An estimated parameter is silently dropped from the simulation mapping when the model cannot give it a value

[wshlavacek]

A parameter that the model class leaves out of get_free_parameter_ids_with_values() is dropped from the simulation mapping even when the parameter table says to estimate it. The optimizer varies it, the simulator never sees it, and nothing reports a problem.

This is the silent sibling of #519, which covers the same root cause reached through the condition table, where it crashes instead.

What happens

get_optimization_to_simulation_parameter_mapping seeds the mapping from the model:

simulation_parameters = dict(model.get_free_parameter_ids_with_values())

_apply_parameter_table then walks the parameter table and skips any row that is not already in that mapping:

for row in parameter_df.itertuples():
    if row.Index not in par_mapping:
        # The current parameter is not required for this condition
        continue

For a parameter whose initial assignment cannot be reduced to a number, the seeding step leaves it out, so the skip treats an estimated parameter as one this condition does not need.

Minimal example, on 0.9.0 and on current main:

import pandas as pd
import petab.v1 as petab
from petab.v1.models.sbml_model import SbmlModel
C = petab.C

problem = petab.Problem()
problem.model = SbmlModel.from_antimony(
    "k_base = 1.5; k_derived = k_base + 1; conc = 1; conc' = -k_derived * conc;"
)
problem.condition_df = pd.DataFrame(
    index=pd.Index(["condition1"], name=C.CONDITION_ID)
)
problem.observable_df = pd.DataFrame(
    {C.OBSERVABLE_FORMULA: ["conc"], C.NOISE_FORMULA: [1.0]},
    index=pd.Index(["obs_conc"], name=C.OBSERVABLE_ID),
)
problem.measurement_df = pd.DataFrame({
    C.OBSERVABLE_ID: ["obs_conc"],
    C.SIMULATION_CONDITION_ID: ["condition1"],
    C.TIME: [1.0],
    C.MEASUREMENT: [0.5],
})
problem.parameter_df = pd.DataFrame(
    {
        C.PARAMETER_SCALE: [C.LIN, C.LIN],
        C.LOWER_BOUND: [0.1, 0.1],
        C.UPPER_BOUND: [10.0, 10.0],
        C.NOMINAL_VALUE: [1.5, 2.5],
        C.ESTIMATE: [1, 1],
    },
    index=pd.Index(["k_base", "k_derived"], name=C.PARAMETER_ID),
)

print(dict(problem.model.get_free_parameter_ids_with_values()))
print(problem.x_ids)
print(problem.get_optimization_to_simulation_parameter_mapping()[0][1])
print(petab.lint_problem(problem))

Output:

{'k_base': 1.5}
['k_base', 'k_derived']
{'k_base': 'k_base'}
False

So k_derived is optimized and has no effect on the objective. lint_problem returns no error and no warning is raised.

Why it matters

An optimizer told to estimate a parameter that cannot change the simulation will still move it, so the fit reports a value for k_derived that means nothing, and the parameter's bounds and prior are silently ignored. The larger the fraction of the model that is written with initial assignments, the more of the parameter table disappears this way. Unlike #519 there is nothing to notice: no exception, no log line, no lint failure.

This is not specific to SBML. BioNetGen models have the same shape, where a parameter may be defined by an expression over other parameters, and that is where I came across it while working on #517.

Possible directions

The full answer is presumably #518, where the model exposes the expression so it can be evaluated after the parameter table is applied. Short of that, the silence seems worth fixing on its own. Either the mapping keeps a parameter-table row the model did not report, so a consumer can see it, or lint fails when a parameter marked for estimation is one the model cannot accept a value for. The second is small and would turn a wrong fit into a clear error message.

Reported from petab 0.9.0 and from main at 847b16f.