PEtab import: condition targetValue that is a parameter/expression (per-condition estimated IC) is rejected

[wshlavacek]

Summary

PEtab import fails when a condition sets a target to a parameter (or expression) value rather than a number. PyBNF's condition importer (ADR-0028) accepts a numeric value, a base pin, or a surrogate relative op, but not a parameter-valued / expression targetValue — a common PEtab idiom for a per-condition estimated initial condition.

Error

NotImplementedError: Condition targetValue 'init_b10_1' for 'init_b10' is an expression,
not a base pin, a surrogate relative op, or a number. Evaluating PEtab condition formulae needs the sympy ...

Root cause (verified)

Bruno_JExpBot2016 condition table sets a target to a parameter:

conditionId   init_b10    init_bcar   ...
model1_data1  init_b10_1  0           ...
model1_data2  0           init_bcar1  ...

init_b10_1 / init_bcar1 are estimated parameters — the condition assigns the model's initial value init_b10 to a per-condition estimated parameter. Bertozzi_PNAS2020 is the same shape (I0_ = I0_CA).

Affected (PEtab benchmark collection)

Bertozzi_PNAS2020, Bruno_JExpBot2016.

Suggested fix

Support a parameter-valued (and, more generally, an expression) condition targetValue: bind the target to the referenced parameter — the per-condition estimated-initial-condition case — evaluating PEtab condition formulae (via sympy, as the error hints) when the value is not a plain number / pin / relative op.

[wshlavacek]

Fixed in ba251fa (on main), design in ADR-0076.

A condition targetValue that names a parameter now imports as a per-condition estimated initial condition: a new condition-perturbation value kind (pset.Mutation(is_param_ref=True)) resolved from the current PSet at apply time, threaded through all four mutant apply paths (net BNG-emit, bngsim net/NF, bngsim SBML/Antimony, RoadRunner). Import binds a free-parameter target to a reference and inlines a fixed one; export emits it back verbatim (PEtab-legal, no surrogate split — byte round-trips); a multi-symbol expression stays the deferred boundary; the gradient path raises GradientNotSupported rather than emit a silent zero column.

Verified end-to-end: the real Bruno_JExpBot2016 now imports and loads into a runnable config. Full test gate green (3218 passed).

Bertozzi_PNAS2020 imports on the condition side but hits a separate, orthogonal limitation — two observables (y_I_NY, y_I_CA) mapping to one model column I_ with per-observable noise — tracked in #503.