…385) Layer H of the #385 gradient-plumbing epic: lift the measurement-model gate clause so an expression-`observableFormula` fit (the SBML/Antimony / new-era PEtab path) is differentiable end to end. Two pieces: 1. SBML/Antimony backend forward sensitivities. BngsimSbmlModelNoTimeout (and Antimony, by inheritance) now mirrors the net backend's gradient path: enable_output_sensitivities(params, ic) (capability-gated), the bind-by-id sensitivity_entity_namespace() (global params -> parameter axis, each species as its own bare IC), sensitivity_params/sensitivity_ic threaded into the ODE Simulator (refused for ssa), and a `species:`-keyed OutputSensitivities tensor attached in _result_to_data (the IC axis rescaled by the species' PyBNF-value -> concentration unit factor). Confirmed feasible: bngsim's SBML loader computes per-species sensitivities (GH #205). 2. Measurement-layer derivative. MeasurementModel.prediction_sensitivity is the vectorized-formula analog of PerMeasurementModel's: the formula's exact symbolic gradient chained through each referenced column's sensitivity plus any fit parameter it names directly (an observation-model scale/offset, which enters the prediction and so lands in the residual-Jacobian -- a square). The assembly's raw_sens accessor recognizes a materialized measurement column and delegates to it (calling back into itself for referenced columns, so routing / normalization fold in); _selector_for learns the `species:` prefix. A plain column collapses to the tensor sensitivity, so the no-measurement path is byte-identical. Gate: drop the whole-objective `self.measurement is not None` clause from _require_gradient_supported; the assembly raises GradientNotSupported cleanly for a referenced column with no forward-sensitivity column. Tests (test_gradient_assembly.py): numpy chain-rule (direct scale column = a square; numeric-only touches the model axis; gate now accepts) + FD oracles -- measurement layer on the net backend (isolates the derivative), a bare-species SBML fit (isolates the backend `species:` path), and the full seam: a small SBML model fit with `obs = w*A + B` over abc.xml. Backend unit tests (test_bngsim_sbml_bridge.py): the namespace contract, Simulator threading, and ssa refusal. Full bngsim pre-push gate green (535 passed), ruff clean. Docs: gradient_fitting.rst new "Measurement-model layer (SBML / Antimony)" section; removed from the not-yet-supported list. Unblocks #386. Closes #455.

Uh oh!

Recover from empty sorted_params_backup #205

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions