Default Options
PEtab.jl supports several gradient and Hessian computation methods, as well as the ODE solvers available in OrdinaryDiffEq.jl. As a result, there are many possible choices when creating a PEtabODEProblem
. To simplify usage, PEtab.jl has benchmark derived heuristics to select appropriate default options based on the size of the parameter estimation problem. This page discusses these default options when creating a PEtabODEProblem
.
The default options are based on model size, which is determined by the number of ODEs and the number of parameters to estimate. This is because there is typically no "one-size-fits-all" solution: ODE solvers that perform well for small models may not perform well for large models, and gradient methods that are effective for small models may not be suitable for larger ones. It should also be noted that the defaults are based on benchmarks for stiff biological models. For information on how to configure the PEtabODEProblem
for models outside of biology, see this page.
These defaults often work well, but they may not be optimal for every model as each problem is unique.
Small Models ($\leq 20$ Parameters and $\leq 15$ ODEs)
ODE solver: For small stiff models, the Rosenbrock Rodas5P()
solver is typically the fastest and most accurate. While Julia's BDF solvers like QNDF()
can perform well, they tend to be less reliable and accurate compared to Rodas5P()
in this regime.
Gradient method: For small models, forward-mode automatic differentiation via ForwardDiff.jl is usually the fastest method, often being twice as fast as the forward-sensitivity equations approach. For :ForwardDiff
, it is possible to set the chunk size, which can improve performance. However, determining the optimal value can be challenging, and thus we plan to add automatic tuning.
Hessian method: For small models, computing the full Hessian via ForwardDiff.jl is often computationally feasible. Benchmarks have shown that using the full Hessian improves convergence.
Overall, for small models, the default configuration is:
petab_prob = PEtabODEProblem(model; odesolver=ODESolver(Rodas5P()),
gradient_method=:ForwardDiff,
hessian_method=:ForwardDiff)
If a model has many condition-specific parameters that only appear in a subset of simulation conditions (see this tutorial), runtime can be improved by setting split_over_conditions=true
in the PEtabODEProblem
. For more details, see [this] example.
Medium-Sized Models ($\leq 75$ Parameters and $\leq 75$ ODEs)
ODE solver: For medium-sized stiff models, multi-step BDF solvers like QNDF()
are generally fast and accurate [23]. However, they can fail for models with many events when using low tolerances. In such cases, KenCarp4()
is a reliable alternative.
Gradient method: As with small models, the most efficient gradient method for medium-sized models is forward-mode automatic differentiation via ForwardDiff.jl.
Hessian method: For medium-sized models, computing the full Hessian via ForwardDiff.jl is often computationally infeasible. Instead, we recommend the Gauss-Newton Hessian approximation, which in behcmarks frequently outperforms the commonly used (L)-BFGS approximation [1].
Overall, for medium models, the default configuration is:
petab_prob = PEtabODEProblem(model; odesolver=ODESolver(QNDF(), abstol=1e-8, reltol=1e-8),
gradient_method=:ForwardDiff,
hessian_method=:GaussNewton)
If an optimization algorithm computes both the gradient and Hessian simultaneously, and the Hessian is computed using the Gauss-Newton approximation, it is possible to reuse quantities from gradient computations by setting gradient_method = :ForwardEquations
and reuse_sensitivities = true
. For more information, see this page on the Fides optimizer.
Large Models ($\geq 75$ Parameters and $\geq 75$ ODEs)
While PEtab.jl provides default settings for large models, we recommend benchmarking different methods. This is because selecting the best ODE solver and gradient configuration can substantially impact runtime.
ODE solver: For efficiently simulating large models, we recommend benchmarking various ODE solvers designed for large problems, such as QNDF()
, FBDF()
, KenCarp4()
, and CVODE_BDF()
. Further, we recommend trying a sparse Jacobian (sparse_jacobian = true
) and testing different linear solvers, such as CVODE_BDF(linsolve=:KLU)
. For more information on solving large stiff models in Julia, see this tutorial.
Gradient method: For large models, the most efficient gradient method is adjoint sensitivity analysis (gradient_method=:Adjoint
). PEtab.jl supports the InterpolatingAdjoint()
, GaussAdjoint()
, and QuadratureAdjoint()
algorithms from SciMLSensitivity.jl. The default is InterpolatingAdjoint(autojacvec = EnzymeVJP())
, but we strongly recommend benchmarking different adjoint methods and different autojacvec
options. For further details on adjoint options, see the SciMLSensitivity.jl documentation.
Hessian method: For large models, computing sensitivities (Gauss-Newton) or a full Hessian is not computationally feasible. Therefore, using an L-(BFGS) approximation is often the best option. BFGS support is built into most available optimizers such as Optim.jl, Ipopt.jl, and Fides.py.
Overall, for large models, the default configuration is:
petab_prob = PEtabODEProblem(model, odesolver=ODESolver(CVODE_BDF()),
gradient_method=:Adjoint,
sensealg=InterpolatingAdjoint(autojacvec=EnzymeVJP()))
References
- [1]
- F. Fröhlich and P. K. Sorger. Fides: Reliable trust-region optimization for parameter estimation of ordinary differential equation models. PLoS computational biology 18, e1010322 (2022).
- [23]
- P. Städter, Y. Schälte, L. Schmiester, J. Hasenauer and P. L. Stapor. Benchmarking of numerical integration methods for ODE models of biological systems. Scientific reports 11, 2696 (2021).