Extensions

HybridKinetics has nine package extensions. Each loads automatically when its trigger packages are loaded in the same session. The Catalyst input, the Symbolics output, and the ModelingToolkit export are exported (or documented) entry points of the package; the others are experimental and unexported: call them as HybridKinetics.name and expect their interfaces to change.

Trigger packageExtensionWhat it adds
CatalystHybridKineticsCatalystExtnetwork_from_reactionsystem
SymbolicsHybridKineticsSymbolicsExtsymbolic
Latexify (with Symbolics)HybridKineticsLatexifyExtlatexify of discovered rates
ModelingToolkitHybridKineticsModelingToolkitExtHybridKinetics.export_mtk_system
CUDAHybridKineticsCUDAExtHybridKinetics.to_device and HybridKinetics.gpu_execute for ExecutionConfig(backend = :gpu)
PlotsHybridKineticsPlotsExtHybridKinetics.plot_training
SBMLHybridKineticsSBMLExtHybridKinetics.import_sbml_network
SBMLToolkit and CatalystHybridKineticsSBMLToolkitExtHybridKinetics.import_sbmltoolkit_network
DataDrivenSparseHybridKineticsDataDrivenSparseExtthe HybridKinetics.DataDrivenSparseSTLSQ discovery backend

Catalyst input

network_from_reactionsystem(rs; unknown) converts a Catalyst ReactionSystem into a BiologicalNetwork: the same species in Catalyst's order, the known rate laws compiled to the matching terms, and the reaction named by unknown (an index into Catalyst.reactions(rs) or the string of its description metadata; a vector names several, one per node) as the unknown destruction terms; unknown = nothing compiles everything as known. Supported rates are a parameter, k * Y for a species that is not a substrate, hill(Y, v, K, n) with a literal integer n, and mm(Y, v, K); anything else raises an error naming the reaction and its rate. The How-to recipes page has the full example, and the test suite checks that the tutorial network converted from Catalyst simulates and discovers exactly as the fixture written by hand.

Symbolics output and Latexify

symbolic(candidate, names), symbolic(result::DiscoveryResult, names), symbolic(result::DiscoveryRun; node) and symbolic(result[:S]) return the discovered rational rate as a Symbolics.Num in the named variables (a per-term result uses its network's state names). With Latexify loaded, latexify(result), latexify(result, names), and latexify(candidate, names) render the same expression. The test suite checks that the expression evaluated on the sample grid matches the numeric candidate to floating-point tolerance.

GPU

ExecutionConfig(backend = :gpu) moves the arrays of each experiment to the GPU with cu before calling the user function. There is no batched GPU ODE integration and no GPU training loop.

Plotting

HybridKinetics.plot_training(training, times, observations, dense_times, truth, prediction; state_names) draws one panel per state with the ground truth, the observations, and the model prediction, plus the objective history.

ModelingToolkit

HybridKinetics.export_mtk_system(model; name, discovered) converts the compiled production and destruction terms into a ModelingToolkit ODESystem whose states carry the network's node names. Neural terms appear as placeholder variables nn_i(t) unless discovered (a DiscoveryResult, a candidate, a DiscoveryRun or one of its per-term results) is given, in which case the discovered rational rate replaces the placeholder of that unknown term (a DiscoveryRun replaces all of them) and the system is complete: ODEProblem(complete(sys), u0, tspan, p) solves it. For a fully known network the exported right-hand sides equal Catalyst's own Catalyst.ode_model(rs) (convert(ODESystem, rs) before Catalyst 16), which the test suite checks.

SBML

HybridKinetics.import_sbml_network(path) builds a BiologicalNetwork from the species, reactions, stoichiometry, and modifiers of an SBML file. Kinetic laws (MathML) are not parsed; reactions with an explicit kinetic law compile as unknown neural terms. HybridKinetics.import_sbmltoolkit_network(path) goes through SBMLToolkit and Catalyst and can recognize mass-action reactions. Neither importer produces Hill or Michaelis-Menten metadata.

DataDrivenSparse

DiscoveryConfig(backend = HybridKinetics.DataDrivenSparseSTLSQ(threshold = 1e-3)) uses DataDrivenSparse's STLSQ solver for the explicit coefficient fit. The graph-local library and the implicit rational discovery stay in HybridKinetics. At the time of writing DataDrivenSparse does not resolve against the ModelingToolkit versions this package allows, so the extension cannot be loaded in the package environment; benchmark/probe_datadriven.jl reports the current state.

Fisher identifiability

Not an extension, but also unexported and experimental: HybridKinetics.assess_identifiability(model, p, data, times, u0, tspan) computes a Gauss-Newton Fisher information matrix over the physical parameters at a fit, with parameter correlations, condition number, and credible intervals (HybridKinetics.parameter_credible_intervals, HybridKinetics.estimate_parameter_uncertainty). It is local arithmetic at one parameter point and is not structural identifiability.