[jcschaff]

Problem. pyvcell had no model for VCell events, so every read and write dropped them. simulate_ode writes the model back before solving, so any model with events ran without them.

Found through biosimulations/compose-api#364, where pbsim-common's VCellUTCStep runs SED-ML archives beside COPASI and Tellurium. On BioModel 297 (Ciliberto 2003), cell division and budding are events. VCell tracked COPASI until the first trigger (t ≈ 73 min), then diverged by up to 6× a variable's range.

Data model (pyvcell/vcml/models_app.py)

• Application.events: list[Event], one per VCML <BioEvents><BioEvent> (VCell's BioEvent).
• Event:
  • name;
  • trigger_type: an EventTriggerType (VCell's seven: general, single time, observable above or below a threshold, linear or log range of times, list of times);
  • use_values_from_trigger_time;
  • parameters: a list of EventParameter;
  • assignments: a list of EventAssignment (variable, expression).
  • trigger and delay are convenience properties for the general-trigger function and the delay.
• EventParameter: name, role (an EventParameterRole, VCell's BioEventParameterType), expression and optional unit.
• Application.add_event(name, trigger, assignments, delay=None, use_values_from_trigger_time=True) adds a general-trigger event.
• All are exported from pyvcell.vcml.

Reading and writing

• The reader (visit_BioEvent) follows VCell's XmlReader.getBioEvents.
  • It reads the current form: TriggerType, role-tagged <Parameter>s, <EventAssignment>s.
  • It reads both legacy forms as general triggers: VCell 5.3's <Trigger>expr</Trigger> and the first-pass 5.4 <TriggerParameters triggerClass="TriggerGeneral">, each with <Delay>.
  • It keeps VCell's quirks: a non-general legacy trigger becomes (t > -1.0), which never rises, and a missing UseValuesFromTriggerTime reads as false.
• The writer emits <BioEvents> after the simulations, as VCell's Xmlproducer does. VCell regenerates the math events from them.

Results

• Ciliberto:
  • Round trip: VCell's raw import and pyvcell's rewrite now have the same 4 BioEvents, 10 assignments, 4 math events, 21 ODEs, 140 parameters and 63 functions. Before: 0 events, 19 ODEs, 132 parameters, 59 functions.
  • Against COPASI: VCell now agrees on all 201 points within 3.5e-6 of each variable's range.
• Tests, in tests/vcml/test_events.py:
  • every trigger type and parameter role round-trips;
  • legacy forms are read as VCell reads them;
  • an event imported from SBML survives the round trip and fires: the same as the event-free model up to t = 5, then the same decay restarted from its reset value.
  • Locally the event and ODE tests pass (23), and mypy is clean.
• Corpus sweep, the same 200 models as feat: run non-spatial simulations on VCell's Sundials ODE solver (simulate_ode) #65's: each original document and pyvcell's round trip of it are solved, and the results compared.
  • 108 now give the same answer, against 103 on 0.5.1. Five models moved from "different variables" to "same", and none got worse.
  • The two models that crashed the solver on 0.5.0 now fail with 0.5.1's zero-equation message, or solve with different variables.
  • One of them, biomodel_98130468, loses all its equations in pyvcell's round trip. That's a separate gap for a follow-up.

Not modelled: RDF and annotation metadata (<Annotation>, vcml:uriBinding, ...) is still dropped on a round trip. It doesn't change the math.

🤖 Generated with Claude Code