assignments_to_experiments: convert assignments that switch more than once

[FFroehlich]

Extends assignments_to_experiments (#527) to assignments that switch more than once.

Why. Of the benchmark-collection problems with time-dependent piecewise assignments, main converts two (Isensee_JCB2018, Raimundez_PCB2020) and rejects three. Two of those have an exact form as experiment periods:

• Brannmark_JBC2010, a sum of doses given at different times:
  insulin := insulin_dose_1 * piecewise(0, time - insulin_time_1 < 0, 1) + insulin_dose_2 * piecewise(0, time - insulin_time_2 < 0, 1)
• Oliveira_NatCommun2021, phases one after another as nested piecewise functions:
  beta := piecewise(beta_0, time < t_1, piecewise(beta_1, time < t_2, ...))

What. An assignment may now depend on time through any number of piecewise conditions — in a sum, in several pieces, or nested — within the scope agreed in #527, applied per condition:

• every condition is a single inequality that can be re-arranged into t </<=/>/>= ... (no &&/||, no equality, linear in time with a non-zero slope);
• every time-dependent piecewise function has an explicit "otherwise" piece;
• time enters only through those conditions;
• every switching time evaluates to a real number in each experiment: it may depend on constant model parameters, on parameters fixed in the parameter table, or on parameters the condition table sets, but not on estimated parameters;
• all other symbols are truly constant model parameters.

The value on each interval between switches is the assignment with every condition replaced by whether it holds there. A switch that changes the value becomes a period; one that does not (e.g. two phases with equal values) adds nothing.

The value before every switch follows from each condition's slope and is kept in the model. If it refers to parameter table symbols (e.g. an estimated d1 in piecewise(d1, time < 10, d2)), it is also set at the start of the first period, since the model can only hold a number for it. (On main, such an assignment fails with a TypeError.)

Switching times are checked up front, in _supported_time_switches: one that depends on an estimated parameter is rejected with an error naming those parameters. Otherwise, the symbols of a switching time that no condition sets take their value from the parameter table, if fixed there, or else from the model. On main, a switching time given by a constant model parameter is rejected. A condition that sets a switching time to an expression in estimated parameters is still rejected per experiment.

Everything that converts on main converts as before: all existing tests pass unchanged except in test_assignments_to_experiments_unsupported, where the two cases asserting that these two shapes are rejected are replaced by a conjunction (time < 10 && time > 5), still rejected, and the estimated switching time now expects the new error. The docstring's scope is updated accordingly; the preequilibration warning and the final lint are unchanged.

Checked

• New tests: a sum of doses, several pieces, nested pieces, a switch that does not change the value, and condition-specific doses as in Brannmark (each experiment gets its own switching times and amounts); switching times given by a constant model parameter, a parameter fixed in the parameter table, or the condition table; a condition that sets a switching time to an estimated parameter (rejected); a pre-switch value that refers to estimated parameters.

• On the benchmark problems (assignments_to_experiments after petab.v2.Problem.from_yaml, then simulated with AMICI's PEtab v2 importer at the nominal parameters, atol 1e-10):

  problem	main	this branch
  Brannmark_JBC2010	rejected	converted (17 periods); llh −141.889201 → −141.889187
  Oliveira_NatCommun2021	rejected (shape)	rejected: its switching times t_1, t_2 are estimated, which no period time can be; the error names them
  Fujita_SciSignal2010	rejected	rejected: EGF_rate*time puts time in a value
  Isensee_JCB2018	converted, Δllh −1.8e-3	unchanged
  Raimundez_PCB2020	converted, Δllh 4.7e-5	unchanged (simulated with ExperimentsToSbmlConverter: write one flat initial assignment per target #546)

  So Oliveira does not become convertible — it now fails for the reason that actually prevents it, rather than for its shape.

  The fixed switching times and the pre-switch value were added after the simulations above. They leave the converted Brannmark, Isensee and Raimundez problems identical (experiment table, condition table, SBML), so those log-likelihoods still hold.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MNJePorYs9zmPYMam6g4EP

[codecov-commenter]

Codecov Report

❌ Patch coverage is 86.00000% with 7 lines in your changes missing coverage. Please review.
✅ Project coverage is 78.07%. Comparing base (8bbe776) to head (79aa38e).

Files with missing lines	Patch %	Lines
petab/v2/converters.py	86.00%	3 Missing and 4 partials ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #556   +/-   ##
=======================================
  Coverage   78.07%   78.07%           
=======================================
  Files          68       68           
  Lines        7793     7823   +30     
  Branches     1389     1399   +10     
=======================================
+ Hits         6084     6108   +24     
- Misses       1224     1227    +3     
- Partials      485      488    +3     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[dilpath]

Looks neat! Thanks.

I agree Oliveira can't be done. For Fujita it could be possible like
Current:

EGF = EGF_rate * time + EGF_0    if    time <= EGF_end
      0                          otherwise

Possible solution:

indicator = 1    if    time <= EGF_end
            0          otherwise

EGF = (EGF_rate * time + EGF_0) * indicator

But this requires a bit more code to create new SBML assignment rules for indicator variables of piecewises.

[dilpath]

Regarding #557, do you prefer that solution or the indicator suggestion above? I think the indicator method is more generally applicable, i.e. it does not require a linear dependence on time.

[FFroehlich]

Regarding #557, do you prefer that solution or the indicator suggestion above? I think the indicator method is more generally applicable, i.e. it does not require a linear dependence on time.

I actually do like the indicator suggestion better!

[FFroehlich]

Regarding #557, do you prefer that solution or the indicator suggestion above? I think the indicator method is more generally applicable, i.e. it does not require a linear dependence on time.

I actually do like the indicator suggestion better!

Thinking about this again, I am not entirely happy with either solution and the solution in #557 is actually closer to the goal we are trying to achieve: the PEtab owns the experiment definition and the model owns the description of the biology.