Skip to contents

Patch Pharmpy's nlmixr execution backend

patch_pharmpy_nlmixr_results.Rd

Applies Python monkey-patches to pharmpy.tools.external.nlmixr.run, the module every nlmixr candidate fit dispatched by a Pharmpy tool (bootstrap, modelsearch, covsearch, iivsearch, amd) goes through. Three upstream bugs make it fail for essentially every real dataset, so those tools are unusable against nlmixr-format models without this patch (InsightRX/pharmr.extra#121):

Usage

patch_pharmpy_nlmixr_results()

Value

NULL, invisibly. Called for its side effect on the Python session.

Details

  1. rdata["thetas"].loc[get_thetas(model).names] — Parameters$names is a tuple in Pharmpy >= 2.0, and pandas reads a tuple passed to .loc as a multi-axis indexer, so this raises IndexingError: Too many indexers. (In Pharmpy 1.8 names is a list and the line happens to work.)

  2. predictions.set_index(model.dataset[model.dataset["DV"] != 0].index) — nlmixr2 returns one prediction row per observation record, but this indexes the dataset by "DV is non-zero". Any dataset with an observation whose DV is exactly 0 (BLQ imputed to 0, a baseline 0 sample, ...) therefore has fewer index labels than prediction rows and pandas raises ValueError: Length mismatch.

  3. execute_model() writes the candidate's dataset with write_dataset() / write_csv(), which names the file after the datainfo path, but the R script it generates reads <model name>.csv. Whenever the model carries a datainfo path — which every modelsearch candidate does, since they are derived from the input model — the two names differ and the R run dies with cannot open file '.../<model name>.csv', leaving no results for the tool to rank.

The patch fixes (1) by wrapping the get_thetas that run.py imported so names is always a list, (2) by handing the parser a model whose DV column is 1 on observation records and 0 elsewhere — so the upstream DV != 0 mask selects exactly the observation rows, with the dataset's own index labels preserved (observation records are identified from EVID / MDV; masks are tried in turn and the first one the parser accepts wins, and if none work the unpatched call is made so the original Pharmpy error surfaces) — and (3) by wrapping the dataset writer so that, when it is handed a directory, it writes <model name>.csv: the name the generated R script actually reads, keeping the caller's force so a leftover file in a reused run folder still raises.

Because upstream only accepts a mask whose row count matches the number of prediction rows, a mask that parses matched on count alone. If a different mask selects the same number of (different) rows, the choice is ambiguous and a RuntimeWarning is emitted rather than returning silently misaligned predictions.

The patched parse_modelfit_results is installed both on pharmpy.tools.external.nlmixr.run and on the pharmpy.tools.external.nlmixr package alias — the latter is the binding pharmpy.tools.read_modelfit_results() resolves, so without it reading an nlmixr fit outside execute_model() stays broken. It also tolerates the strict= keyword that alias passes but run.py does not accept.

Idempotent, and a no-op on a Pharmpy release that has fixed these bugs. Called automatically by call_pharmpy_tool() for nlmixr-format models.