which is how the next step gets the SIDs. Do not scrape them out of the prose.

--parent-folder NAME nests that folder inside another one. Use it when a project accumulates several conversions: a folder named for the day holding one subfolder per model keeps each run's evidence together, and the runs sort chronologically. Without it a project's root fills up with model names and nothing says which run each belongs to.

Decisions

Population (convert_trial.py). --population design (default) is editable in Jinkō afterwards. phi replays the fitted subjects. sampled draws from the full multivariate normal represented by $OMEGA; a finite sample's realised covariance has sampling error and must be measured rather than called exact.

Random effects (both converters). --vpop-mode etas (default) represents the target $OMEGA covariance through ETA marginals and correlations. parameters puts marginals on the derived parameters instead: it reads better in the UI but is equivalent only without covariate effects.

Dosing (convert_trial.py). --dosing auto (default) uses arms when the data set has an arm structure and per-patient slots when it does not. A study where every subject has an individualised history — neonatal phenobarbital, weight-banded single doses — has no arm structure, and per-patient puts each patient's own dose times and amounts in the vpop. Needs --population phi or sampled, because the schedule rides on the patient table.

Uncertainty (convert_trial.py). --eps-clones N gives a per-patient predictive cone. Needs --population phi or sampled. Not a VPC — see references/residual-error.md.

Refused, with the record named: steady-state (SS) dosing, $MIX, nested random effects, $OMEGA/$SIGMA as SD, correlation or Cholesky, $PRED-only models, abbreviated code that cannot be read (MPAST, a variable ETA() subscript, CALL, include), a parameter the kinetics need that reads a record-level data item such as DV or EVID, and IF(AMT.GT.0) TDOS = TIME — which is refused precisely because substituting simulation time for it looks right and silently gives zero time-since-dose.

Reported, not refused: $PRIOR and $NONPARAMETRIC shaped the estimation, not the forward model, so the estimates convert as given.

Converted with a stated approximation: DOSE = AMT points at the model's dose parameter, so a subject whose doses differ in amount reads only the first; time-varying covariates keep their baseline only; a covariate selecting which random effect applies cannot be one vpop marginal; a combined error model written through a weight (Y = F + W*EPS(1), W = F*θ + θ) is diagnosed with its coefficients but the observable is emitted without residual error.

References

• references/workflow.md — commands, flags, worked sequence
• references/validation.md — the numerical checks and what each proves
• references/reporting.md — equivalence reports, data tables, overlays
• references/conversion-map.md — what each NONMEM construct becomes
• references/residual-error.md — $ERROR, $SIGMA, the clone approximation

Generic mechanics belong to the lower-level skills: jinko-model for components, jinko-vpop for populations, jinko-protocol for arms, jinko-trial for runs, jinko-data-table for tables, jinko-trial-viz for visualisations, jinko-document for documents.