HybridKinetics.jl

Hybrid kinetic models for small biochemical networks: learn the unknown rate laws from time-series data, then recover them symbolically.

HybridKinetics fits hybrid models of small biochemical networks. You give it a known interaction graph and known kinetics (mass action, linear decay, Hill, Michaelis-Menten saturation, competitive binding, or a custom rate); it compiles those into a production-destruction ODE

\[\frac{du_i}{dt} = P_i(u) - D_i(u)\,u_i .\]

Any number of destruction terms, on distinct nodes, may be marked unknown. Each is replaced by its own small neural network (a universal differential equation), trained jointly on time-series data from one or more initial conditions, and then approximated symbolically by sparse rational regression (implicit SINDy) over a library built only from that node's graph neighbours. Two unknown terms on the same node are refused, and an unknown production term is out of scope.

The package reports three things: whether each unknown term is practically identifiable from the data (a Fisher-information and scale-collinearity diagnostic), how well the hybrid model reproduces observed and held-out trajectories, and which symbolic terms are recovered. It is a research tool for small networks with a known graph (the benchmarks cover 2- to 6-state networks), not a general-purpose network-inference tool or reaction-network solver.

Version 0.10. The public API may still change before 1.0; see the changelog.

Where to go

  • Getting started: install the package and run a complete fit in a few minutes.
  • Tutorial: the unknown-inhibition example, step by step.
  • Concepts: the model form, the identifiability diagnostic, symbolic discovery, and the reference protocol with its train/holdout split.
  • How-to recipes: CSV data, marking an edge unknown, resimulating, the SciML solve surface, checkpoints, execution backends.
  • Benchmarks: what the recovery benchmarks run and what they show.
  • API reference: every exported name.
  • Extensions: experimental GPU, plotting, ModelingToolkit, SBML, and DataDrivenSparse integrations.
  • Scope and limitations: what the package does not do.

Requirements

Julia 1.10 or newer. Installing the dependencies needs network access to the Julia package server or to GitHub.